Semantic segmentation of JPEG blocks using a deep CNN for non-aligned JPEG forgery detection and localization

Abstract

In this paper, a new approach is proposed for non-aligned JPEG forgery detection and localization. Our method is based on the semantic pixel-wise segmentation of JPEG blocks using a deep neural network. Semantic segmentation is the process of assigning each pixel of an image to a class label. We train a deep Convolutional Neural Network (CNN) to segment the boundaries of JPEG blocks. The trained deep CNN can accurately detect block boundaries related to various JPEG compressions. Therefore, non-aligned JPEG forgeries can be easily detected and localized by detecting irregularities in the segmented block boundaries. The proposed approach can detect and localize JPEG forgeries with the same and different quantization matrices as well as image forgeries with several compression stages. We tested the proposed algorithm with various forged and authentic images and compared the results with the state-of-the-art approaches. Experimental results showed that the proposed CNN-based algorithm performs well for non-aligned JPEG forgery detection and localization.