Citation
Preliminary study on medium format cameras comparison via photo response non-uniformity

Material Information

Title:
Preliminary study on medium format cameras comparison via photo response non-uniformity
Creator:
Frtiz, Edgar
Place of Publication:
Denver, CO
Publisher:
University of Colorado Denver
Publication Date:
Language:
English

Thesis/Dissertation Information

Degree:
Master's ( Master of science)
Degree Grantor:
University of Colorado Denver
Degree Divisions:
Department of Music and Entertainment Industry Studies, CU Denver
Degree Disciplines:
Recording arts
Committee Chair:
Grigoras, Catalin
Committee Members:
Smith, Jeff M.
Golemboski, Carol

Notes

Abstract:
This is a preliminary study on Medium Format cameras. This study utilizes Photo Response Non-Uniformity (PRNU), which is part of the background noise inherent in all digital camera sensors, to compare between Medium Format cameras of the same make and model. PRNU is comprised of tiny imperfections at the pixel level and is sometimes referred to as the camera’s “fingerprint.” This background noise can be extracted and utilized to compare between a reference image and a source camera. Two different methods of PRNU extraction were utilized in this study, a Gaussian method, and a Wiener method. The inter-variability and intra-variability between a reference image and sets of natural scene images was used to compare cameras.

Record Information

Source Institution:
University of Colorado Denver
Holding Location:
Auraria Library
Rights Management:
Copyright Edgar Fritz. Permission granted to University of Colorado Denver to digitize and display this item for non-profit research and educational purposes. Any reuse of this item in excess of fair use or other copyright exemptions requires permission of the copyright holder.

Downloads

This item has the following downloads:


Full Text
PRELIMINARY STUDY ON MEDIUM FORMAT CAMERAS COMPARISON
VIA PHOTO RESPONSE NON-UNIFORMITY by
EDGAR FRITZ
B.A., University of California, San Diego, 2001
A thesis submitted to the Faculty of the Graduate School of the University of Colorado in partial fulfillment of the requirements for the degree of Master of Science Recording Arts Program
2019


This thesis for the Master of Science degree by
Edgar Fritz
has been approved for the Recording Arts Program by
Catalin Grigoras, Chair Jeff M. Smith Carol Golemboski
Date: December 14, 2019


Fritz, Edgar (M.S., Recording Arts Program)
Preliminary Study on Medium Format Cameras Comparison Via Photo Response Non-Uniformity Thesis directed by Associate Professor Catalin Grigoras
ABSTRACT
This is a preliminary study on Medium Format cameras. This study utilizes Photo Response Non-Uniformity (PRNU), which is part of the background noise inherent in all digital camera sensors, to compare between Medium Format cameras of the same make and model. PRNU is comprised of tiny imperfections at the pixel level and is sometimes referred to as the camera's "fingerprint." This background noise can be extracted and utilized to compare between a reference image and a source camera. Two different methods of PRNU extraction were utilized in this study, a Gaussian method, and a Wiener method. The inter-variability and intra-variability between a reference image and sets of natural scene images was used to compare cameras.
The form and content of this abstract are approved. I recommend its publication.
Approved: Catalin Grigoras


To my wife and kids, for always encouraging me to reach my goals.
IV


ACKNOWLEDGEMENTS
Thank you to my professors at the University of Colorado Denver.
v


TABLE OF CONTENTS
CHAPTER
I. DIGITAL PHOTOGRAPHY BACKGROUND.........................................................1
Digital Noise.....................................................................1
Photo Response Non-Uniformity.....................................................3
PRNU and the Daubert Standard for Evidence Admissibility..........................4
Prior Research....................................................................5
Medium Format Sensors.............................................................7
II. CAMERA FINGERPRINT...................................................................10
Materials and Methods............................................................10
PRNU Suitability Test............................................................11
PRNU Extraction..................................................................16
III. SOURCE CAMERA COMPARISON............................................................21
FUJIFILM GFX 50s.................................................................22
FUJIFILM GFX 100.................................................................23
Hasselblad X1DII 50c.............................................................25
Hasselblad H6D 100c..............................................................26
Phase One IQ3 100MP..............................................................27
Phase One IQ4 150MP..............................................................28
IV. CONCLUSION...........................................................................30
REFERENCES...............................................................................32
APPENDIX.................................................................................34
VI


LIST OF FIGURES
FIGURE
1 FUJIFILM Sensor Size Comparison...................................................................8
2 Flasselblad Sensor Size Comparison..................................................................9
3 Phase One Sensor Size Comparison....................................................................9
4 RAW to TIFF Conversion............................................................................10
5 FUJIFILM GFX 50s PRNU Suitability Test ............................................................11
6 FUJIFILM GFX 100 PRNU Suitability Test.............................................................12
7 Flasselblad X1DII 50c PRNU Suitability Test........................................................13
8 Flasselblad H6D 100c PRNU Suitability Test.........................................................14
9 Phase One IQ3 100MP PRNU Suitability Test..........................................................15
10 Phase One IQ4 150MP PRNU Suitability Test..........................................................16
11 FUJIFILM GFX 50s Extracted PRNU Pattern...........................................................17
12 FUJIFILM GFX 100 Extracted PRNU Pattern...........................................................18
13 Flasselblad X1DII 50c Extracted PRNU Pattern.......................................................18
14 Flasselblad H6D 100c Extracted PRNU Pattern........................................................19
15 Phase One IQ3 100MP Extracted PRNU Pattern.........................................................19
16 Phase One IQ4 150MP Extracted PRNU Pattern.........................................................20
17 Gaussian Method FUJIFILM GFX 50s Log Likelihood Ratio Results.....................................22
18 Wiener Method FUJIFILM GFX 50s Log Likelihood Ratio Results.......................................23
19 Gaussian Method FUJIFILM GFX 100 Log Likelihood Ratio Results.....................................24
20 Wiener Method FUJIFILM GFX 50s Log Likelihood Ratio Results.......................................24
21 Gaussian Method Flasselblad X1DII 50c Log Likelihood Ratio Results................................25
22 Wiener Method Flasselblad X1DII 50c Log Likelihood Ratio Results..................................25
vii


23 Gaussian Method Hasselblad H6D 100c Log Likelihood Ratio Results................................26
24 Wiener Method Hasselblad H6D 100c Log Likelihood Ratio Results..................................26
25 Gaussian Method Phase One IQ3 100MP Log Likelihood Ratio Results................................27
26 Wiener Method Phase One IQ3 100MP Log Likelihood Ratio Results..................................28
27 Gaussian Method Phase One IQ4 150MP Log Likelihood Ratio Results................................29
28 Wiener Method Phase One IQ4 150MP Log Likelihood Ratio Results..................................29
viii


LIST OF TABLES
TABLE
1 Cross Sample of Typical Cameras and Specifications.....................................6
2 Medium Format Sensors Included in this Study and their Specifications..................8
3 FUJI FILM GFX 50s Reference Image from snEdFr71002897 Results.........................22
4 FUJIFILM GFX 100 Reference Image from snEGF76A02867 Results...........................23
5 Flasselblad X1DII 50c Reference Image from snEGF76A02867 Results......................25
6 Flasselblad H6D 100c Reference Image from sn54520225540E Results......................26
7 Phase One IQ3 100MP Reference Image from snFIP001078 Results..........................27
8 Phase One IQ4 150MP Reference Image from snJD010673 Results...........................28
IX


CHAPTER I
DIGITAL PHOTOGRAPHY BACKGROUND
Modern digital cameras are ubiquitous in our society today. The digital revolution transformed photography forever. Digital photography eliminated the wait of developing and printing images before confirming if you even captured the shot. It brought users instant feedback and the ability to review images wherever they were. But digital cameras brought much more than that, they include a wealth of information never before captured in the film era. Digital images contain data embedded within the file. This is called the metadata, and it contains valuable information that can be invaluable from a forensic standpoint. Embedded within the file you may find information about when the image was taken, the make, model, and serial number of the camera, geolocation, camera settings, lens information, and even if an image has been edited after it was captured. Information not only can be obtained from the file's metadata, but the images themselves can be analyzed and inferences can be made from them. Digital images can be used forensically to identify locations from elements within the image, dates and times from shadows, subject identification from biometrics, size of objects from known objects comparison, and many more. You can even ascertain which camera took an image from the inherent background noise of a digital sensor. Today, you'd be hard pressed to find someone without some sort of image capturing device on them at all times. Digital images can be the one piece of corroborating evidence that ties together an entire court case. They are a big part of today's criminal justice system, that is why forensic scientists have been studying their properties since the dawn of the digital age.
Digital Noise
A digital camera works by allowing light to enter though an opening in an optical lens (aperture) for a specified length of time (shutter speed). The photons of light that enter are focused on a digital sensor which is made up of an array of pixels. Each pixel is made of a photosensitive material responsible for gathering photons of a specific wavelength of light. Photons captured by the
1


photosensitive part of a pixel stimulate the emission of electrons, one for each captured photo. The accumulated photoelectrons are counted and proportionally converted to a voltage. That number can then be amplified depending on the sensitivity to light set on the camera (ISO setting). This voltage is then digitized in an analog to digital converter (ADC) and converted into Is and Os to be written out to a file. This binary code can be thought of as the RAW image data. Additional information will be written to the file by the camera as metadata for future decoding of the image.
The process described above would be an ideal camera with ideal components and processes. But no camera can perform to the ideal, hence no image is an exact representation of the photons that struck the sensor. All digital images will possess some inherent noise, which comes from several sources during the image acquisition process. Noise can be thought of as variations in color or intensity from what the true scene was. This noise can be grouped into two general types: Random Noise, and Fixed Pattern Noise.
Each electronic circuit in the signal processing chain suffers from random fluctuations. These fluctuations can come from the photosensitive components of the sensor, the ISO gain, and/or the ADC digitization. Horizontal and Vertical Banding Noise (HVBN) is an example of random noise in the image capturing process. HVBN manifests itself as linear color and intensity fluctuations in the rows or columns and look like bright bands in the image. HVBN noise can be caused by errors in the sensor readout; that is when rows of pixels are activated and readout down the pipeline. ISO settings will also introduce HVBN as the ISO is increased. ISO is a scalar multiplier of the photoelectron count, so higher settings will amplify errors. ADC digitization can introduce HVBN noise with errors in the digital conversion process (1). There are several other sources of random noise that can affect a digital image.
Fixed Pattern Noise is a pattern that will be stamped in all images from the same sensor under the same conditions. There are many factors that will contribute to Fixed Pattern Noise at different stages of the image capturing process. Thermal agitation of electrons in the photoreceptors of a sensor
2


can liberate some electrons. These freed electrons will be counted as if they were photons that were
present in the image and will be added to the RAW data. Thermal electrons are freed at a relatively constant rate creating constant imperfections in an image. Thermal Noise will increase with exposure time and will show up as "hot pixels" in an image. These hot pixels are overexposed pixels that will show up in the same spot repeatedly.
Pixel Non-Uniformity is a general term for the imperfections in the micron sized pixels of a sensor, it is comprised of two types. Dark Signal Non-Uniformity (DSNU) is the offset from the average across the sensor at a particular temperature and integration time without external illumination. Basically, this noise is found when an image is taken in complete darkness. DSNU will be fixed in all images taken in total darkness. Pixel Non-Uniformity is also found in illuminated pixels. Pixels have acceptable levels of variation in capturing efficiency, photon counting accuracy; variations in pixel size, shape, and substrate material. Not all pixels in a sensor are made the same, these tiny variations are called the Photo Response Non-Uniformity (PRNU). PRNU comes from the gain or ratio between optical power on a pixel versus the electrical signal output. When the photoreceptive cells in a sensor are illuminated evenly, each cell should ideally output the same voltage, but they don't. PRNU will be ingrained in all images as part of the background noise and will be an ever-present fixed pattern. That is why PRNU is often referred to as the camera's "fingerprint." PRNU noise grows in direct proportion to the exposure level, so brighter pixels will show more PRNU noise. PRNU noise will be reproduced for all images taken under the same conditions and settings (2).
Photo Response Non-Uniformity
PRNU can be used for several forensic methods. One example is to test for the presence of a specific PRNU in an image, which can achieve reliable device identification (prove that a certain camera took a given image). PRNU can also help prove two images were taken by the same device (device linking). The presence of PRNU is indicative of the fact that the image is natural and not a computer
3


rendering. By establishing the absence of the PRNU in individual image regions, it is possible to discover
maliciously forgery attempts. By detecting the strength or form of the PRNU, it is possible to reconstruct some of the processing that was done to an image. PRNU is used as a template to estimate geometrical processing, such as scaling, cropping, or rotation. Non-geometrical operations are also going to influence the strength of the PRNU in the image and thus can be potentially detected. The spectral and spatial characteristics of the PRNU can be used to identify the camera model or distinguish between a scan and a digital camera image (the scan will exhibit spatial anisotropy) (3).
PRNU and the Daubert Standard for Evidence Admissibility PRNU has proven to be extremely valuable as a forensic tool. During the court proceedings of the United States of America v. Nathan Allen Railey, United States Court of Appeals for the Southern District of Alabama, prosecutors showed child pornography images in question were created by the defendant's camera by comparing the PRNU. This identification was challenged in an admissibility hearing to verify the technology met the Daubert standard. It met the challenge as follows (4):
Whether the theory or technique employed is generally accepted in the scientific community: PRNU source camera identification was first introduced in 2005 by Lukas, Fridrich, and Goljan, where they detailed how to determine a digital image's origin using sensor pattern noise (5). Numerous papers have subsequently been written expanding on these methods and techniques.
Whether the technique has been subjected to peer review or publications: There have been numerous peer-reviewed publications on the subject appearing in such periodicals as IEEE, SPIE, the International Conference on Image Processing, EURASIP Journal on Image and Video Processing, International Conference on Computer Science and its Applications, etc.
Whether the technique has been tested or can be tested: There exist many techniques for PRNU extraction and comparison. These methods and algorithms are tested and peer reviewed as they
4


are developed. Validation is a typical part of the scientific process, and some validation has been conducted by the Law Enforcement Analysis Facility, an independent testing company in NY (6).
Whether an error rate can be determined: There have been several large-scale tests conducted to test identification error and false positive rates. One such study was conducted by Miroslav Goljan on more than a million images. It tested PRNU image identification via peak to correlation energy, with a false positive rate at less than .00024%, and the false rejection rate at less than .0238% (7).
PRNU met the Daubert standard in this court case, and expert testimony regarding PRNU implementation was admitted into court. US v Railey provided support for future PRNU implementation against Daubert challenges since the theory or technique employed met the standards set and is generally accepted in the scientific community as a valid method.
Prior Research
Camera identification has been studied extensively for decades now. Researchers have found several ways of matching images with the camera that produced them. A variety of features have been employed, including color filter array design, pixel color value interpolation algorithm, image sensor anomalies, lens characteristics and anomalies, and image processing pipeline characteristics (8).
PRNU extraction for camera comparison has also been widely studied. There have been various attempts to improve the accuracy of digital camera identification. Several approaches aim to achieve this by improving the estimate of PRNU. Some apply Maximum Likelihood Estimation (MLE) for estimation of a multiplicative factor from reference images (9). Other research suppresses unwanted artifacts from sensor pattern noise. This is achieved by preprocessing the PRNU by subtracting Zero Mean and applying Wiener filtering (10). Some research proposes extracting PRNU per color channel in a Color Decoupled PRNU extraction method (11). The effect of denoising filter has been investigated using sparse 3D transform-domain collaborative filtering for more accurate noise extraction (12). Another documented method uses edge adaptive PRNU predictor context adaptive interpolation (13,
5


