Login| Sign Up| Help| Contact|

Patent Searching and Data


Title:
IDENTIFYING COMPUTER GRAPHICS FROM DIGITAL PHOTOGRAPHS
Document Type and Number:
WIPO Patent Application WO/2008/089330
Kind Code:
A3
Abstract:
Computer graphics may be detected in digital images by extracting a first set of features from an input digital image, extracting a second set of features from a prediction- error image derived from the input digital image, and applying a classification algorithm to the first set of features and the second set of features to determine if the combined sets of features indicate that the input digital image corresponds to computer graphics.

Inventors:
SHI YUN QING (US)
CHEN WEN (US)
Application Number:
PCT/US2008/051314
Publication Date:
October 02, 2008
Filing Date:
January 17, 2008
Export Citation:
Click for automatic bibliography generation   Help
Assignee:
NEW JERSEY TECH INST (US)
SHI YUN QING (US)
CHEN WEN (US)
International Classes:
G06V10/52; G06V10/56
Other References:
FARID H ET AL: "How Realistic is Photorealistic?", IEEE TRANSACTIONS ON SIGNAL PROCESSING, IEEE SERVICE CENTER, NEW YORK, NY, US, vol. 53, no. 2, 1 February 2005 (2005-02-01), pages 845 - 850, XP011125221, ISSN: 1053-587X
S. LYU: "Natural image statistics for digital image forensics", August 2005, PH.D. THESIS, DARMOUTH COLLEGE, DARMOUTH, XP002489495
CUTZU F ET AL: "Estimating the photorealism of images: distinguishing paintings from photographs", PROCEEDINGS 2003 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION. CVPR 2003. MADISON, WI, JUNE 18 - 20, 2003; [PROCEEDINGS OF THE IEEE COMPUTER CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION], LOS ALAMITOS, CA : IEEE COMP. SOC, US, vol. 2, 18 June 2003 (2003-06-18), pages 305 - 312, XP010644687, ISBN: 978-0-7695-1900-5
YUANHAO CHEN ET AL: "Automatic Classification of Photographs and Graphics", MULTIMEDIA AND EXPO, 2006 IEEE INTERNATIONAL CONFERENCE ON, IEEE, PI, 1 July 2006 (2006-07-01), pages 973 - 976, XP031033000, ISBN: 978-1-4244-0366-0
SHI Y Q ET AL: "Image Steganalysis Based on Moments of Characteristic Functions Using Wavelet Decomposition, Prediction-Error Image, and Neural Network", MULTIMEDIA AND EXPO, 2005. ICME 2005. IEEE INTERNATIONAL CONFERENCE ON AMSTERDAM, THE NETHERLANDS 06-06 JULY 2005, PISCATAWAY, NJ, USA,IEEE, 6 July 2005 (2005-07-06), pages 269 - 272, XP010844364, ISBN: 978-0-7803-9331-8
Attorney, Agent or Firm:
AMERNICK, Burton, A. et al. (1875 Eye Street NWSuite 110, Washington DC, US)
Download PDF:
Claims:

CLAIMS

What is claimed is:

1. A method of identifying computer graphics in digital images, the method comprising: extracting a first set of features from an input digital image, said extracting comprising: applying a discrete transformation to decompose the input digital image into a set of sub-bands; and obtaining sets of moments of characteristic functions associated with the sub-bands and with the input digital image, wherein said sets of moments are used to form said first set of features; extracting a second set of features from a prediction-error image derived from said input digital image; and

applying a classification algorithm to said first set of features and said second set

of features to determine if the combined sets of features indicate that the input digital image corresponds to computer graphics.

2. The method of Claim 1, wherein said extracting a second set of features from a prediction-error image comprises: applying a discrete transformation to decompose the prediction-error image into a set of sub-bands; and

obtaining sets of moments of characteristic functions associated with the sub-

bands and with the prediction-error image, wherein said sets of moments are used to form

said second set of features.

3. The method of Claim 1, wherein said discrete transformation comprises a discrete wavelet transform.

4. The method of Claim 3, wherein said discrete wavelet transform comprises a Haar discrete wavelet transform.

5. The method of Claim 1, wherein said extracting a first set of features from an input digital image further comprises: forming an image histogram associated with each sub-band; and forming the characteristic function associated with each sub-band based on the associated image histogram of that sub-band.

6. The method of Claim 1 , further comprising: forming said prediction-error image, wherein each pixel of the image is predicted based on three of its neighboring pixels.

7. The method of Claim 1, further comprising: decomposing said input digital image according to a color model prior to performing said extracting the first set of features and extracting the second set of

features, wherein said extracting the first set of features and extracting the second set of

features are applied individually to each component of said color model to obtain a set of

features associated with said component.

