Distinguishing Computer Graphics From Photographic Images Using Local Binary Patterns

With the rapid development of the Internet and the rapid popularization of digital cameras of the low-cost digital image capture device, people can contact with a large number of digital images via ubiquitous Internet, and also can easily generate digital photos with cameras and other devices. With the development of computer rendering technology, it is easy to generate digital images by rendering software such as3D Studio Max, Accurender and Photoshop, that people can not distinguish the images with eyes, but with the emerging "Photo" fake events, the "seeing is believing " traditional concept has been overturned, which results in a negative impact on politics, culture, news, science and other aspects of authenticity. In recent years, the classification of Computer Graphics and Photographic Images based on digital forensics technology has received the attention of scholars both at domestic and foreign, and there are many methods have been developed used for CG from PG automatic classification.According to the above problem, this paper studied the algorithm of realizing the CG and PG automatic classification, the specific work is as follows:1. Proposed a new feature based on image texture Local Binary Patterns (LBP) characteristics to distinguish CG and PG. Select the YCbCr color model and analysis LBP feature correlation of each color component of the images, extract LBP texture features from the CG and PG images’ Y, Cr color components and their prediction error components, respectively. Further the LBP feature dimension is from256dimension down to59dimension, greatly reduces the complexity of operations. Ultimately complete the CG and PG automatic classification by means of support vector machine (SVM). A large number of experimental results show that, LBP algorithm can distinguish the CG and PG effectively, the classification accuracy reaches up to94%, higher than the existing several excellent classification algorithm.2. Proposed an algorithm based on multi-resolution LBP feature extraction. The algorithm considers8neighbors that form a circle with radius of1and2to calculate the central pixel’s LBP feature value, then analysis the classification accuracy of12kinds of multi-resolution LBP features value of YCbCr and their prediction-error components. On this basis, also do some optimized combination of multi-solution LBP features. Experimental results show that multi-resolution LBP features compared to single LBP feature further improves the CG and PG classification accuracy, up to95%and achieved good forensics results.3. Study the robustness of the algorithm based on LBP features. Respectively attack the CG and PG images with JPEG compression using different compression quality factors and different proportions cropping, then extract the LBP features of the attacked images, finally analyze and compare the robustness of the algorithm. The experimental results show that the algorithm for JPEG compression and cropping has good robustness.