14). Another technique is to pre-process the estimated PRNU for removal of non-unique artifacts.
These techniques work by identifying and suppressing the peaks according to local characteristics in the magnitude spectrum of the reference PRNU with spectrum equalization, improving the detection accuracy further (15). Furthermore, other methods improve performance by using better statistical detection. Some use peak to correlation energy to attenuate hidden periodic patterns (16). Others use correlation over circular cross-correlation norm to further decrease the false-positive rate (14).
Though there are several methods and techniques that explore source camera identification via PRNU comparison, few use RAW image data to achieve this. Those that do use RAW data only focus on either identification of source device of the same model, or different devices of the same model but not different models or cameras (17). If you browse through PRNU publications, you will find researchers typically utilize JPEG compressed images from mass manufactured consumer level cameras. The sensor quality and size used in typical research overlooks the best available sensors in the market today. One paper tried to specifically address this issue but only obtained data from "mid-end and high-end popular consumer digital cameras"(18), not take into consideration professional level medium format cameras. Below is a cross sample of typical cameras utilized in PRNU literature:
Table 1 Cross Sample of Typical Cameras and Specifications
6


Medium Format Sensors
We have discussed how PRNU has been studied by numerous people with various methods and under various conditions for accuracy, and robustness, but no research has been conducted on the top quality sensors in the market. The objective of this thesis is to look beyond consumer level cameras and study professional level medium format cameras. Medium format refers to any sensor larger than a full frame DSLR (or the traditional 35mm sensor at 24 x 36mm). The advantage of medium format sensors comes not just from their size, but from the larger pixel area, and meticulously calibrated sensors. Medium format sensors provide a wider dynamic color range, larger field of view, and shallower depth of field. Larger pixels have more light-gathering capability. Finer and more detailed tonal information can be produced, delivering more information in the final file. But more importantly for camera comparison, better build, quality control, and smaller production scale.
This thesis focuses specifically on today's top end medium format camera manufacturers, FUJIFILM, Hasselblad, and Phase One. The following cameras were used in this research as being recent top end cameras from said manufacturers: FUJIFILM GFX 50S with a 50MP sensor, FUJIFILM GFX 100 with a 100MP sensor; Hasselblad H6D 100c camera with a 100MP sensor, Hasselblad XlDII-50c mirrorless camera with a 50MP sensor; Phase One IQ3 with a 101MP sensor, Phase One IQ4 with a 151MP sensor. Medium format sensors provide much larger images with better resolution and more mega pixels than any sensor found in prior studies into PRNU. The Pixel Area (size of each pixel) is also much larger allowing for greater light gathering capability. And the RAW recording allows for 14 and even 16 bits of depth yielding millions of more colors, that is 1.6 x 107 for an 8 bit image vs. 2.8 x 1014 for a 16 bit image. The following table shows the specifications for the cameras used in this study.
7


Table 2 Medium Format Sensors Included in this Study and their Specifications
Make Model Sensor Type Sensor Size Sensor Resolution Pixel Area Mega Pixels File Format
FUJIFILM GFX50S CMOS 43.8mm x 32.9mm 8256x6192 28.2 pm2 51.4 MP RAW 14 bit RAF
FUJIFILM GFX 100 CMOS BSI 43.8mm x 32.9mm 11648 x 8736 14.14 pm2 101.9 MP RAW 16 bit RAF
Hasselblad X1DII-50C CMOS 43.8 x 32.9mm 8272 x6200 27.98 pm2 51.3 MP RAW 16 bit 3FR
Hasselblad H6D-100C CMOS 53.4 x 40.0 mm 11600 x 8700 21.16 pm2 100 MP RAW 16 bit 3FR
Phase One IQ3-100MP CMOS 53.7 x 40.4mm 11608 x 8708 21.44 pm2 101.1 MP RAW 16 bit HQ
Phase One IQ4-150MP CMOS BSI 53.4 x 40.0mm 14204 x10652 14.14 pm2 151 MP RAW 16 bit HQ
Here is what the manufacturers have to say about their cameras.
The FUJIFILM GFX 50S features a 43.8x32.9mm CMOS medium format sensor, boasting an effective resolution of 51.4 million pixels. The FUJIFILM GFX 100 pairs a newly-developed back-illuminated 102MP imaging sensor with a blazingly fast X Processor 4 processing engine to create a combination capable of outputting 16-bit images with amazing color fidelity, rich shadow detail, and incredible dynamic range. Its back-illuminated structure enhances image quality by bringing the exposure plane closer to the color filter array. Sharpness is also enhanced, while moire and false colors are near eliminated through the omission of an optical low-pass filter (19).
Format size comparison
Figure 1 FUJIFILM Sensor Size Comparison
8


Hasselblad camera sensors are individually calibrated to ensure maximum performance in
any given situation. A massive amount of data is gathered for each unit in the production stage in order to study the variations that can occur and see how the sensor performs under different circumstances. Through Hasselblad's rigorous sensor testing, all irregularities are corrected automatically when shooting with any Hasselblad camera (20).
H6D-100C X1D-50C ’FULL FRAME' DSLR
53 X 40mm 44 x 33mm 36 x 24mm
Figure 2 Hasselblad Sensor Size Comparison
The full frame medium format sensors found in the Phase One IQ Digital Backs are 2.5x larger than that of full frame DSLR cameras, and 1.5x larger than cropped sensor mirrorless medium format. The use of a larger sensor means more light, and therefore more information is available when converting captured light into a digital file. The availability of more data allows Phase One to honor the quality of that data, translating to more efficient use of information and less interpolation (21).
Figure 3 Phase One Sensor Size Comparison
9


CHAPTER II
CAMERA FINGERPRINT Materials and Methods
Flat field (evenly illuminated neutral wall) images were collected from each camera. The images were slightly underexposed, so no pixels maxed out their threshold. A few natural scene images were also collected for comparative analysis. In order to obtain the full quality of each sensor, RAW images were collected. RAW images would avoid the undesired addition of JPEG artifacts into the background noise. In-camera noise reduction was not applied to any of the images collected to maintain proper PRNU signatures. ISO 100 was used for all images collected to limit random noise. All calculations were performed on a Windows 10, 64-bit operating system with Intel Core 17-8550U CPU at 2.0 GHz and 16.0 GB of RAM. Matlab R2019b version 9.7.0 build 1190202 was utilized for all computations.
RAW files were converted into uncompressed TIFF files in Adobe Bridge CC 2019 v9.1.0.338 with Adobe Photoshop CC 2019 v 20.0.7 for the Hasselblad and FUJI FILM cameras. The Image Processor tool converted folders of RAW images into uncompressed TIFF files. Note this process does not provide bit depth options and automatically converted the images into 8 bits per channel. Phase One images were processed using Capture One Pro vl2.1.1.19 since Adobe does not support I IQ. file format. The Process Recipe tool was used in Capture One to convert RAW images into uncompressed TIFF files. This process does allow for 16 bit TIFF conversion, but files were converted to 8 bits per channel for consistency.
Figure 4 RAW to TIFF Conversion
10


PRNU Suitability Test
Once the images were converted to TIFF files, they were tested for PRNU suitability with a denoising filter prior to PRNU extraction. The Wavelet method was uses to denoise the images. This is a spatial adaptive statistical model for wavelet image coefficients as zero-mean Gaussian random variables with high local correlation. This method assumes a marginal prior distribution on wavelet coefficients variances and estimates them using an approximate maximum a posteriori probability rule. Then an approximate minimum mean squared error estimation procedure was applied to restore the noisy wavelet image coefficients (22). Simply put, the wavelet method analyses data at a local level as well as at a global level. The result tells us if the images in question are suitable for PRNU extraction.
The flat field reference images collected were tested with a Wiener method for suitability of PRNU extraction. The following figures show each set of images and the cross correlation of the intra/inter-variability. This process plots the first image in the set as the evidence and compares the inter-variability (the variation within each file), and the intra-variability (the variation between all the files), and calculates the cross correlation amongst them for PRNU suitability. Note some reference flat field image sets were reduced to 20 images for computational considerations. All reference flat field image sets were found to be suitable for PRNU analysis.
8256x6192-Fuji-GFX-50S-snEdFr71002897
Figure 5 FUJIFILM GFX 50s PRNU Suitability Test
11


Computes CC1, CC2... ans =
0.0011 0.6982 0.0004 0.7434
20 images, very good for PRNU analysis, 100*k2=0.10807 < kl=0.69817
8256x6192-Fuji-GFX-50S-snEdFr71002897
Figure 5 Continued
11648x8736-Fuji-GFX-100-snEdFr76A02867
Computes CC1, CC2... ans =
-0.0005 0.7289 -0.0003 0.7695
20 images, very good for PRNU analysis, 100*k2=-0.054285 < kl=0.72888
11648x8736-Fuji-GFX-100-snEdFr76A02867
Figure 6 FUJIFILM GFX 100 PRNU Suitability Test
12


8272x6200-Hassleblad-X1 DII-50C-snEdFr565101 AB529F
Computes CC1, CC2... ans =
0.0007 0.7781 0.0011 0.8505
24 images, very good for PRNU analysis, 100*k2=0.067744 < kl=0.77806
8272x6200-Hassleblad-XlDII-50C-snEdFr565101AB529F
Figure 7 Hasselblad X1DII 50c PRNU Suitability Test
13


11600x8700-Hassleblad-H6D-100C-snEdFr54520225540E
Computes CC1, CC2... ans =
-0.0008 0.7376 -0.0008 0.7840
30 images, very good for PRNU analysis, 100*k2=-0.075368 < kl=0.73763
11600x8700-Hassleblad-H6D-100C-snEdFr54520225540E
Figure 8 Hasselblad H6D 100c PRNU Suitability Test
14


11608x8708-PhaseOne-IQ3-100MP-snEdFrHP001078
Computes CC1, CC2... ans =
0.0001 0.7378 -0.0002 0.7824
20 images, very good for PRNU analysis, 100*k2=0.0072872 < kl=0.73777
11608x8708-PhaseOne-IQ3-100MP-snEdFrHP001078
Figure 9 Phase One IQ3 100MP PRNU Suitability Test
15


14204x10652-PhaseOne-IQ4-150MP-snEdFrJD010673
Computes CC1, CC2... ans =
0.0007 0.7781 0.0011 0.8505
24 images, very good for PRNU analysis, 100*k2=0.067744 < kl=0.77806
8272x6200-Hassleblad-XlDII-50C-snEdFr565101AB529F
Figure 10 Phase One IQ4 150MP PRNU Suitability Test PRNU Extraction
Once confiremed the flat field images were suitable, the PRNU was extracted and displayed.
This was achieved by taking the noise residuals of the reference flat field images and convert them to a grayscale signal. First, the fingerprint is estimated in each of the three color channels (red, blue, green). Then, the channels are combined using the common linear transformation RGB -> grayscale, the sample mean is subtracted and averages of each column and each row in each of four sub-sampled 2-D signals
16


corresponding to four types of pixels in the Bayer CFA are subtracted from all elements. This "Zero-
Mean" procedure removes a large portion of Non Unique Artifacts introduced by demosaicking (that is noise that can be found in cameras from the same make and model, making them not unique to the sample camera). To remove any residual artifacts from the estimated fingerprint, the fingerprint was finally filtered using an adaptive Wiener filter in the frequency domain (7).
The following images are the extracted PRNU noise from the reference flat field image set.
Figure 11 FUJIFILM GFX 50s Extracted PRNU Pattern
17


Figure 12 FUJIFILM GFX 100 Extracted PRNU Pattern
Figure 13 Hasselblad X1DII 50c Extracted PRNU Pattern
18


Figure 14 Hasselblad H6D 100c Extracted PRNU Pattern
Figure 15 Phase One IQ3 100MP Extracted PRNU Pattern
19