8. The method of Claim 7, wherein said color model is selected from the group consisting of the HSV color model and the YC 1 C b color model.

9. The method of Claim 7, wherein said color model is selected using a genetic algorithm.

10. The method of Claim 1, further comprising: applying a genetic algorithm to said first and second sets of features obtained based on a digital image database to obtain a revised set of features prior to said applying a classification algorithm; and selecting a subset of said first and second sets of features obtained from said input digital image based on said revised set of features; wherein said applying a classification algorithm applies the classification algorithm to said subset as said combined sets of features.

11. The method of Claim 1 , further comprising: downloading over a communication network software code that, when executed by a processor, causes the processor to implement said extracting a first set of features, said extracting a second set of features, and said applying a classification algorithm.

12. A processor-readable medium containing software code that, when executed by a processor, causes the processor to implement a method of identifying computer graphics in digital images, the method comprising: extracting a first set of features from an input digital image, said extracting comprising: applying a discrete transformation to decompose the input digital image into a set of sub-bands; and obtaining sets of moments of characteristic functions associated with the sub-bands and with the input digital image, wherein said sets of moments are used to form said first set of features; extracting a second set of features from a prediction-error image derived from said input digital image; and applying a classification algorithm to said first set of features and said second set of features to determine if the combined sets of features indicate that the input digital image corresponds to computer graphics.

13. The medium of Claim 12, wherein said extracting a second set of features from a prediction-error image comprises: applying a discrete transformation to decompose the prediction-error image into a set of sub-bands; and

obtaining sets of moments of characteristic functions associated with the sub-

bands and with the prediction-error image, wherein said sets of moments are used to form said second set of features.

14. The medium of Claim 12, wherein said discrete transformation comprises a discrete wavelet transform.

15. The method of Claim 14, wherein said discrete wavelet transform comprises a Haar discrete wavelet transform.

16. The medium of Claim 12, wherein said extracting a first set of features from an input digital image further comprises: forming an image histogram associated with each sub-band; and forming the characteristic function associated with each sub-band based on the associated image histogram of that sub-band.

17. The medium of Claim 12, wherein said software code, when executed by a processor, further causes the method implemented by the processor to comprise: forming said prediction-error image, wherein each pixel of the image is predicted based on three of its neighboring pixels.

18. The medium of Claim 12, wherein said software code, when executed by a processor, further causes the method implemented by the processor to comprise:

decomposing said input digital image according to a color model prior to

performing said extracting the first set of features and extracting the second set of features, wherein said extracting the first set of features and extracting the second set of features are applied individually to each component of said color model to obtain a set of features associated with said component.

19. The medium of Claim 18, wherein said color model is selected from the group consisting of the HSV color model and the YC 1 -C b color model.

20. The medium of Claim 18, wherein said color model is selected using a genetic algorithm.

21. The medium of Claim 12, wherein said software code, when executed by a processor, further causes the method implemented by the processor to comprise: applying a genetic algorithm to said first and second sets of features obtained

based on a digital image database to obtain a revised set of features prior to said applying a classification algorithm; and

selecting a subset of said first and second sets of features obtained from said input digital image based on said revised set of features; wherein said applying a classification algorithm applies the classification algorithm to said subset as said combined sets of features.

Description:

IDENTIFYING COMPUTER GRAPHICS FROM DIGITAL

PHOTOGRAPHS

Field of the Invention

[001] Embodiments of the present invention may be directed to the detection of computer graphics from digital photographs.

Background of the Invention

[002] For the purposes of this application, "computer graphics" will be used to refer to creating pictures via computer technology, also sometimes referred to as "rendering." With the ongoing development of rendering technology, computer-generated graphic images not only have come to appear more realistic in basic features, but they have also come to be more realistic in terms of such features as color, lighting/shadowing, etc. As a result, it has become more difficult over the years to distinguish image forgery. Therefore, detecting computer graphics in images has become a significant issue in such areas as criminal investigation, journalism, and intelligence.

Summary of Embodiments of the Invention

[003] Various embodiments of the invention may be directed to identifying computer graphics in digital images and may include extracting a first set of features from an input digital image; extracting a second set of features from a prediction-error image derived from the input digital image; and applying a classification algorithm to the first set of

features and the second set of features to determine if the combined sets of features indicate that the input digital image corresponds to computer graphics. One or both of the feature extractions may include applying a discrete transformation to decompose the

input digital image into a set of sub-bands; and obtaining sets of moments of characteristic functions associated with the sub-bands and with the input digital image, wherein the sets of moments are used to form the set of features. [004] Various embodiments of the invention may be implemented in hardware, software, firmware, and/or combinations thereof.

