A statistical feature based approach to distinguish PRCG from photographs☆
Introduction
With the rapid growth and widespread using of digital cameras and mobile phones, photos and pictures can be obtained when and where. At the same time, rapid development of computer vision and computer graphic technologies, make it is more convenient to generate the photorealistic computer graphics (PRCG). These PRCG are so realistic that it is very difficult to differentiate them from photographs and pictures by human eyes. See Fig. 1. The old adage of “seeing is believing” is no longer true. This trend has brought new issues and challenges for us concerning the authenticity of digital images. Therefore, identifying PRCG from photographs has become an important topic in the field of media application.
Photographs are those pictures that are captured from real-world by imaging device such as camera. PRCG are scene images that does not exist in reality; it is generated only by computers and graphic software through the following process: first, a 3D polygon model is built to simulate a desire shape, then color, texture and simulate light irradiation are given to this model. At last, PRCG is imaged using a virtual camera. Recent advances in computer graphics are bring an amazing result, that is, it is possible to create models that are physically accurate without visual artifacts and lighting inconsistencies. Therefore, distinguishing PRCG from photographs is a very challenging problem.
There are many established methodologies to distinguish PRCG from photographs in recent years. The majority of existing methods in literature are based on statistical learning model, in which, the features of inconsistency between photographs and PRCG are used for classification. Therefore, the main difference of existing approaches lies on the features extraction, and features extraction methods can be classified into two categories, the first is based on transform domain and the second is based on the physical characteristics of the imaging equipment.
Taking a general survey on existing methods, there are some problems such as lower detection precision, weaker robustness and unsatisfactory stability. In our work, we aim at improving the detection accuracy, robustness, and generalization capability, and seek for new features that are able to distinguish photographs and PRCG. We find that the texture changes are different between photographs and computer-generated images under same Homomorphic filtering transformation, so we use Homomorphic filtering to highlight the image detail information. Considering that the Contourlet transformation has multi-direction selectivity and anisotropy, and can be used to describe the inherent characteristics of image, we use the statistic features of the Contourlet sub-bands to construct the distinguishing features. We define a customized statistical feature, named texture similarity, and combine it with the statistical features extracted from the co-occurrence matrix of differential matrixes to construct forensics features. Then we develop a statistical model, and use the least squares SVM as a classifier to classify PRCG from photographs. Experimental results show that the proposed method not only possesses satisfactory detection accuracy, also, it is robust to tolerate content-preserving manipulations, and provides satisfactory generalization capability for different image data sets.
The rest of this paper is organized as follows. In Section 2, we introduce the relational works. The proposed method is described in Section 3. Section 4 contains experimental results. Finally, a short summary are presented in Section 5.
Section snippets
Related works
The earliest approaches to distinguish photographs and PRCG were proposed by Ng and Chang [1] and Lyu and Farid [2], respectively. Work [1] presented a classification approach that used the natural image statistics features composed of 33-dimensional (33-D) power spectrum features, 24-D local patch features and 72-D wavelet higher-order statistics. Subsequently, a geometry-based approach [3] was proposed by the same authors to model the physical differences between PRCG and photographs, in
Proposed method
Although PRCG are so realistic that it is very difficult to be differentiated from photographs by human’s eyes, but the imaging mechanism and imaging process are completely dissimilar. The imaging process of the photographs is very complex since it is inevitably influenced by both natural environment and acquisition equipment, while the process of generating a PRCG only primitively simulates the imaging process of photographs. These differences will lead to different characteristics in many
Experimental results and analysis
In this section, we use experimental results to demonstrate the performance of the proposed method. Our algorithm is implemented and tested using MATLAB2010a, and our experiment is performed on the computer with Processor Pentium(R) Dual-Core CPU, E5400 @ 2.70 GHz, 2.00 GB RAM.
Conclusions
Nowadays, computer imagery and computer generated images touch many aspects of our daily life. Computer graphics are getting so photorealistic that it has brought new challenges towards the credibility of digital images. Identifying photorealistic computer graphics from photographs has become an important topic in the field of media application.
In this paper, we aimed at improving the detection accuracy, robustness, and generalization capability, simultaneously, proposed a new approach to
Acknowledgments
We would like to thank Dr. Zhaohong Li for her kindness by providing us with their codes. We also would like to acknowledge the helpful comments and kindly suggestions provided by anonymous referees.
References (28)
- et al.
Perceptual discrimination of computer generated and photographic faces
Digital Invest.
(2012) - T. Ng, S. Chang, Classifying photographic and photorealistic computer graphic images using natural image, Technical...
- et al.
How realistic is photorealistic?
IEEE Trans. Signal Process.
(2005) - T. Ng, S. Chang, J. Hsu, L. Xie, MP. Tsui, Physics motivated features for distinguishing photographic images and...
- Y. Wang, P. Moulin, On discrimination between photo realistic and photo graphic images, in: IEEE International...
- W. Chen, Y.Q. Shi, G.R. Xuan, Identifying computer graphics using HSV color model and statistical moments of...
- X. Cui, X. Tong, G. Xuan, Discrimination between photo images and computer graphics based on statistical moments in the...
- et al.
Identifying computer graphics from digital photographs based on coherence of adjacent pixels
J. Comput. Res. Dev.
(2009) - G. Sankar, V. Zhao, Y.H. Yang, Feature based classification of computer graphics and real images, in: Proc. IEEE...
- W. Li, T. Zhang, E. Zheng, X. Ping, Identifying photorealistic computer graphics using second-order difference...
Cited by (26)
Distinguishing natural and computer generated images using Multi-Colorspace fused EfficientNet
2022, Journal of Information Security and ApplicationsCitation Excerpt :Differences in the generation of natural images and computer graphics, camera or device properties, etc., are considered for the traditional feature based classification works [6–8]. Many other feature based works using color, texture, and shape based statistical features followed this area of image forensics study [9–11]. Leveraging features from various image transformation domains like wavelet [12], contourlet [13], quaternion wavelet [14], etc., is another approach that can be seen in the traditional feature based works.
Distinguishing computer-generated images from photographic images using two-stream convolutional neural network
2021, Applied Soft ComputingCitation Excerpt :In this method also, an SVM classifier was applied. Wang et al. [26] proposed a method to distinguish CG images from PG images based on the homomorphic filtering transformation and SVM classifier. LBP and SVM based method was demonstrated by Li et al. [18].
A Systematic Survey on Photorealistic Computer Graphic and Photographic Image Discrimination
2023, International Journal of Image and GraphicsComputer Graphic and Photographic Image Classification Using Transfer Learning Approach
2022, Traitement du Signal
- ☆
This work was supported by the National Basic Research Program of China (973 Program) under Grant No. 2012CB316400; the National Natural Science Foundation of China under Grant No. 61075007 and No. 61273252.