Figure 16 Phase One IQ4 150MP Extracted PRNU Pattern
Note the subtle outline still visible in Figure 16. This was due to the flat field images containing part of some furniture that was overlooked when colleting the flat field images.
20


CHAPTER III
SOURCE CAMERA COMPARISON
The extracted PRNU was then used to compare a natural scene reference image taken with the reference flat field cameras, and a small open source image database created for this research. PRNU noise was obtained by suppressing scene content from an image with a denoising filter and then averaging the noise residuals. In this study, Gaussian denoising filter, and Wiener denoising filter were used to suppress scene images. The Gaussian filter removes noise by blurring an image. It achieves this by distributing pixels in Gaussian (bell-shaped) distribution across the entire image. The Wiener filter both recovers an image from blurring of a lowpass filter (inverse filtering) and reduces noise by smoothing at the same time, it removes additive noise and inverts the blurring simultaneously (23). Two methods were used to study the accuracy and reliability in comparing between different cameras of the same make and model. Note the appendix will contain figures from all camera comparisons.
In order to conduct this analysis, open source RAW images and some JPEG images were obtained from www.dpreview.com.www.photographvblog.com.www.flickr.com.www.phaseone.com. www.hasselblad.com. and www.FUJIFILM.com. to build a small database of images. All open source RAW images were converted to TIFF files for comparison. The open source images collected were of natural scenes with normal dynamic range and exposure.
Correlation Coefficient (CC) was used to establish the relationship between the source camera and the reference image. Mathematically, CC shows the relationship between two sets of data by evaluating how the values change with relation to each other. CC returns values between -1 and 1, where 1 indicates a strong positive relationship, 0 indicates no correlation, and -1 indicates a strong negative relationship. The intra-variability and inter-variability was calculated. Log Likelihood Ratio (LLR) was used to compare the likelihood the reference image was taken with the sample camera or the open source camera(s). This calculation was then used to determine if the hypothesis the reference
21


image was taken with the suspect camera either has support, limited support, is inconclusive, or has no
support. A larger LLR number represents a greater likelihood the reference images came from that camera. Result tables were color coded green to represent a correct attribution, or correct elimination. Color coded red represents an incorrect identification called false positive, or incorrect elimination called false rejection. Limited support and inconclusive results were color coded gray.
FUJIFILM GFX 50s
A natural scene reference image (_DSF1055.tif) taken with a FUJIFILM GFX 50S camera was compared against flat field images from the source camera (snEdFr71002897), and a set of natural scene images from an open source camera of the same make and model (snDPRdb71001454). The Gaussian method showed limited support for both cameras. It did not correctly attribute the source camera, nor correctly eliminate the open source camera. The Wiener method was inconclusive for the open source camera, and correctly attributed the reference image to the source camera. The Wiener method performed better than the Gaussian method.
Table 3 FUJIFILM GFX 50s Reference Image from snEdFr71002897 Results
Camera used Gaussian Wiener
FUJIFILM GFX 50S
snDPRdb71001454 Limited Support Inconclusive
snEdFr71002897 source Limited Support Has Support
------------Evidence-------------
Evidence file: _DSF1055.tif
------------Reference------------
2.2268 => 8256x6192-Fuji-GFX-50S-snDPRdb71001454 2.0514 => 8256x6192-Fuji-GFX-50S-snEdFr71002897
Figure 17 Gaussian Method FUJIFILM GFX 50s Log Likelihood Ratio Results
22


-------------Evidence-------------
Evidence file: _DSF1055.tif
-------------Reference------------
920.1153 => 8256x6192-Fuji-GFX-50S-snEdFr71002897 1.793 => 8256x6192-Fuji-GFX-50S-snDPRdb71001454
Figure 18 Wiener Method FUJIFILM GFX 50s Log Likelihood Ratio Results
FUJIFILM GFX 100
A natural scene reference image (_DSF1093.tif) taken with a FUJIFILM GFX 100 camera was compared against flat field images from the source camera (snEGF76A02867), and seven sets of natural scene images from open source cameras of the same make and model. Six open source cameras were utilized; two cameras with RAW files (sn92001309, sn920001255), three cameras with JPEG files (sn92001020, sn92001023, sn92001142), and one camera with both RAW and JPEG files (sn92A01040). The Gaussian method correctly eliminated the cameras with JPEG files, and showed limited support for two cameras with RAW files. It had a false positive for an open source camera with RAW files (sn920001255), and falsely rejected the source camera. The Wiener method correctly eliminated the cameras with JPEG files, and was inconclusive for two cameras with RAW files. It correctly attributed the reference image to the source camera, but had a false positive for an open source camera with RAW files (sn920001255). Both methods correctly eliminated all JPEG images, yet both methods had a false positive. The Wiener method performed better than the Gaussian method.
Table 4 FUJIFILM GFX 100 Reference Image from snEGF76A02867 Results
Camera used FUJIFILM GFX 100 Gaussian Wiener
sn92A01040JPEG No Support No Support
sn92001020 JPEG No Support No Support
sn92001023JPEG No Support No Support
sn92001142 JPEG No Support No Support
23


sn92A01040 RAW Limited Support Inconclusive
sn92001309 RAW Limited Support Inconclusive
sn920001255 RAW Has Support Has Support
snEGF76A02867 source No Support Has Support
------------Evidence-------------
Evidence file: _DSF1093.tif
------------Reference------------
7.923 => 11648x8736-Fuji-GFX-100-snDPRall920001255
2.184 => 11648x8736-Fuji-GFX-100-snDPRall92001309
2.0761 => 11648x8736-Fuji-GFX-100-snDPRall92A01040
0.65465 => 11648x8736-Fuji-GFX-100-snEdFr76A02867
0.0058786 => 11648x8736-FUJIFILM-GFX-100-sn92001023-outdoor-20190524
0.0014498 => 11648x8736-FUJIFILM-GFX-100-sn92001023-outdoor-20190705
0.0013964 => 11648x8736-FUJIFILM-GFX-100-sn92001020-indoor-20190627
0.00040234 => 11648x8736-FUJIFILM-GFX-100-sn92A01040-outdoor-20190531
0.00026888 => 11648x8736-FUJIFILM-GFX-100-sn92A01040-outdoor-20190602
0.00019191 => 11648x8736-FUJIFILM-GFX-100-sn92A01040-outdoor-20190529
4.0999e-05 => 11648x8736-FUJIFILM-GFX-100-sn92001020-outdoor-20190627
Figure 19 Gaussian Method FUJIFILM GFX 100 Log Likelihood Ratio Results
------------Evidence-------------
Evidence file: _DSF1093.tif
------------Reference------------
4.1249 => 11648x8736-Fuji-GFX-100-snEdFr76A02867
3.409 => 11648x8736-Fuji-GFX-100-snDPRall920001255
1.6035 => 11648x8736-Fuji-GFX-100-snDPRall92001309
1.367 => 11648x8736-Fuji-GFX-100-snDPRall92A01040
0.036042 => 11648x8736-FUJIFILM-GFX-100-sn92001023-outdoor-20190705
0.031236 => 11648x8736-FUJIFILM-GFX-100-sn92001020-outdoor-20190627
0.026273 => 11648x8736-FUJIFILM-GFX-100-sn92001023-outdoor-20190524
0.0082169 => 11648x8736-FUJIFILM-GFX-100-sn92A01040-outdoor-20190529
0.0046518 => 11648x8736-FUJIFILM-GFX-100-sn92A01040-outdoor-20190531
0.0033626 => 11648x8736-FUJIFILM-GFX-100-sn92001020-indoor-20190627
0.002312 => 11648x8736-FUJIFILM-GFX-100-sn92A01040-outdoor-20190602
Figure 20 Wiener Method FUJIFILM GFX 50s Log Likelihood Ratio Results
24


Hasselblad X1DII 50c
A natural scene reference image (B7040919.tif) taken with a Hasselblad X1DII 50c camera was compared against flat field images from the source camera (sn565101AB529F), and sets of natural scene images from two open source cameras of the same make and model (sn565101AB5307, sn565101AB530C). The Gaussian method correctly eliminated both open source cameras, and was inconclusive for the source camera. The Wiener method correctly eliminated both open source cameras, and was inconclusive for the source camera. Neither method was able to positively identify the source camera.
Table 5 Hasselblad X1DII 50c Reference Image from snEGF76A02867 Results
Camera used Gaussian Wiener
Hasselblad X1DII 50c
sn565101AB530C No Support No Support
sn565101AB5307 No Support No Support
sn565101AB529F source Inconclusive Inconclusive
-------------Evidence-------------
Evidence file: B7040919.tif -------------Reference------------
1.3958 => 8272x6200-Hassleblad-XlDII-50C-snEdFr565101AB529F 0.3017 => 8272x6200-Hassleblad-XlDII-50C-sn565101AB5307 0.27916 => 8272x6200-Hassleblad-XlDII-50C-sn565101AB530C
Figure 21 Gaussian Method Hasselblad X1DII 50c Log Likelihood Ratio Results
-------------Evidence-------------
Evidence file: B7040919.tif -------------Reference------------
1.3201 => 8272x6200-Hassleblad-XlDII-50C-snEdFr565101AB529F 0.4069 => 8272x6200-Hassleblad-XlDII-50C-sn565101AB530C 0.36728 => 8272x6200-Hassleblad-XlDII-50C-sn565101AB5307
Figure 22 Wiener Method Hasselblad X1DII 50c Log Likelihood Ratio Results
25


Hasselblad H6D 100c
A natural scene reference image (B0000031.tif) taken with a Hasselblad H6D 100c camera was compared against flat field images from the source camera (sn54520225540E), and a set of natural scene images from an open source camera of the same make and model (sn54520225536). The Gaussian method correctly eliminated the open source camera, and correctly attributed the reference image to the source camera. The Wiener method correctly eliminated the open source camera, and correctly attributed the reference image to the source camera. Both methods correctly attributed the source camera and eliminated the open source camera.
Table 6 Hasselblad H6D 100c Reference Image from sn54520225540E Results
Camera used Gaussian Wiener
Hasselblad H6D 100c
sn54520225536 No Support No Support
sn54520225540E source Has Support Has Support
------------Evidence-------------
Evidence file: B0000031.tif ------------Reference------------
6.0045 => 11600x8700-Hassleblad-H6D-100C-snEdFr54520225540E 0.048243 => 11600x8700-Hasselblad-H6D-100c-snEdFr54520225536
Figure 23 Gaussian Method Hasselblad H6D 100c Log Likelihood Ratio Results
------------Evidence-------------
Evidence file: B0000031.tif ------------Reference------------
3.3766 => 11600x8700-Hassleblad-H6D-100C-snEdFr54520225540E 0.026363 => 11600x8700-Hasselblad-H6D-100c-snEdFr54520225536
Figure 24 Wiener Method Hasselblad H6D 100c Log Likelihood Ratio Results
26


Phase One IQ3 100MP
A natural scene reference image (CF008259.tif) taken with a Phase One IQ3 100MP camera was compared against flat field images from the source camera (snHP001078), a set of natural scene images from an open source camera (snlG011142), and both flat field and natural scene sets of images from a camera of the same make and model (snHP001634). The Gaussian method correctly eliminated one of the open source cameras. It correctly eliminated the flat field set of images from the second open source camera, but had limited support for the natural scene set of images from that camera (snHP001634). It correctly attributed the reference image to the source camera. The Wiener method correctly eliminated one of the open source cameras. It correctly eliminated the flat field set of images from the second open source camera, but had a false positive for the natural scene set of images from that camera (snHP001634). It correctly attributed the reference image to the source camera. Both methods correctly identified the source camera, but the Wiener method had a false positive.
Table 7 Phase One IQ3 100MP Reference Image from snHP001078 Results
Camera used Phase One IQ3 100MP Gaussian Wiener
snlG011142 No Support No Support
snHP001634 (flat field) No Support No Support
snFIP001634 (natural scene) Limited Support Flas Support
snFIP001078 source Flas Support Flas Support
------------Evidence------------
Evidence file: CF008259.tif ------------Reference-----------
10.4939 => 11608x8708-PhaseOne-IQ3-100MP-snEdFrHP001078 2.2616 => 11608x8709-PhaseOne-IQ3-100MP-snTEWHP001634out 0.51505 => 11608x8708-PhaseOne-IQ3-100MP-snTEWHP001634 0.0014213 => 11608x8709-PhaseOne-IQ3-100MP-snlRIG011142
Figure 25 Gaussian Method Phase One IQ3 100MP Log Likelihood Ratio Results
27


------------Evidence------------
Evidence file: CF008259.tif ------------Reference-----------
14.0131 => 11608x8709-PhaseOne-IQ3-100MP-snTEWHP001634out 3.0122 => 11608x8708-PhaseOne-IQ3-100MP-snEdFrHP001078
0.3793 => 11608x8708-PhaseOne-IQ3-100MP-snTEWHP001634 0.0028195 => 11608x8709-PhaseOne-IQ3-100MP-snlRIG011142
Figure 26 Wiener Method Phase One IQ3 100MP Log Likelihood Ratio Results
Phase One IQ4 150MP
A natural scene reference image (Capturell829.tif) taken with a Phase One IQ4 150MP camera was compared against flat field images from the source camera (snEdFrJD010673), and a set of natural scene images from an open source camera of the same make and model (snJD020741). Note this 151 megapixels sensor was the largest one studied, which caused some computing memory issues in the processing. The original TIFF conversion files were 14204x10652 pixels and around 840 megapixels file size each. This caused a memory error when conducting the camera comparison tests. Neither TIFF lossless compression nor JPEG compression were able to solve the memory issues. The error may arise from an array size limitation on the version of Matlab used in this study. As a solution, images from this camera were down-sampled to 1601x8700 pixels, which allowed the algorithm to process the images without any further errors. The Gaussian method had a false positive for the open source camera. It correctly attributed the reference image to the source camera. The Wiener method was inconclusive for the open source camera. It correctly attributed the reference image to the source camera. Both methods correctly identified the source camera, but the Gaussian method had a false positive.
Table 8 Phase One IQ4 150MP Reference Image from snJD010673 Results
Camera used Gaussian Wiener Phase One IQ4 150MP
snJD020741 Flas Support Inconclusive
snJD010673 source Flas Support Flas Support
28


