Robust detection for object removal with post-processing by exemplar-based image inpainting

Abstract

Exemplar-based image inpainting can be maliciously used for object removal forgery without leaving any perceptual clues. Especially, post-processing might further brings challenges for its blind forensics. In the paper, a robust forensics approach is presented to detect object removal tamper by exemplar-based image inpainting with post-processing, such as JEPG compression, blurring, imnoise, and so on. Object removal changes local texture and gradient smoothness, which destroys the inherent properties of nature images. Two local texture descriptors including LBP (Local Binary Patterns) and GLCM (Gray-level Co-occurrence Matrix) are exploited to measure texture variation, and image gradient is used to describe the structure change. Fourteen statistical features including Zernike zero-order moment, min, max, mean, variance and standard deviation are extracted from them. Then, support vector machine (SVM) is exploited as pattern classifier to determine whether an image has been suffered from object removal or not. Experimental results show the detection robustness for object-removal forgery with post-processing.