Brief Description of the Drawings

[005] Various embodiments of the invention will be described below in conjunction with the attached drawings, in which:

[006] Figure 1 shows a conceptual block representation/flowchart of various embodiments of the invention;

[007] Figure 2 presents a conceptual flowchart of feature extraction according to various embodiments of the invention;

[008] Figure 3 depicts the use of image pixels to obtain a pixel prediction according to some embodiments of the invention;

[009] Figure 4 gives a further conceptual presentation of feature extraction according to various embodiments of the invention; and

[0010] Figure 5 shows a conceptual block diagram of a system that may be used to implement various embodiments of the invention.

Detailed Description of Various Embodiments of the Invention

[001 1 ] The detection of computer graphics may be addressed as a pattern recognition problem. This particular problem may be decomposed into a two-part process, as shown

in Figure 1 : to extract a set of appropriate features 11 and to pass those features through an appropriate classifier 12 to distinguish computer graphics (CG) from photographic

images (PI).

[0012] In one approach to the extraction of features, moments of wavelet characteristic functions may be used to form, for example, 78-dimensional feature vectors for use in image detection. This approach to extracting features is shown in Figure 2.

[0013] As noted in Figure 2, there may be a set of features derived based on the image and a set of features derived from a prediction-error image that may be derived from the image. These, together, may form the overall set of features.

[0014] As shown in Figure 2, the input images may undergo wavelet decomposition 21.

In the exemplary embodiment, in which 78-dimensional feature vectors may be obtained, a three-level wavelet transform may be used; however, the invention need not be limited to three-level transforms. However, for the purposes of this discussion, a three-level discrete wavelet transform (DWT) will be the assumed example. Such a three-level

DWT may be used to provide twelve sub-bands, denoted by LL 1 , HL 1 , LH 1 , and HH,, where i—1, 2, 3. In an exemplary embodiment, a three-level ηaar DWT may be used; however, the invention need not be limited to this example.

[0015] The histogram of a digital image or a wavelet sub-band may be used to reflect the probability mass function (pmf) of the image or wavelet sub-band, if the image values or the wavelet coefficient values, respectively, are treated as a random variable. Therefore, the process may next take the digital image and its sub-bands, obtained via the wavelet decomposition 21, obtain corresponding histograms, and take their characteristic functions 22, where a characteristic function may be obtained by taking the discrete

Fourier transform (DFT) of a histogram. A characteristic function (corresponding to one of the thirteen sets of image data being considered in this exemplary embodiment, the image and its sub-bands), will be denoted H(J), where/may represent frequency. [0016] After obtaining characteristic functions 22, statistical moments of the characteristic functions may be obtained 23 to provide the desired features. The «-th statistical moment, M n , of a characteristic function, H(f), may be defined as follows.

where |H(/Jt)| is the magnitude of the characteristic function and N is the total number of discrete value levels of the image pixels.

[0017] In an exemplary embodiment, a set of 39 derived features may include the I s ', 2 nd and 3 rd moments of the characteristic functions of 13 sub-bands (the image itself, LLi,

HL 1 , LH 1 , HH 1 , LL 2 , HL 2 , LH 2 , HH 2 , LL 3 , HL 3 , LH 3 , HH 3 ), as discussed above.

[0018] A second set of 39 features, again, using the exemplary embodiment, may be similarly generated for a prediction-error image. One exemplary prediction algorithm, found in M. Weinberger et al., "LOCOI: A Low-Complexity Context-Based Lossless

Image Compression Algorithm," Proc. IEEE Data Compression Conf., 1996, pp. 140-

149, that may be used in some embodiments of the invention may be expressed as follows:

max(α,b) c ≤ min(a,b) mϊn(a,b) c ≥ max(a,b) (2) a + b - c others

where x is the prediction value of the pixel x under the consideration. The pixels, a, b, and c, are three neighboring pixels within an image 31, as shown in Figure 3. The locations of a, b, and c for this algorithm can be illustrated as shown by the sub-image 32. Prediction error may be obtained by taking the difference between the pixel value in the input image and x corresponding to that pixel value. In such a manner, a prediction-

error image may thus be obtained 24.

[0019] Following the generation of the prediction-error image 24, wavelet decomposition 25, generation of characteristic functions for histograms 26, and obtaining moments of the characteristic functions 27 may be performed similarly to the corresponding operations for the input image. In the exemplary embodiment, this results in another 39 features (i.e., first, second, and third moments for each of the prediction-error image and twelve sub-bands obtained via wavelet decomposition), for a total of 78 features obtained from a given image.