-------------Evidence------------
Evidence file: Capturell829.tif -------------Reference------------
4.5279 => 11601x8700 down sampled from 14204xl0652-PhaseOne-IQ4-150-snDPRallJD020741 3.1936 => 11601x8700 down sampled from 14204xl0652-PhaseOne-IQ4-150MP-snEdFrJD010673
Figure 27 Gaussian Method Phase One IQ4 150MP Log Likelihood Ratio Results
------------Evidence-------------
Evidence file: Capturell829.tif ------------Reference------------
4.0585 => 11601x8700 down sampled from 14204xl0652-PhaseOne-IQ4-150MP-snEdFrJD010673 1.3039 => 11601x8700 down sampled from 14204xl0652-PhaseOne-IQ4-150-snDPRallJD020741
Figure 28 Wiener Method Phase One IQ4 150MP Log Likelihood Ratio Results
29


CHAPTER IV
CONCLUSION
This was a preliminary study of medium format sensors and the background Photo Response Non-Uniformity with flat field images. Some expected and some unexpected results arose during the analysis of these sensors. The denoising filters used were the Gaussian and Wiener methods. Overall, the Wiener method performed better than the Gaussian method for camera comparison, but not as well as in other studies. The Gaussian method correctly attributed or eliminated cameras 12 times with 3 false positive over 21 tests (57.2% success rate, 14.3% error rate). The Wiener method correctly attributed or eliminated cameras 14 times with 1 false positive over 21 tests (66.7% success rate, 4.78% error rate). Both methods correctly excluded all JPEG images included in the tests. The Wiener method was more reliable than the Gaussian method, but clearly both methods have some limitations when dealing with medium format sensors, either from the file size or the higher quality of these sensors. Other methods have been developed that may handle camera comparison better. Future research should include possibly a Wavelet methods (3), a Block Matching and 3D (BM3D) filter method (24), and/or an Anisotropic Diffusion (AD) filtering schemes (25). Camera comparison was done with cross correlation, but other methods like Peak to Correlation Energy (PCE) or Maximum Likelihood Estimate (MLE) exist that may prove to be more reliable in future research.
One of the experimental goals of this thesis was working with RAW image data to analyze the full capabilities of these superior sensors. Matlab does not ingest RAW files directly, so they had to be converted into uncompressed TIFF files. An issue with TIFF files is they tend to grow in size when converting from RAW files, growing from 50 MB - 150 MB to 146 MB - 868 MB. This not only makes collecting large sets of images difficult in file management but also working with these files is extremely time consuming. Another limitation in TIFF conversion was the degradation in bit depth from the 14 or
30


16 bits per channel to only 8 bits. This defeats one of the advantages of medium format cameras over previously researched consumer level cameras.
Creating a proper database is essential for source camera comparison. This proved to be a challenge for this type of cameras. Medium format cameras are not consumer level, they tend to range in cost from $5,000 USD to $52,000 USD. The high sticker price means they are mostly sold in specialty camera stores, and places that do have them will tend to only have one model on display, making it difficult to collect multiple source samples. Sample images can be found on open source sites, but RAW files are not typically available. Open source samples are often limited in quantity of images and in the variety of same make and model cameras. Also, no flat field images were found on open source sites for these cameras. For proper comparison, it would be ideal to have a database of multiple cameras of flat field images, with 50 or more images from each camera. One last limitation in creating a comparative database is all the medium format cameras used in this research had their own unique sensor resolution. That is, none of the sensors analyzed had the same number of pixels as any other. This undoubtably introduced some intra-make/model fixed pattern noise which made camera comparison less reliable.
In conclusion, medium format cameras produce unrivaled images in quality, definition, clarity, dynamic range, and color tones. The higher build quality of these massive sensors produces less background noise than consumer cameras. One might think a larger file would have more background noise for camera comparison, but that isn't the case. From a forensic perspective, the higher quality of medium format sensors makes correct attribution and elimination of a source camera more difficult than a cheaper consumer level camera.
31


REFERENCES
1. Rista J. What causes banding noise in CMOS sensors? 2019 [Available from: https://photo.stackexchange.com.
2. Martinec E. Noise, Dynamic Range and Bit Depth in Digital SLRs 2008 [Available from: http://theory.uchicago.edu/~eim/pix/20d/tests/noise/.
3. Fridrich J. Digital Image Forensics Using Sensor Noise
4. United States of America v Nathan Allen Railey. United States District Court: Southern District of Alabama; 2011.
5. Said A, Lukas J, Apostolopoulos JG, Fridrich J, Goljan M. Determining digital image origin using sensor imperfections. Image and Video Communications and Processing 20052005.
6. Sargent JTG, A. Test Report Approval, nonpublished communication to the FBI. Law Enforcement Analysis Facility. 2009.
7. Goljan MF, J.; Filler, T. Large Scale Test of Sensor Fingerprint Camera Identification. Proc SPIE, Media Forensics and Security. 2009;7254.
8. Rosenfeld KS, HusrevTaha. A Study of the Robustness of PRNU-based Camera Identification.
Proc SPIE, Media Forensics and Security. 2009;72540M.
9. M. Chen JF. Determining image origin and integrity using sensor noise. IEEE Transactions on Information Forensics and Security. 2008;3:74-90.
10. Li C. Source camera identification using enhanced sensor pattern noise. IEEE Trans Image Process. 2010;5:280-7.
11. Li C, Li, Y. Color-decoupled photo response non-uniformity for digital image forensics. IEEE Trans Circuits System Video Technology. 2012;22(260-271).
12. Cortiana A, Conotter, Valentina, Boato, G., Natale, F.G.B. Performance comparison of denoising filters for source camera identification. SPIE Electronic Imaging. 2011.
13. Hu Y, Jian, C., Li, C. Source camera identification using large components of sensor pattern noise. Proceedings of Int Conf Computer Science Applications. 2009;23:1-5.
14. Kang X, Jiansheng, C., Kerui, I. A context-adaptive spn predictor for trustworthy source camera identification. EURASIP Journal on Image and Video Processing. 2014;1.
15. Lin X, Li, C.T. Preprocessing reference sensor pattern noise via spectrum equalization. IEEE Transactions on Information Forensics and Security. 2016;11:126-40.
16. Luka J, Fridrich J, Goljan M. Digital Camera Identification From Sensor Pattern Noise. IEEE Transactions on Information Forensics and Security. 2006;1(2):205-14.
32


17. Mehrish A, Subramanyam AM, Emmanuel S. Robust PRNU estimation from probabilistic raw measurements. Signal Processing: Image Communication. 2018;66:30-41.
18. Thing YCaVLL. A study on the photo response non-uniformity noise pattern based image forensics in real-world applications. Institute for Infocomm Research. 2012.
19. Fujifilm-x 2019 [Available from: https://fuiifilm-x.com/en-us/products/cameras/.
20. Hasselblad Medium Format 2019 [Available from: https://www.hasselblad.com/medium-format/.
21. PhaseOne 2019 [Available from: https://www.phaseone.com/en/Photographv/Camera-Technology/Full-Frame-Medium-Format.
22. M. K. Mihcak IK, K. Ramchandran, and P. Moulin. Low-complexity image denoising based on statistical modeling of wavelet coefficients. IEEE Signal Process Lett. 1999(6(12)):300-3.
23. Benjamin Anderson-Sackaney AA-D. Evaluation of Sensor Pattern Noise Estimators for Source Camera Identification. International Journal of Computer and Information Engineering. 2016;10(12).
24. K. Dabov AF, V. Katkovnik, and K. Egiazarian. Image denoising by sparse 3-D transform-domain collaborative filtering. IEEE Transactions on Image Processing. 2007;16(8):2080-95.
25. W.V. Houten ZG. Using Anisotropic Diffusion for Efficient Extraction of Sensor Noise in Camera Identification. Journal of Forensic Sciences. 2012;57:521-7.
33


PDF PDF
APPENDIX
Graphs from camera comparison tests.
FUJIFILM GFX 50s
DSF1055.tifvs 8256x6192-Fuji-GFX-50S-snDPRdb71001454
-0.02 -0.01 0 0.01 0.02 0.03 0.04
CCs
DSF 1055.tif vs 8256x6192-Fuji-GFX-50S-snDPRdb71001454
-0.03 -0.02 -0.01 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07
CCs
34


PDF PDF
dSF1 055.tif vs 8256x6192-Fuji-GFX-50S-snEdFr71002897
0 0.05 0.1 0.15 0.2
CCs
DSF1055.tif vs 8256x6192-Fuji-GFX-50S-snEdFr71002897
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18
CCs
35


PDF PDF
FUJIFILM GFX 100
dSF1 093.tif vs 11648x8736-Fuji-GFX-100-snDPRall920001255
-0 01 -0.005 0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04
CCs
dSF1 093.tif vs 11648x8736-Fuji-GFX-100-snDPRall920001255
-0.02 -0.01 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07
CCs
36


PDF PDF
DSF1093.tif vs 11648x8736-Fuji-GFX-100-snEdFr76A02867
CC RGB=0.059373; logLR=0.65465; Reference images=31
CCs
dSF1 093.tif vs 11648x8736-Fuji-GFX-100-snEdFr76A02867
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16
CCs
37


PDF PDF
SF1093.tif vs 11648x8736-FUJIFILM-GFX-100-sn92A01040-outdoor-20190529
-0.05 0 0.05 0.1 0.15 0.2
CCs
dSF1 093 tif vs 11648x8736-FU JIFILM-GFX-100-sn92A01040-outdoor-20190529
0 0.02 0.04 0.06 0.08 0.1
CCs
38


PDF PDF
. SF1093 tif vs 11648x8736-FU JI FI LM-GFX-100-sn92A01040-outdoor-20190531
0 0.05 0.1 0.15 0.2 0.25
CCs
dSF1 093 tif vs 11648x8736-FUJI FI LM-GFX-100-sn92A01040-outdoor-20190531
0 0.05 0.1 0.15 0.2
CCs
39


PDF PDF
SF1093.tif vs 11648x8736-FUJIFILM-GFX-100-sn92A01040-outdoor-20190602
0 0.05 0.1 0.15 0.2
CCs
dSF1 093 tif vs 11648x8736-FU JIFILM-GFX-100-sn92A01040-outdoor-20190602
CCs
40


PDF PDF
DS F1093.tif vs 11648x8736-FUJIFILM-GFX-100-sn92001020-indoor-20190627
CC RGB=-0.00014346; logLR=0.0013964; #reference images=5
CCs
QS F1093.tif vs 11648x8736-FUJIFILM-GFX-100-sn92001020-indoor-20190627
0 0.02 0.04 0.06 0.08 0.1 0.12
CCs
41


PDF PDF
. SF1093 .tif vs 11648x8736-FUJIFILM-G FX-100-sn92001020-outdoor-20190627
0 0.05 0.1 0.15
CCs
DSF1093 tif vs 11648x8736-FUJIFILM-GFX-100-sn92001020-OUtdoor-20190627
-0.02 0 0.02 0.04 0.06 0.08
CCs
42


PDF PDF
. SF1093 .tif vs 11648x8736-FUJIFILM-G FX-100-sn92001023-outdoor-20190524
-0.05 0 0.05 0.1 0.15
CCs
DSF1093 tif vs 11648x8736-FUJIFILM-GFX-100-sn92001023-OUtdoor-20190524
-0.01 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08
CCs
43


PDF PDF
. SF1093 .tif vs 11648x8736-FUJIFILM-G FX-100-sn92001023-outdoor-20190705
-0.02 0 0.02 0.04 0.06 0.08
CCs
DSF1093 tif vs 11648x8736-FUJIFILM-GFX-100-sn92001023-OUtdoor-20190705
-0.005 0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045
CCs
44


PDF PDF
dSF1 093.tif vs 11648x8736-Fuji-GFX-100-snDPRall92A01040
-0.01 0 0.01 0.02 0.03 0.04 0.05
CCs
dSF1 093.tif vs 11648x8736-Fuji-G FX-100-snDPRall92A01040
-0.01 0 0.01 0.02 0.03 0.04 0.05 0.08
CCs
45


PDF PDF
DSF1093.tif vs 11648x8736-Fuji-GFX-100-snDPRall92001309
-2 0 2 4 6 8 10 12 14 16
CCs -io-3
dSF1 093.tif vs 11648x8736-Fuji-GFX-100-snDPRall92001309
0 0.005 0.01 0.015 0.02 0.025
CCs
46


PDF PDF
Hasselblad X1D II 50c
B7040919.tif vs 8272x6200-Hassleblad-X1 DII-50C-snEdFr CC RGB=0.14935; logLR=1.3958; Reference images=24 The hypothesis that the evidence image was taken with this camera has inconclusive support.
--------Inlra-variabilily {R-R}
--------Inter-variability {R-l)
—““ Evidenoe (R-E)
0 0.05 0.1 0.15 0.2 0.25 0.3
CCs
B7040919.tif vs 8272x6200-Hassleblad-X1 DII-50C-snEdFr
CCs
47


