Computer Graphic and Photographic Image Classification using Local Image descriptors

With the tremendous development of computer graphic rendering technology, photorealistic computer graphic images are difficult to differentiate from photo graphic images. In this article, a method is proposed based on discrete wavelet transform based binary statistical image features to distinguish computer graphic from photo graphic images using the support vector machine classifier. Textural descriptors extracted using binary statistical image features are different for computer graphic and photo graphic which are based on learning of natural image statistic filters. Input RGB image is first converted into grayscale and decomposed into sub-bands using Haar discrete wavelet transform and then binary statistical image features are extracted. Fuzzy entropy based feature subset selection is employed to choose relevant features. Experimental results using Columbia database show that the method achieves good detection accuracy. Keyword : Binary statistical image feature; Image forensics; Digital image forensic; Discrete wavelet transform; Natural image; Fuzzy entropy measure; Computer graphic image; Photo graphic image Figure 1. Examples of photographic images (first column)1 and photorealistic computer graphic images (second column)2. Received : 25 May 2016, Revised : 05 June 2017 Accepted : 21 June 2017, Online published : 06 November 2017 Defence Science Journal, Vol. 67, No. 6, November 2017, pp. 654-663, DOI : 10.14429/dsj.67.10079  2017, DESIDOC BIRAJDAR & MANKAR : COMPuTER GRAPHIC AND PHOTOGRAPHIC IMAGE ClASSIFICATION uSING lOCAl IMAGE DESCRIPTORS 655 in active forensic approach to verify the authenticity of the received image. In passive or blind forensic, the received image is used for assessing its originality without any extra information. Various methods are proposed in the literature to distinguish PG and CG images. These methodologies can be classified into two categories (a) perceptual methods and (b) statistical learning (feature based) methods. Feature based methods are further classified as (1) transform domain methods and (2) methods based on physical characteristics of the imaging equipment7. Perceptual methods are based on human observers to examine PRCG and photographic images8. In case of large number of images, this method becomes unusable. Hence, automatic computer based method is widely used to distinguish photographic images and photorealistic CG images. 1.1 Motivation and our Work Photographic image or natural image is any image originated from a digital imaging sensor (e.g., camera). A computer graphic or synthetic image is any scene (2D or 3D models) rendered by the software either partially or totally9. Based on the survey by existing approaches, there is still a major limitation such as lower detection accuracy. Our proposed work is primarily focused on enhancing the detection accuracy and finding the new features to differentiate the CG and PG images. In this paper, a method is proposed to discriminate photorealistic computer graphic and real photographic image using discrete wavelet transform (DWT) based binarised statistical image features (BSIF). learning based binary BSIF image descriptors are different for PG and CG images and can be used to classify CG and PG images. Binary codes using BSIF are obtained for a set of natural image patches using corresponding filters with different sizes. First, the input image is decomposed into various subbands with different levels using DWT. BSIF features are then extracted from approximation sub-band after second level DWT decomposition. Fuzzy entropy based feature sub-set selection algorithm is employed to select the most relevant and informative features from the input feature space. Finally, support vector machine (SVM) classifier is used to classify CG and PG images. Experimental results show that the proposed method has satisfactory detection accuracy. 2. rELAtEd WorK At present, various methods are proposed for the classification of computer graphic photorealism from natural image. These methods are briefly reviewed in this section. First-order and higher order wavelet statistical characteristics like mean, variance, skewness and kurtosis of the coefficient histograms of four wavelet sub-band and prediction error sub-band are used as features4. First, the input image is decomposed into three levels and from each R, G and B channel statistical features are extracted. Geometry-based features by means of the fractal geometry at the finest scale and the differential geometry at the inter-mediate scale based on textural characteristics is presented by Ng10, et al. Authors found the differences between geometric object models of photographic images and computer-generated images and an average detection accuracy of 83.5 per cent was reported. Approach based on exploiting differences of image acquisition in a digital camera and the generative algorithms used by computer generated graphics is proposed11. This difference is captured using the properties of the residual image (pattern noise in digital camera images) which is extracted by a wavelet transform based denoising filter. Chen12, et al. introduced the statistical moments of characteristic function of the HSV image and wavelet sub-bands as the distinguishing features for detection. 234-D HSV colour model features demonstrated better performance compared to RGB model. Same authors used genetic algorithm based feature selection13 and achieved feature dimension reduction from 234 to 100 with increased detection accuracy. Features based on fractional lower order moments in the image wavelet domain are employed14 resulting in an accuracy of 81.85 per cent and with a feature length of 135. Colour histogram feature, moment-based features, local patch statistics feature, features based on texture interpolation combination is used in by Sankar15, et al for CG and PG classification with detection accuracy of 90 per cent. In another study, features based on the variance and kurtosis of second-order difference signals and the first four order statistics of predicting error signals are used by li16, et al. with detection accuracy of 90.2 per cent. Conotter and Cordin17 proposed transform domain approach in which wavelet transform domain features4 and sophisticated pattern noise statistics feature fusion are used for classification with detection accuracy of 85.3 per cent. Hidden Markov Tree (HMT) based feature extraction approach was proposed by Pan and Huang18 to classify the natural images and computer graphics images and obtained average detection accuracy of 84.6 per cent. Daubechie wavelet was adopted to construct the HMT in the experiment and the classification was based on SVM classifier. Zang and Wang proposed a method based on imaging features and visual features extracted from wavelet sub-bands19. First, the wavelet coefficients in high frequency sub-bands (different sub-images) are separated by a threshold T similar to image denoising approach. Finally, statistical characteristics and cross correlation of wavelet coefficients from sub-band components are used as features to differentiate real photographic and PRCG images. A scheme to classify natural and fake image based on multi-resolution decomposition using 2D-DWT and higher order local autocorrelations is proposed20. Support vector machine (SVM) is employed for the classification, resulting in a detection accuracy of 76.82 per cent. Face asymmetry information is extracted to develop a geometric-based method to classify computer generated and natural human faces21 using two datasets. Dataset 1 contains very realistic images, which are almost undetectable by human and Dataset 2 contains more images, related to real situations. Peng22, et al. combined statistical, textural and physical characteristic features. Statistical parameters such as the mean, variance, kurtosis, skewness and median of the histograms of grayscale image in the spatial and wavelet domain are selected and the fractal dimensions of grayscale image are selected as statistical features. In addition to this, wavelet sub-bands are extracted as visual features and the mean, variance, skewness, kurtosis