[0020] Color images may be represented in a variety of different color spaces, such as RGB (red, green, blue), GLHS (general LHS (lightness, hue and saturation)), HSV (hue, saturation, and values (brightness)), HLS (hue, lightness, and saturation), and HSI (hue, saturation, and intensity), and YC 1 -C b (luminance and chrominance). While various embodiments of the invention may use HSV and/or YC r C b as their color models, the invention need not be limited to only those models.

[0021] The feature extraction process is illustrated in Figure 4 for the case in which the

HSV model is used (as noted above, the invention is not intended to be thus limited, and YCrC b or other models may also be used). The color test image (for example, photo or graphics) 41 may first be represented in H (hue), S (saturation) and V (values (brightness)) channels 42. Then, for each channel, its prediction error image may be obtained 42 based on the prediction model of Equation (2), above, or on another appropriate prediction model. The three-level wavelet decomposition may then be performed for each channel and for its associated prediction error image, resulting, again in the case of a three-level DWT (to which the invention is not limited) in thirteen sub- bands (including the original channel/prediction error image). For each sub-band, the first three moments of characteristic function may be derived, as in Equation (1), resulting in 39 features 43; note that the invention is not limited to the case of the first three moments, and other numbers of moments may be used. Thus, for the three-level DWT case, one

may obtain, for the three channels, a total of 234 features 44.

[0022] It may be additionally shown that feature efficiency may vary for different color models used with a given image or type of image. Therefore, one may apply a genetic optimization algorithm to determine an optimal color space.

In an exemplary embodiment, one may derive a color model from the RGB color model using the following formula:

Cl a\ al α3 R ύilO

Cl = aA a5 aβ G + al l

C3 al aS a9 B all

There are twelve matrix elements al, a2, ... , a 12, in the above formula. That is to say,

one need only determine a finite number of matrix elements in order to determine a specific color model from the RGB model. One may use, for example, a continuous genetic algorithm to search for an optimal color model that can lead to the best performance in detecting computer graphics from photographic images based on a graphics/photographs database sufficient to cover a sufficiently broad range of possible images (i.e., based on the images to which one expects to apply the particular embodiment of the invention). According to the genetic algorithm, one may use "crossover" followed by "mutation" procedures iteratively to select optimal matrix elements. The "fitness measure" in this framework may be the correct detection rate, the balance between true positive and true negative, etc., and/or combinations thereof; the invention is not limited to any particular fitness measure.

[0023] Furthermore, once the features have been derived, it is possible that there may be redundant or unnecessary features among the derived features. Genetic algorithms may also be applied to the derived feature set to obtain a reduced feature set. In one particular example, in which a YC r Cb was used, the 234-dimensional feature set was reduced to a set of 117 features without significant degradation in classification performance. [0024] In a particular embodiment of the invention, this may, once again, be done based on a sufficient graphics/photographs database, which may be similar to that discussed above. In this application of a genetic algorithm, one may use "crossover" followed by "mutation" procedures iteratively to select an optimal subset of features. The "fitness measure" in this framework may be the correct detection rate, the balance between true positive and true negative, etc., and/or combinations thereof; the invention is not limited

to any particular fitness measure. The occurrence of each feature may then be counted,

and the most frequently occurring features may be selected to form an optimal feature

subset.

[0025] Returning now to Figure 1, classification 12 may be performed on the features extracted in feature extraction 11. In some embodiments of the invention, feature classification 12 may be implemented by means of a support vector machine (SVM) with a polynomial kernel. However, the invention need not be limited to this particular form of classifier, and, for example, other kernels and/or other classifiers (e.g., neural network-

based classifiers, etc.) may be employed.

[0026] Various embodiments of the invention may comprise hardware, software, and/or firmware. Figure 5 shows an exemplary system that may be used to implement various forms and/or portions of embodiments of the invention. Such a computing system may include one or more processors 52, which may be coupled to one or more system memories 51. Such system memory 51 may include, for example, RAM, ROM, or other such machine-readable media, and system memory 51 may be used to incorporate, for example, a basic I/O system (BIOS), operating system, instructions for execution by processor 52, etc. The system may also include further memory 53, such as additional RAM, ROM, hard disk drives, or other processor-readable media. Processor 52 may also be coupled to at least one input/output (I/O) interface 54. I/O interface 54 may include one or more user interfaces, as well as readers for various types of storage media and/or connections to one or more communication networks (e.g., communication interfaces

and/or modems), from which, for example, software code may be obtained.

[0027] Various embodiments of the invention have been presented above. However, the

invention is not intended to be limited to the specific embodiments presented, which have been presented for purposes of illustration. Rather, the invention extends to functional equivalents as would be within the scope of the appended claims. Those skilled in the art, having the benefit of the teachings of this specification, may make numerous modifications without departing from the scope and spirit of the invention in its various aspects.