PDF PDF
B7040919.tif vs 8272x6200-Hassleblad-X1 DII-50C-sn565101AB530C
0 0.005 0.01 0.015 0.02 0.025
CCs
B7040919.tif vs 8272x6200-Hassleblad-X1 DII-50C-sn565101AB530C
0 0.005 0.01 0.015 0.02 0.025
CCs
48


B7040919.tif vs 8272x6200-Hassleblad-X1 DII-50C-sn565101 AB5307 CC RGB=0.00068653; logLR=0.3017; #reference images=5
CCs
B7040919.tif vs 8272x6200-Hassleblad-X1 DII-50C-sn565101AB5307
-0.01 0 0.01 0.02 0.03 0.04 0.05 0.06
CCs
49


PDF
Hasselblad H6D 100c
B0000031 tif vs 11600x8700-Hassleblad-H6D-100C-snEdFr54520225540E
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16
CCs
B0000031.tif vs 11600x8700-Hasselblad-H6D-100c-snEdFr54520225536
CCs /W3
50


B0000031.tif vs 11600x8700-Hasselblad-H6D-100c-snEdFr54520225536
-1 01 234567
CCs -10-3
B0000031 tif vs 11600x8700-Hassleblad-H6D-100C-snEdFr54520225540E
0 0.05 0.1 0.15
CCs
51


Phase One IQ3 100 MP
CF008259 tif vs 11608x8709-PhaseOne-IQ3-100MP-snTEWHP001634out
0 0.02 0.04 0.06 0.08 0.1 0.12
CCs
CF008259.tif vs 11608x8708-PhaseOne-IQ3-100MP-snEdFrHP001078 CC RGB=0.14603; logLR=10.4939; Reference images=29 The hypothesis that the evidence image was taken with this camera has support.
T
=n
— Inira-variabillly (R-R)
— Inter-variability (R-l)
— Evidence (R-E)
____I____
0.06
___I___
0.08
CCs
52


CF008259.tif vs 11608x8708-PhaseOne-IQ3-100MP-snEdFrHP001078
LL
Q
Q_
CC RGB=0.12204; logLR=3.0122; Reference images=29 The hypothesis that the evidence image was taken with this camera has support.
— Inira-va riabilily (R-R)
— Inter-variability {R-l)
— Evidence (R-E)
0.08
CCs
CF008259.tif vs 11608x8708-PhaseOne-IQ3-100MP-snTEWHP001634
0 0.1 0.2 0.3 0.4 0.5 0.6
CCs
53


PDF PDF
CF008259 tif vs 11608x8708-PhaseOne-IQ3-1 OOMP-snTEWH PO01634
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
CCs
CF008259.tif vs 11608x8709-PhaseOne-IQ3-100MP-snlRIG011142 CC RGB=-0.00013582; logLR=0.0014213; Reference images=5
54


CF008259.tif vs 11608x8709-PhaseOne-IQ3-100MP-snlRIG011142 CC RGB=8.3688e-05; logLR=0.0028195; Reference images=5
CCs
CF008259 tif vs 11608x8709-PhaseOne-IQ3-100MP-snTEWHP001634out
0 0.01 0.02 0.03 0.04 0.05 0.06
CCs
55


PDF PDF
Phase One IQ4 150 MP
Capture11829.tif vs 11601x8700 downsampled from 14204x10652-PhaseOne-IQ4-150-snDPRallJD020741
0 0.01 0.02 0.03 0.04 0.05
CCs
Capture11829 tif vs 11601x8700 downsampled from 14204x10652-PhaseOne-IQ4-150MP-snEdFrJDOl 0673
CCs
56


PDF PDF
Capture11829 tifvs 11601x8700 downsampled from 14204x10652-PhaseOne-IQ4-150MP-snEdFrJD010673 CC RGB=0.095275; logLR=4.0585; Reference images=20
CCs
Capture11829.tif vs 11601x8700 downsampled from 14204x10652-PhaseOne-IQ4-150-snDPRallJD020741
0 0.005 0.01 0.015 0.02 0.025
CCs
57


Full Text

PAGE 1

PRELIMINARY STUDY ON MEDIUM FORMAT CAMERAS COMPARISON VIA PHOTO RESPONSE NON UNIFORMITY by EDGAR FRITZ B.A., University of California, San Diego, 2001 A thesis submitted to the Faculty of the Graduate School of the University of Colorado in part ial fulfillment of the requirements for the degree of Master of Science Recording Arts Program 201 9

PAGE 2

ii This thesis for the Master of Science d egree by Edgar Fritz h as been approved for the Recording Arts Program by Catalin Grigoras , Chair Je ff M. Smith Carol Golemboski Date: December 1 4 , 2019

PAGE 3

iii Fritz, Edgar ( M.S., Recording Arts Program ) Preliminary Study o n Medium Format Cameras Comparison Via Photo Response Non Uniformity Thesis directed by Associate Professor Catalin Grig oras ABSTRACT This is a preliminary study on Medium Format c ameras. This study utilizes Photo Response Non Uniformity (PRNU), which is part of the background noise inherent in all d igital camera sensors , to compare between Medium Format cameras of the same make and model . PRNU is comprised of tiny imperfections at the pixel level and is sometimes referred to as the camera’s “fingerprint.” This background noise can be extracted and utilized to compare between a reference image and a source camera. Two different methods of PRNU extraction were utilized in thi s study , a G aussian method, and a W iener method. The i nter varia bility and intravariability between a reference image and sets of natural scene images was used to compare cameras. The form and content of this abstract are approved. I recommend its publication. Approved: Catalin Grigoras

PAGE 4

iv To my wife and kids, for always encouraging me to reach my goals.

PAGE 5

v ACKNOWLEDGEMENTS Thank you to my professors at the University of Colorado Denver .

PAGE 6

vi TABLE OF CO NTENTS CHAPTER I. DIGITAL PHOTOGRAPHY BACKGROUND .......................................................................................... 1 Digital Noise................................................................................................................................ 1 Photo Response N on Uniformity...... 3 PRNU and the Daubert Standard for Evidence Admissibility ...................................................... 4 Prior Re search ............................................................................................................................. 5 Medium Format Sensors............................................................................................................. 7 II. CAMERA FINGERPRINT..................................................................................................... 1 0 Materials and Methods ....................................................................................................... ...... 10 PRNU S uitability Test .. .. 11 PRNU Extraction 1 6 III. SOURCE CAMERA COMPARISON ..... 2 1 FUJIFILM GFX 50s .. . 2 2 FUJIFILM GFX 100 ..... 2 3 Hasselblad X1DII 50c ... 2 5 Hasselblad H6D 100c .. . 2 6 Phase One IQ3 100MP ...... 2 7 Phase One IQ4 150 MP . .. 2 8 IV. CONCLUSION ..... 3 0 REFERENCES ......................................................................................................................................... 3 2 APPENDIX . 34

PAGE 7

vii LIST OF FIGURES FIGURE 1 FUJIFILM Sensor Size Comparison .. .... 8 2 Hasselblad Sensor Size Comparison .. . 9 3 P hase One Sensor Size Comparison .. . 9 4 RAW to TIFF C onversion . . 10 5 FUJIFILM GFX 50 s PRNU S uitability T est .. 1 1 6 FUJI FILM GFX 100 PRNU S uitability T est .. 1 2 7 Hasselblad X1DII 50c PRNU S uitability T est ...... 1 3 8 Hasselblad H6D 100c PRNU S uitability T est .. . 1 4 9 Phase One IQ3 100MP PRNU S uitability T est ... 1 5 10 Phase One IQ4 150MP PRNU S uitability T est .. ... 1 6 11 FUJIFILM GFX 50s E xtracted PRNU P atter n ....... 1 7 12 FUJIFILM GFX 100 Extracted PRNU Pattern .. 1 8 13 Hasselblad X1DII 50c Extracted PRNU Pattern ..... 1 8 14 Hasselblad H6D 100c Extracted PRNU Pattern .. 19 15 Phase One IQ3 100MP Extracted PRNU Pattern 19 1 6 Phase One IQ4 150MP Extracted PRNU Pattern . .. 2 0 17 Gaussian Method FUJIFILM GFX 50s Log Likelihood Ratio Results .. ... 2 2 18 Wiener Method FUJIFILM GFX 50s Log Likelihood Ratio Results .. ... 2 3 19 Gaussian Method FUJIFILM GFX 100 Log Likelihood Ratio Results .. 2 4 20 Wiener Method FUJIFILM GFX 50s Log Likelihood R atio Results .. ... 2 4 21 Gaussian Method Hasselblad X1DII 50c Log Likelihood Ratio Results ...... 2 5 22 Wiener Method Hasselblad X1DII 50c Log Likelihood Rati o Results .. ...... 2 5

PAGE 8

viii 23 Gaussian Method Hasselblad H6D 100c Log Likelihood Ratio Results ... .. 2 6 24 Wiener Method Hasselblad H6D 100c Log Likelihood Ratio Res ults ..... 2 6 25 Gaussian Method Phase One IQ3 100MP Log Likelihood Ratio Results .. .... .. 2 7 26 W iener Method Phase One IQ3 100MP Log Likelihood Ratio Results ... 2 8 27 Gaussian Method Phase One IQ4 150MP Log Likelihood Ratio Results 29 28 Wiener Method P hase One IQ4 150MP Log Likelihood Ratio Results . .. 29

PAGE 9

ix LIST OF TABLES TABLE 1 Cross Sample of Typical Cameras and Specifications .... 6 2 Medium Format Sensors Included in this Study and their Specifications . 8 3 FUJIFILM GFX 50s R eference I mage from snEdFr71002897 R esults . 2 2 4 FUJIFILM GFX 100 Reference I mage from snEGF76A02867 R esults . ... 2 3 5 Hasselblad X1DII 50c R eference I mage from snEGF76A02867 Results . . 2 5 6 Hasselblad H6D 100c R eference I mage from sn54520225540E Results . .. 2 6 7 Phase One IQ3 100MP Reference I mage from snHP001078 R esults .. . 2 7 8 Phase One IQ4 150MP Reference I mage from snJD010 673 Results .. .. 2 8

PAGE 10

1 CHAPTER I DIGITAL PHOTOGRAPHY BACKGROUND Modern digital cameras are ubiquitous in our society today. The digital revolution transformed photography forever. Digital p hotography eliminated the wait of developing and print ing image s before confirming if you even capture d the shot. It brought users instant feedback and the ability to review images wherever they wer e. But d igital cameras brought much more than that , they include a wealth of information never before captured in the film era. Digital images contain data embedded within the file . This is called the metadata , and it contains valuable information tha t can be invaluable from a forensic standpoint . Embedded within the file you may find info rmation about when the image was taken, the make, model, and serial number of the camera , geolocation, camera setting s, lens information, and even if an image has been edited after it was captured. Information not only can be obtained from the file’s metadata, but the images themselves can be analyzed and infere nces can be made from them . Digital images can be used forensically to identify locations from elements within the image , dates and times from shadows , subject identification from biometrics, size of objects from known objects comparison , and many more . You can even ascertain which camera took an image from the inherent background noise of a digital sensor. Today, you’d be hard presse d to find someone without some sort of image capturing device on them at all times . Digital images can be the one piece of corroborating evidence that ties to gether an entire court case. They are a big part of today’s criminal justice system, that is why forensic scientists have been studying their properties since the dawn of the digital age. Digital Noise A digital camera works by allowing light to enter though an opening in an optical lens (aperture) for a specified length of time (shutter s peed) . The photons of light that enter are focus ed on a digital sensor which is made up of an array of pixels . Each pixel is made of a photosensitive material responsible for gathering photons of a specific wavelength of light. Photons captured by the

PAGE 11

2 photosensitive part of a pixel stimulate the emission of electrons, one for each captured photo. T he accumulated photoelectrons are counted and proportionally co nverted to a voltage. That number can then be amplified depending on t he sensitivity to light set on the camera (ISO setting ) . This voltage is then digitized in an analog to digital converter (ADC) and converted into 1s and 0s to be written out to a file . This binary code can be thought of as the RAW image data. Additional information will be written to the file by the camera as metadata for future decoding of the image. The process described a bove would be an ideal camera with ideal components and processes. But no camera can perform to the ideal, hence no image is an exact representation of the photons that struck the sensor. A ll digital images will possess some inherent noise , which comes f rom several sources during the image acquisition process. Noise can be thought of as variations in color or intensity from what the true scene was . This noise can be grouped into two general types: Random N oise, and F ixed P attern N oise. Each electronic circuit in the signal processing chain suffers from random fluctuations . These fluctuations can come from the photosensitive components of the sensor , the ISO gain, and /or the ADC digitization. Horizontal and Vertical Banding Noise (HVBN) is an example of r andom noise in the image capturing process . HVBN manifests itself as linear color and intensity f luctuations in the rows or columns and look like bright bands in the image. HVBN noise can be caused by errors in the sensor readout ; that is when rows of pixels are activated and readout down the pipeline . ISO settings will also introduce HVBN as the ISO is increased . ISO is a scalar multiplier of the photoelectron count, so higher settings will amplify errors . ADC digitization can introduce HVBN noise with errors in the digital conv ersion process ( 1 ) . There are several other sources of random noise that can affect a digital image. Fixed Patter n Noise is a pattern that will be stamped i n all images from the same sensor under the same conditions. There are many factors that will contribute to Fixed Pattern Noise at different stages of the image capturing process. Thermal agitation of electr ons in the photoreceptors of a sensor

