Robust Median Filtering Forensics by CNN-Based Multiple Residuals Learning

Abstract

Median filtering (MF), a non-linear smoothing operation, has often been utilized as a means of image denoising to protect image edges and hide the traces of image tampering. Although the existing MF forensics methods have achieved excellent performance without post-processing, there is still a challenge of detecting MF in small image size and JPEG compression scenario. To meet this challenge, a robust MF forensic method using convolutional-neural-network (CNN)-based multiple residuals learning is proposed in this paper. Firstly, to reveal the traces left by MF, we use multiple high-pass filters to initialize the weights of the pre-processing layer, and obtain discriminative residuals to characterize MF artifacts in various aspects. Then the output of the pre-processing layer is employed as the input of CNN, which is elaborately designed to extract rich hierarchical features for further classification. Furthermore, Batch Normalization (BN) is introduced as a regularization method to help accelerate convergence of the entire network. The extensive experimental results on the composite database demonstrate that the proposed method is superior to the state-of-the-art methods when detecting MF in both JPEG compressed and small size images.