PAGE 12

3 can l iberate some electrons. These freed electrons will be counted as if they were photons that were present in the image and will be added to the RAW data. Thermal electrons are freed at a relatively constant rate c reating constant imperfections in an image . Thermal Noise will increase with exposure time and will show up as “hot pixels” in an image . These hot pixels are overexposed pixels that will show up in the same spot repeatedly . Pixel Non Uniformity is a g eneral term for the imperfections in the mi cron sized pixels of a sensor , it is comprised of two types. Dark Signal Non Uniformity (DSNU) is the offset from the average across the sensor at a particular temperature and integration time without external illumination. Basically, this noise is found when an image is taken in complete darkness. DSNU will be fixed in all images taken in total darkness. Pixel Non Uniformity is also found in illuminated pixels. Pixels have acceptable levels of variation in c apturing efficiency , photon counting accuracy; variations in pixel size, shape, and substrate material. Not all pixels in a sensor are made the same, these tiny variations are called the Photo Response Non Uniformity (PRNU). PRNU comes from the gain or ratio between optical power on a pixel vers us the electrical signal output. When the photoreceptive cell s in a sensor are illuminated evenly, each cell should ideally output the same voltage, but they don’t . PRNU will be ingrained in all images as part of the background noise and will be an ever present fixed pattern. That is why PRNU is often referred to as the camera’s “fingerprint.” PRNU noise grows in direct proportion to the exposure level, so bri ghter pixels will show more PRNU noise. PR NU noise will be reproduced for all images taken under the same conditions and settings ( 2 ) . Photo Response Non Uniformity PRNU can be used for several forensic methods . One example is to test for the presence of a specific PRNU in an image, wh i ch c an achieve reliable device identification ( prove that a certain camera took a given image) . PRNU can a lso help prove two images were taken by the same device (device linking). The presence of PRNU is indicative of the fact that the image is natural and not a computer

PAGE 13

4 rendering. By establishing the absence of the PRNU in individual image regions, it is possible to discover maliciously forge ry attempts . By detecting the strength or fo rm of the PRNU , it is possible to reconstruct some of the processing that was done to an image. PRNU is used as a template to estimate geometrical processing, such as scaling, cropping, or rotation. Non geometrical operations are also going to influence the strength of the PRNU in the image and thus can be potentially detected. The spectral and spati al characteristics of the PRNU can be used to identify the camera model or distinguish between a scan and a digital camera image (the scan will exhibit spatial anisotropy) ( 3 ) . PRNU and the Daube rt Standard for Evidence Admissibility PRNU has proven to be extremely valuable as a forensic tool. During the court proceedings of the United States of America v. Nathan Allen Railey , United States Court of Appeals for the Southern District of A labama, p rosecutors showed child pornography images in question were created by the defendant ’s camera by comparing the PRNU . This identification was challenged in an admissibility hearing to verify the technology met the Daubert s tandard . It met the challenge as follows ( 4 ) : Whether the theory or technique employed is generally accepted in the scientific community: PRNU source camera identification was first introduced in 2005 by Lukas, Fridrich, and Goljan, where they detailed how to determine a digital image ’s origin using sensor pattern noise ( 5 ) . Numerous papers have subsequently been written expanding on the se methods and techniques. Whether the technique has been subjected to peer review or publications : There have been numerous peer reviewed publications on the subject appearing in such periodicals as IEEE, SPIE, the International Conference on Image Processing, EURASIP Journal on Image and Video Processing, International Conference on Com puter Science and its Applications, etc. Whether the technique has been tested or can be tested: There exist many techniques for PRNU extraction and co mparison. These methods and algorithms are tested and peer reviewed as they

PAGE 14

5 are developed. Validation is a typical part of the scientific process, and some validation has been conducted by the Law Enforcement Analysis Facility , an independent testing company in NY ( 6 ) . Whether an error rate can be determined: There have been several larg e scale tests conducted to test identification error and false positive rates . One such study was conducted by Miroslav Goljan on more than a million images. It tested PRNU image identification via peak to correlation energy, with a false positive rate at less than .000 2 4 % , and the false rejection rate at less than .0238 % ( 7 ) . PRNU met the Daubert standard in this court ca se, and expert testimony regarding PRNU implementation was admitted into court. US v Railey provided support for future PRNU implementation against Daubert challenges since the theory or technique employed met the standards set and is generally accepted in the scientific community as a valid method . Prior Research Camera identification has been studied extensively for decades now . Researchers have found several ways of matching images with the camera that produced them. A variety of features have been employed, including color filter array design, pixel color value interpolation algorithm, image sensor anomalies, lens characteristics and anomalies, and image processing pipeline characteristics ( 8 ) . PRNU extraction for camera comparison has also been widely studied . There have been various attempts to improve the accuracy of digital camera identification. Several approaches aim to achieve this by improv ing the estimate of PRNU. Some apply Maximum Likelihood Estimation (MLE) for estimation of a multiplicative factor from reference images ( 9 ) . Other research suppresses unwan ted artifacts from sensor pattern noise . This is achieved by preprocessing the PRNU by subtracting Zero Mean and applying Wiener filtering ( 10) . Some research proposes extracting PRNU per c olor channel in a Color Decoupled PRNU extraction method ( 11 ) . The effect of denoising filter has been investigated using sparse 3 D transform domain collaborative filtering for mo re accurate noise extraction ( 12 ) . Another documented method uses edge adaptive PRNU predictor context adaptive interpolation ( 13 ,

PAGE 15

6 14) . Ano ther technique is to pre process the estimated PRNU for removal of non unique artifacts. These techniques work by identifying and suppressing the peaks according to local characteristi cs in the magnitu de spectrum of the reference PRNU with spectrum equalization, improving the detection accuracy further ( 15 ) . Furthermore, other methods improve per formance by using better statistical detection. Some use peak to correlation energy to attenuate hidden periodic patterns ( 16) . Others use correlation over circular cross correlation norm to further decrease the fals e positive rate ( 14 ) . Though there are several methods and techniques that explore source camera identification via PRNU comparison , few use RAW image data to achieve this. Those that do u se RAW data only focus on either identification of source device of the same model, or different devices of the same model but not different models or cameras ( 17 ) . If yo u browse through PRNU publications , you will find researche r s typically utilize JPEG compressed images from mass manufac tured consumer level cameras . The sensor quality and size used in typical research overlooks the best available sensors in the market today . One paper tried to specifically address this issue but only obtain ed data from “mid end and high end popular consumer digital cameras” ( 18) , not take into consider ation professional level medium format cameras. Below is a cross sample of typica l cameras utilized in PRNU literature : Table 1 Cross Sample of Typical Cameras and Specifications Make Model Sensor Type Sensor Size Sensor Resolution Pixel Area Mega Pixels File Format Nokia 6600 Cell Phone Not Provided 6 40 x 480 Not Provided .3 MP JPEG 8 bit Canon PowerShot A430 CCD 4.8 x 3.6 mm 2272 x 1704 4.33 m 2 4.1 MP JPEG 8 bit FUJIFILM FinePix J5 0 CCD 5.75 x 4.32 mm 3264 x 2448 3.03 m 2 8.2 MP JPEG 8 bit Olympus MJU 7050 CCD 6.16 x 4.62 mm 4315 x 3244 2.04 m 2 1 4 MP JPEG 8 bit Casio EX Z150 CCD 5.75 x 4.32 mm 3264 x 2448 2.99 m 2 8.29 MP JPEG 8 bit Canon EOS 60D CMOS 22.3 x 14.9 mm 5196 x 3464 18.4 m 2 18 MP RAW 14 bit C2R

PAGE 16

7 Medium Format Sensors We have discussed how PRNU has been studied by numerous people wit h various methods and under various conditions for accuracy, and robustness , but no research has been conducted on the top quality sensors in the market . The objective of this thesis is to look beyond consumer level cam eras and study professional level medium format cameras. Medium format refers to any sensor larger than a full frame DSLR (or the traditional 35mm sensor at 24 x 36mm). The advantage of medium format sensors comes not just from the ir s ize, but from the larger pixel area, and meticulously calibrated sensors. Medium format sensors provide a wide r dynamic color range , larger field of view, and shallower depth of field. Larger pixels have more light gathering capability . Finer and more detailed tonal information can be produced, delivering more information in the final file. But more importantly for camera comparison, better build, quality control, and smaller production scale. This thesis focus es specifically on tod ay’s top end medium format camera manufacturers, FUJIFILM, Hasselblad, and Phase One. The following cameras were used in this research as being recent top end cameras from said manufacturers : FUJIFILM GFX 50S with a 50MP sensor , FUJIFILM GFX 100 with a 10 0MP sensor ; Hasselblad H6D 100c camera with a 100MP sensor, Hasselblad X1D II50 c mirrorless camera with a 50MP sensor; Phase One IQ3 with a 10 1 MP sensor, Phase One IQ4 with a 15 1 MP sensor. Medium format sensors provide much larger images with better resolution and more mega pixels than any sensor found in prior studies into PRNU. The Pixel Area (size of each pixel) is also much larger allowing for greater light gathering capability. And the RAW recording allo ws for 14 an d even 16 bits of depth yielding millions of more colors , that is 1 . 6 x 107 for an 8 bit image vs. 2.8 x 1014 for a 16 bit image. The following table shows the specifications for the cameras used in this study.

PAGE 17

8 Table 2 Medium Format Se nsors Included in this Study and their Specifications Make Model Sensor Type Sensor Size Sensor Resolution Pixel Area Mega Pixels File Format FUJIFILM GFX 50S CMOS 43.8mm x 32.9mm 8256 x 6192 28.2 m 2 51.4 MP RAW 14 bit RAF FUJIFILM GFX 1 00 CMOS BSI 43.8mm x 32.9mm 11648 x 8736 14.14 m 2 101.9 MP RAW 16 bit RAF Hasselblad X1DII 50c CMOS 43.8 x 32.9mm 8272 x 6200 27.98 m 2 51.3 MP RAW 16 bit 3FR Hasselblad H6D 100c CMOS 53.4 x 40.0 mm 11600 x 8700 21.16 m 2 100 MP RAW 16 bit 3FR Phase One IQ3 100MP CMOS 53.7 x 40.4mm 11608 x 8708 21.44 m 2 101.1 MP RAW 16 bit IIQ Phase One IQ4 150MP CMOS BSI 53.4 x 40.0mm 14204 x 10652 14.14 m 2 151 MP RAW 16 bit IIQ Here is what the manufacturers have to say about their cameras. The FUJIFILM GFX 50S features a 43.832.9mm CMOS medium format sensor, boasting an effective resolution of 51.4 million pixels. The FUJIFILM GFX 100 pairs a newly developed back illuminated 1 02MP imaging sensor with a blazingly fast X Pro cessor 4 processing engine to create a combination capable of outputting 16 bit images with amazing color fidelity, rich shadow detail, and incredible dynamic range. Its back illuminated structure enhances image quality by bringing the exposure plane close r to the color filter array. Sharpness is also enhanced, while moir and false colors are near eliminated through the omission of an optical low pass filter ( 19) . Figure 1 FUJIFILM Sensor Size Comparison

PAGE 18

9 Hasselblad camera sensors are individually calibrated to ensure maximum performance in any given situation. A massive amount of data is gathered for each unit in the production stage in order to study the variations that can occur and see how t he sensor performs under different circumstances. Through Hasselblad’s rigorous sensor testing, all irregularities are corrected automatically when shooting with any Hasselblad camera ( 20) . Figure 2 Hasselblad Sensor Size Comparison The full frame medium format sensors found in the Phase One IQ Digital Backs are 2.5x larger than that of full frame DSLR cameras, and 1.5x larger than cropped sensor mirrorless medium format. The use of a la rger sensor means more light, and therefore more information is available when converting captured light into a digital file. The availability of more data allows Phase One to honor the quality of that data, translating to more efficient use of information and less interpolation ( 21 ) . Figure 3 Phase One Sensor Size Comparison

PAGE 19

10 CHAPTER II CAMERA FINGERPRINT Materials a nd Methods Flat f ield (evenly illuminated neutral wall) images were collected from each camera. The images were slightly underexposed, so no pixels maxed out their threshold. A few natural scene images were also collected for comparative analysis. In order to obtain the full quality of each sensor, RAW images were collected. RAW images would avoid the undesired addition of JPEG artifacts into the background noise . In camera noise reduction was not appli ed to any of the images collected to maintain proper PRNU signatures . ISO 100 was used for all images collected to limit random noise. All calculations were performed on a Windows 10, 64 bit operating system with Intel Core 17 8550U CPU at 2.0 GHz and 16.0 GB of RAM. Matlab R2019b version 9.7.0 build 1190202 was utilized for all computations. RAW files were converted into u ncompressed TIFF fi les in Adobe Bridge CC 2019 v9.1.0.338 with Adobe Photoshop CC 2019 v 20.0.7 for the Hasselblad and FUJIFILM cameras. The Image Processor tool convert ed folders of RAW images into uncompressed TIFF files. Note this process does not provide bit depth options and aut omatically converted the images into 8 bits per channel. Phase One images were processed using C apture One Pro v12.1.1.19 since Adobe does not support IIQ file format . The Process Recipe tool was used in Capture One to convert RAW images into uncompresse d TIFF files. This process does allow for 16 bit TIFF conversion, but files were converted to 8 bits per channel for consistency. Figure 4 RAW to TIFF C onversion

PAGE 20

11 PRNU Suitability Test Once the images were converted to TIFF files, they were tested for PRNU suitability with a denoising filter prior to PRNU extraction. The Wavelet meth od was uses to denoise the images. This is a spatial adaptive statistical model for wavelet image coefficients as zero mean Gaussian random variables with high local correlation. This method assumes a marginal prior distribution on wavelet coefficients v ariances and estimates them using an approximate maximum a posteriori probability rule. Then an approxim ate minimum mean squared error estimation procedure was applied to restore the noisy wavelet image coefficients ( 22 ) . Simply put, the wavelet method analyses data at a local level as well as at a global level. The result tells us if the images in question are suitable for PRNU extraction. The flat field reference images collected were tested with a Wiener method for suitability of PRNU extraction. The following figures show each set of images and the cross correlation of the intra/intervariability. Th is process plots the first image in the set as the evidence and compares the inter variability ( the variation within each file), and the intra variability ( the varia tion between all the files), and calculates the cross correlation amongst them for PRNU suitability. N ote some reference flat field image sets were reduced to 20 images for computational considerations . All reference flat field i mage sets were found to be suitable for PRNU analysis. Figure 5 FUJIFILM GFX 50 s PRNU S uitability T est

PAGE 21

12 Figure 5 Continued Figure 6 FUJIFILM GFX 100 PRNU S uitability T est Comput es CC1, CC2... ans = 0.0011 0.6982 0.0004 0.7434 20 images, very good for PRNU analysis, 100*k2=0.10807 < k1=0.69817 ------------------------------------------------------------------------------8256x6192Fuji GFX 50S snEdFr71002897 Comput es CC1, CC2... ans = 0.0005 0.7289 0.0003 0.7695 20 images, very good for PRNU analysis, 100*k2= 0.054285 < k1=0.72888 ------------------------------------------------------------------------------11648x8736Fuji GFX 100 snEdFr76A02867

PAGE 22

13 Figure 7 Hasselblad X1DII 50c PRNU S uitability T est Com putes CC1, CC2... ans = 0.0007 0.7781 0.0011 0.8505 24 images, very good for PRNU analysis, 100*k2=0.067744 < k1=0.77806 ------------------------------------------------------------------------------8272x6200Hassleblad X1DII 50C snEdFr565101AB529F

PAGE 23

14 Figure 8 Hasselblad H6D 100 c PRNU S uitability T est Computes CC1, CC2... ans = 0.0008 0.7376 0.0008 0.7840 30 images, very good for PRNU analysis, 100*k2= 0.075368 < k1=0.73763 ------------------------------------------------------------------------------11600x8700Hassleblad H6D 1 00C snEdFr54520225540E

PAGE 24

15 Figure 9 Phase One IQ3 100MP PRNU S uitability T est Computes CC1, CC2... ans = 0.0001 0.7378 0.0002 0.7824 20 images, very good for PRNU analysis, 100*k2=0.0072872 < k1=0.73777 ------------------------------------------------------------------------------11608x8708Ph aseOne IQ3 100MP snEdFrHP001078

PAGE 25

16 Figure 10 Phase One IQ4 150 MP PRNU S uitability T est PRNU Extraction Once confiremed the flat field images were suitable , the PRNU was extracted and displayed . This was achieve d by taking the noise residuals of the reference flat field images and conver t them to a grayscale signal. First, the fingerprint is estimated in each of the three color channel s (red, blue, green). Then, the channels are mean is subtracted and average s of each column and each row i n each of four sub sampled 2 D signals Computes CC1, CC2... ans = 0.0007 0.7781 0.0011 0.8505 24 images, very good for PRNU analysis, 100*k2=0.067744 < k1=0.77806 ------------------------------------------------------------------------------8272x 6200 Hassleblad X1DII 50C snEdFr565101AB529F

PAGE 26

17 corresponding to four types of pixels in the Bayer CFA are subtracted from all elements. This “Zero Mean” procedure removes a large portion of Non Unique Artifacts introduced by demosaicking (that is noise that can b e found in cameras from the same make and model, making them not unique to the sample camera) . To remove any residual artifacts from the estimated fingerprint, the fingerprint was finally filtered using an adaptive Wiener filter in the frequency domain ( 7 ) . T he following images are the extracted PRNU noise from the reference flat field image set . Figure 11 FUJIFILM GFX 50s Extracted PRNU Pattern

PAGE 27

18 Figure 12 FUJIFILM GFX 100 Extracted PRNU Pattern Figure 1 3 Hasselblad X1DII 50c Extracted PRNU Pattern

PAGE 28

19 Figure 14 Hasselblad H6D 100c Extracted PRNU Pattern Figure 15 Phase One IQ3 100MP Extracted PRNU Pattern

PAGE 29

20 Figure 16 Phase O ne IQ4 150MP Extracted PRNU Pattern Note the subtle outline still visible in Figure 16. This was due to the flat field images containing part of some furniture that was overlooked when colleting the flat field images.

PAGE 30

21 CH APTER III SOURCE CAMERA COMPARISON The extracted PRNU was then used to compare a natural scene reference image taken with the reference flat field cameras , and a small open source image database created fo r this research . PRNU noise was obtained by suppressin g scene content from an image with a denoising filter and then averaging the noise residuals. In this study, Gaussian denoising filter, and W iener denoising filter were used to suppress scene images . The Gaussian filt er removes noise by blurring an image. It achieves this by distributing pixels in Gaussian (bell shaped) distribution across the entire image. The Wiener filter both recovers an image from blurring of a lowpass filter (inverse filtering) and reduces noise by smoothing at the same time, i t removes additive noise and inverts the blurring simultaneously ( 23 ) . Two methods were used to study the accuracy and reliability in comparing between diff erent cameras of the same make and model. Note the appendix will contain figures from all camera comparisons. In order to conduct this analysis, open source RAW images and some JPEG images were obtained from www.dp review.com, www.photographyblog.com , www.flickr.com , www.phaseone.com , www.hasselblad.com , and www. FUJIFILM .com , to build a small database of images. All open source RAW images were converted to TIFF files for comparison. The open source images collected were of natural scenes with normal dynamic range and exposure. Correlation Coefficient (CC) was used to establish the relationship between the source camera and the reference image. Mathematically, CC shows the relationship between two sets of data by evaluating how the values change with relatio n to each other. CC returns values between 1 and 1, where 1 indicates a strong positive relations hip, 0 indicates no correlation, and 1 indicates a strong negative relationship. The intravariability and inter variability was calculated. Log Lik elihood Ratio ( L L R ) was used to compare the likelihood the reference image was taken with the sample camera or the open sourc e camera(s) . This calculation was then used to determine if the hypothesis the reference

PAGE 31

22 image was taken with the suspect camera either has support, limited support, is inconclusive, or has no support. A larger LL R number represents a greater likelihood the reference images came from th at camera. Result tables were color coded green to represent a correct attribution, or correct elimination . Color coded red represents an incorrect identification called false positive , or inc orrect elimination called false rejection . L imited support and inconclusive results were color coded gray. FUJIFILM GFX 50s A natural scene reference image ( _DSF1055.tif) taken with a FUJIFILM GFX 50S camera was compared against flat field images from the source camera ( snEdFr71002897) , and a set of natural scene images from an open source camera of the same make and model ( snDPRdb71001454) . The Gaussian method showed li mited support for both cameras . It did not correctly attribute the source camera , nor correctly eliminate the open source camera. The Wiener method was inconclusive for the open source camera, and corre ctly attributed the reference image to the source camera . The Wiener method performed better than the Gaussian method. Table 3 FUJIFILM GFX 50s R eference I mage from snEdFr71002897 R esults Camera used FUJIFILM GFX 50 S Gaussian Wiener snDPRdb71001454 Limited Support Inconclusive snEdFr71002897 source Limited Support Has Support Figure 17 Gaussian Method FUJIFI LM GFX 50s Log Likelihood Ratio R esults ------------------Evidence ------------------Evidence file: _DSF1055.tif -------------------Reference-----------------2.2268 => 8256x6192Fuji GFX 50S snDPRdb71001454 2.0514 => 8256x6192 Fuji GFX 50S snEdFr 71002897

PAGE 32

23 Figure 18 Wiener Method FUJIFILM GFX 50s Log Likelihood Ratio R esults FUJIFILM GFX 100 A natural scene reference image ( _DSF1093.tif) taken with a FUJIFILM GFX 100 camera was compared against flat field images from the source camera (snEGF76A02867) , and sev en sets of natural scene images from open source cameras of the same make and model . Six open source cameras were utilized ; t wo cameras with RAW files (sn92001309, sn920001255) , three cameras with JPEG files (sn92001020, sn92001023, sn92001142) , and one camera with b oth RAW and JPEG files (sn92A01040) . The Gaussian method correctly eliminated the cameras with JPEG files , and showed limited support for two cameras with RAW files . It had a false positive for a n open source camera with RAW files (sn920001255) , and falsely rejected the source camera . The Wiener method correctly eliminated the cam eras with JPEG files , a nd was inconclusive for t wo cameras with RAW files . It correctly attributed the reference image to the source camera , but had a false positive for an open source camera with RAW files (sn920001255) . Both methods co rrectly eliminated all JPEG images , yet both methods had a false positive . The Wiener method performed better than the Gaussian method. Table 4 FUJIFILM GFX 100 R eference I mage from snEGF76A02867 R esults Camera used FUJIFILM GFX 100 Gaussian Wien er sn92A01040 JPEG No Support No Support sn9200 1020 JPEG No Support No Support sn9200 1023 JPEG No Support No Support sn9200 1142 JPEG No Support No Support ------------------Evidence ------------------Evidence file: _DSF1055.tif -------------------Reference-----------------920.1153 => 8256x6192Fuji GFX 50S snEdFr71002897 1.793 => 8256x6192 Fuji GFX 50S snDPRdb71001454

PAGE 33

24 sn92A0 1040 RAW Limited Support Inconclusive sn92001309 RAW Limited Support Inconclusive sn9200 0 1255 RAW Has Support Has Support snEGF76A02867 source No Support Has Support Figure 19 Gaussian Method FUJIFILM GFX 100 Log Likelihood Ratio R esults Figure 20 Wiener Method FUJIFILM GFX 50s Log Likelihood Ratio R esults ------------------Eviden ce ------------------Evidence file: _DSF1093.tif -------------------Reference-----------------7.923 => 11648x8736Fuji GFX 100 snDPRall920001255 2.184 => 11648x8736Fuji GFX 100 snDPRall92001309 2.0761 => 11648x8736 Fuji GFX 100 snDPRall92A01040 0.65465 => 11648x8736Fuji GFX 100 snEdFr76A02867 0.0058786 => 11648x8736FUJIFILM GFX 100 sn92001023 outdoor 20190524 0.0014498 => 11648x8736FUJIFILM GFX 100 sn92 001023 outdoor 20190705 0.0013964 => 11648x8736FUJIFILM GFX 100 sn92001020 indoor 20190627 0.00040 234 => 11648x8736 FUJIFILM GFX 100 sn92A01040 outdoor 20190531 0.00026888 => 11648x8736 FUJIFILM GFX 100 sn92A01040 outdoor 20190602 0.00019191 => 11648x8736 FUJIFILM GFX 100 sn92A01040 outdoor 20190529 4.0999e 05 => 11648x8736 FUJIFILM GFX 100 sn92001020 outdoor 20190627 ------------------Evidence ------------------Evidence file: _DSF1093.tif -------------------Reference-----------------4.1249 => 11648x8736 Fuji GFX 100 snEdFr76A02867 3.409 => 11648x8736Fuji GFX 100 snDPRall920001255 1.6035 => 11648x 8736 Fuji GFX 100 snDPRall92001309 1.367 => 11648x8736Fuji GFX 100 snDPRall92A01040 0.036042 => 11648x8736 FUJIFILM GFX 100 sn92001023 outdoor 20190705 0.031236 => 11648x8736 FUJIFILM GFX 100 sn92001020 outdoor 20190627 0.026273 => 11648x8736 FUJIFILM GFX 100 sn92001023 ou tdoor 20190524 0.0082169 => 11648x8736FUJIFILM GFX 100 sn92A01040 outdoor 20190529 0.0046518 => 11648x8736FUJIFILM GFX 100 sn92A01040 outdoor 20190531 0.0033626 => 11648x8736FUJIFILM GFX 100 sn92001020 indoor 20190627 0.002312 => 11648 x8736 FUJIFILM GFX 100 sn92A01040 outdoor 20190602

PAGE 34

25 Hasselblad X1DII 50c A natural scene reference image ( B7040919 .tif) taken with a Hasselblad X1DII 50 c camera was compared against flat field images from the source camera ( sn565101AB529F ) , and sets of natural scene images from two open source cameras of the same make and model ( sn565101AB5307, sn565101AB530C ) . The Gaussian method correctly eliminated both open source cameras , and was inconclusive for the source camera. The Wiener method correctly eliminated both open source cameras , and was inconclusive for the source camera. Neither method was able to positively identify the source camera. Table 5 Hasselblad X1DII 50c R eference I mag e from snEGF76A02867 R esults Camera used Hasselblad X1DII 50c Gaussian Wiener sn565101AB530C No Support No Support sn565101AB5307 No Support No Support sn565101AB529F source Inconclusive Inconclusive Figure 21 Gaussian Method Hasse lb lad X1DII 50c Log Likelihood Ratio R esults Figure 22 Wiener Method Hasselblad X1DII 50c Log Likelihood Ratio R esults ------------------Evidence ------------------Evidence file: B7040919.tif -------------------Reference-----------------1.3958 => 8272x6200Hassleblad X1DII 50C snEdFr565101AB529F 0.3017 => 8272x6200Ha ssleblad X1DII 50C sn565101AB5307 0.27916 => 8272x6200 Hassleblad X1DII 50C sn565101AB530C ------------------Evidence ------------------Evidence file: B7040919.tif -------------------Reference-----------------1.3201 => 8272x6200Hassleblad X1DII 50C snEdFr565101AB529F 0.4069 => 8272x6200Hassleblad X1DII 50C sn565101AB530C 0.36728 => 8272x6200 Hassleblad X1DII 50C sn565101AB5307

PAGE 35

26 Hasselblad H6D 100c A natural scene reference image ( B0000031.tif ) taken with a Hasselblad H6D 100c camera was compared against flat field images from the source camera ( sn 54520225540E ), and a set of natural scene images from an open source camera of the same make and model ( sn 54520225536 ) . The Gaussian method correctly eliminated the open source camera, and correctly attributed the reference image to the source camera . The Wiener method co rrectly eliminated the open source camera, and correctly attributed the reference image to the source camera . Both methods correctly attributed the source camera and eliminated the open source camera. Table 6 Hasselblad H6D 100c R eference I mage from sn54520225540E R esults Camera used Hasselblad H6D 100c Gaussian Wiener sn 54520225536 No Support No Support sn 54520225540 E source Has Support Has Support Figure 23 Gaussian Method Hasselblad H6D 100c Log Likelihood Ratio R esults Figure 24 Wiener Method Hasselblad H6D 100c Log Likelihood Ratio R esults ------------------Evidence ------------------Evidence file: B0000031.tif -------------------Reference-----------------6. 0045 => 11600x8700 Hassleblad H6D 100C snEdFr54520225540E 0.048243 => 11600x8700 Hasselblad H6D 100c snEdFr54520225536 ------------------Evidence ------------------Evidence file: B0000031.tif -------------------Reference-----------------3.3766 => 1160 0x8700 Hassleblad H6D 100C snEdFr54520225540E 0.026363 => 11600x8700 Hasselblad H6D 100c snEdFr54520225536

PAGE 36

27 Phase One IQ3 100MP A natural scene reference image ( CF008259.tif ) taken with a Phase One IQ3 100MP camera was compared against flat field images from the source camera ( sn HP001078 ) , a set of nat ural scene images from an open source camera ( snIG011142) , and both flat field and natural scene sets of images from a camera of the same make and model (snHP001634) . The Gaussian method correctly eliminated one of the open source cameras . It correctly eliminated the flat field set of images from the second open source camera, but had limited support for the natural scene se t of images from that camera ( snHP001634) . It correctly attributed the reference image to the source camera . The Wiener method correctly eliminated one of the open source cameras . It correctly eliminated the flat field set of images from the second open source camera , but had a false positive for the natural scene set of images from that camera ( snHP001634) . It correc tly attributed the reference image to the source camera . Both methods correctly identified the source came ra, but the Wiener method had a false positive . Table 7 Phase One IQ3 100MP R eference I mage from snHP001078 R esults Figure 25 Gaussian Method Phase One IQ3 100MP Log Likelihood Ratio R esults Camera used Phase One IQ3 100MP Gaussian Wiener snIG011142 No Support No Support snHP001634 ( flat field ) No Support No Support snHP001634 ( natural scene ) Limited Support Has Support snHP001078 source Has Support Has Support ------------------Evidence ------------------Evidence file: CF008259.tif -------------------Reference-----------------10.4939 => 11608x8708Phas eOne IQ3 100MP snEdFrHP001078 2.2616 => 11608x8709 PhaseOne IQ3 100MP snTEWHP001634out 0.51505 => 11608x8708PhaseOne IQ3 100MP snTEWHP001634 0.0014213 => 11608x8709 PhaseOne IQ3 100M P snIRIG011142

PAGE 37

28 Figure 26 Wiener Met hod Phase One IQ3 100MP Log Likelihood Ratio R esults Phase One IQ4 150MP A natural scene reference image ( Capture11829.tif) taken with a Phase One IQ4 150MP camera was compared against flat field images from the source camera ( snEdFrJD010673 ) , and a set of natural scene images from an open source camera of the same make and model ( snJD020741) . Note this 151 megapixels sens or was the largest one studied, which caused some computing memory issues in the processing. The original TIFF conversion files were 14204x10652 pixels and around 840 megapixels file size each. This caused a memory error when conduc t ing the camera comparison tests. Neither TIFF l ossless compression nor JPEG compression were able to solve the memory issues. Th e error may arise from an array size limitation on the v ersion of Matlab used in this study. As a solution, images from this camera were down sam pled to 1601x8700 pixels, which allowed the algorithm to process the images without any further errors . The Gaussian method had a false positive for the op en source camera. It correctly attributed the reference image to the source camera . The Wiener method was inconclusive for the open source camera. It correctly attributed the reference image to the source camera . Both methods correctly identified the source camera , but the Gaussian method had a false positive . Table 8 Phase One IQ4 150MP R eference I mage from snJD010673 R esults Camera used Phase One IQ4 150MP Gaussian Wiener snJ D020741 Has Support Inconclusive sn JD010673 source Has Support Has Support ------------------Evidence ------------------Evidence f ile: CF008259.tif -------------------Reference-----------------14.0131 => 11608x8709PhaseOne IQ3 100MP snTEWHP001634out 3.0122 => 11608x8708 PhaseOne IQ3 100MP snEdFrHP001078 0.3793 => 11608x8708 PhaseOne IQ3 100MP snTEWHP001634 0.0028195 => 11608x8709 PhaseOne IQ3 100MP snIRIG011142

PAGE 38

29 Fi gure 2 7 Gaussian Method Phase One IQ 4 1 5 0MP Log Likelihood Ratio R esults Figure 2 8 Wiener Method Phase One IQ 4 1 5 0MP Log Likelihood Ratio R esults ------------------Evidence ------------------Evidence file: Capture11829.tif -------------------Reference-----------------4.5279 => 11601x8700 down sampled from 14204x10652 PhaseOne IQ4 150 snDPRallJD020741 3.1936 => 11601x8700 down sampled from 14204x 10652 PhaseOne IQ4 150MP snEdFrJD010673 ------------------Evidence ------------------Evidence file: Capture11829.tif -------------------Reference-----------------4.0585 => 11601x8700 down sampled from 14204x10652 PhaseOne IQ4 150MP snEdFrJD010673 1.3039 => 11601x8700 down sampled from 14204x10652 PhaseOne IQ4 150 snDPRallJD020741

PAGE 39

30 CHAPTER IV CONCLUSION This was a preliminary study of medium format sensors and th e background Photo Response Non Uniformity with flat field images. Some expected and some unexpected results arose during the analysis of these s ensors . The denoising filters used were the Gaussian and Wiener methods. Overall, the Wiener method per formed better than the Gaussian method for camera comparison, but not as well as in other studies . The Gaussian method correctly attributed or eliminated cameras 12 times with 3 false positive over 21 tests (57.2% success rate, 14.3% error rate). The Wiener method correctly attributed or eliminated cameras 14 times with 1 false positive over 21 tests (66.7% s uccess rate, 4.78% error rate) . Both methods correctly excluded all JPEG images included in the tests. The Wiener method was more reliable than the Gaussian method , but c learly both methods have some limitations when dealing with me dium format sensors, either from the file size or the high er quality of these sensors. Other methods have been developed that may handle camera comparison better. Future research should include possibly a Wavelet methods ( 3 ) , a Bloc k Matching and 3D (BM3D) filter method ( 24 ) , and/ or an Anisotropic Diffusion (AD) filtering schemes ( 25) . Camera comparison was done with cross c orrelation , but other methods like Peak to Correlation Energy (PCE) or Maximum Likelihood Estimate (MLE) exist that may prove to be more reliable in future research. One of the experimental goals of this thesis was working with RAW image data to analyze the full capabilities of these superior sensors . Matlab does not ingest RAW files dire ctly, so they had to be conv erted in to uncompressed TIFF files . A n issue with TIFF files is they tend to grow in size when converting from RAW files, growing from 50 MB – 150 MB to 146 MB – 868 MB . This not only makes collecting large sets of images difficu lt in file management but also working with these files is extremely time consuming. Another limitation in TIFF conversion was the degradation in bit depth from the 14 or

PAGE 40

31 16 bits per channel to only 8 bits. This defeats one of the advantages of medium format cameras over previously researched consumer level cameras. Creating a proper database is essential for source camera comparison . This proved to be a challenge for this type of cam eras. Medium format cameras are not consumer level, they tend to rang e in cost from $5,000 USD to $52,000 USD . The high sticker price means they are mostly sold in specialty camera stores, and places that do have them will tend to only have one mode l on display, making it difficult to collect multiple source samples. Sample images can be found on open source sites, but RAW files are not typically available . Open source s amples are often limited in quantity of images and i n the variety of same make and model cameras . Also, no flat field images were found on open source sites for these cameras . For proper comparison, it would be ideal to have a database of multiple cameras of flat field images , wit h 50 or more images from each camera. One last limitation in creating a comparative database is all the medium format cameras used in this research had their own unique sensor resolution. That is, n one of the sensors analyzed had the same n umber of pixels as any other. This undoubtably introduced some intra make/model fixed pattern noise which made camera comparison less reliable . In conclusion, medium format cameras produce unrival ed images in quality, definition, clarity, dynamic range, and color tones. The higher build quality of these massive sensors produces less background noise than consumer cameras . One might think a larger file would have more background noise for camera comparison , but tha t isn’t the case. From a forensic perspective, the higher quality of medium format sensor s makes correct attribution and elimination of a source camera more difficult than a cheaper con sumer level camera.

PAGE 41

32 REFERENCES 1. Rista J. What causes banding noise in CMOS sensors? 2019 [Available from: https://photo.stackexchange.com . 2. Martinec E. Noise, Dynamic Range and Bit Depth in Digital SLRs 2008 [Available from: http://theory.uchicago.edu/~ejm/pix/20d/tests/noise/ . 3. Fridrich J. Digital Image Forensics Using Sen sor Noise 4. United States of America v Nathan Allen Railey. United States Dist rict Court: Southern District of Alabama; 2011. 5. Said A, Lukas J, Apostolopoulos JG, Fridrich J, Goljan M. Determining digital image origin using sensor imperfections. Image and Video Communications and Processing 20052005. 6. Sargent JTG, A. Test Repor t Approval, nonpublished communication to the FBI. Law Enforcement Analysis Facility. 2009. 7. Goljan MF, J.; Filler, T. Large Scale Test of Sensor Fingerprint Camera Identific ation. Proc SPIE, Media Forensics and Security. 2009;7254. 8. Rosenfeld KS, Husr ev Taha. A Study of the Robustness of PRNU based Camera Identification. Proc SPIE, Media Forensics and Security. 2009;72540M. 9. M. Chen JF. Determining image origin and integrity using sensor noise. IEEE Transactions on Information Forensics and Security. 2008;3:74 90. 10. Li C. Source camera identification using enhanced sensor pattern noise. IEEE Trans Image Process. 2010;5:280 7. 11. Li C, Li, Y. Colordecoupled photo respon se non uniformity for digital image forensics. IEEE Trans Circuits System Video Technology. 2012;22(260271). 12. Cortiana A, Conotter, Valentina, Boato, G., Natale, F.G.B. Performance comparison of denoising filters for source camera identification. SPIE Electronic Imaging. 2011. 13. Hu Y, Jian, C., Li, C. Source camera identificatio n using large components of sensor pattern noise. Proceedings of Int Conf Computer Science Applications. 2009;23:1 5. 14. Kang X, Jiansheng, C., Kerui, I. A contextadaptive spn predictor for trustworthy source camera identification. EURASIP Journal on Image and Video Processing. 2014;1. 15. Lin X, Li, C.T. Preprocessing reference sensor pattern noise via spectrum equalization. IEEE Transactions on Informati on Forensics and Security. 2016;11:12640. 16. Luka J, Fridrich J, Goljan M. Digital Camera Identific ation From Sensor Pattern Noise. IEEE Transactions on Information Forensics and Security. 2006;1(2):205 14.

PAGE 42

33 17. Mehrish A, Subramanyam AV, Emmanuel S. Robu st PRNU estimation from probabilistic raw measurements. Signal Processing: Image Communication. 2018; 66:30 41. 18. Thing YCaVLL. A study on the photo response non uniformity noise pattern based image forensics in realworld applications. Institute for Info comm Research. 2012. 19. Fujifilm x 2019 [Available from: https://fujifilm x.com/en us/products/cameras/ . 20. Hasselblad Medium Format 2019 [Available from: https://www.hasselblad.com/medium format/ . 21. PhaseOne 2019 [Available from: https://www.phaseone.com/en/Photography/Camera Technology /Full Frame Medium Format . 22. M. K. Mihcak IK, K. Ramchandran, and P. Moulin. Low complexity image d enoising based on statistical modeling of wavelet coefficients. IEEE Signal Process Lett. 1999(6(12)):300 3. 23. Benjamin Anderson Sackaney AA D. Evaluatio n of Sensor Pattern Noise Estimators for Source Camera Identification. International Journal of Compu ter and Information Engineering. 2016;10(12). 24. K. Dabov AF, V. Katkovnik, and K. Egiazarian. Image denoising by sparse 3 D transform domain collaborativ e filtering. IEEE Transactions on Image Processing. 2007;16(8):2080 95. 25. W.V. Houten ZG. Using Ani sotropic Diffusion for Efficient Extraction of Sensor Noise in Camera Identification. Journal of Forensic Sciences. 2012;57:521 7.

PAGE 43

34 APPENDIX Graphs from camera comparison tests. FUJIFILM GFX 50s

PAGE 44

35

PAGE 45

36 FUJIFILM GFX 100

PAGE 46

37

PAGE 47

38

PAGE 48

39

PAGE 49

40

PAGE 50

41

PAGE 51

42

PAGE 52

43

PAGE 53

44

PAGE 54

45

PAGE 55

46

PAGE 56

47 Hasselblad X1D II 50c

PAGE 57

48

PAGE 58

49

PAGE 59

50 Hasselblad H6D 100c

PAGE 60

51

PAGE 61

52 Phase One IQ3 100 MP

PAGE 62

53

PAGE 63

54

PAGE 64

55

PAGE 65

56 Phase One IQ4 150 MP

PAGE 66

57