Feature Pyramid Deep Matching and Localization Network for Image Forensics

Abstract

To advance the state of the art of image forensics technologies, a new formulation of splicing localization is proposed, which aims to obtain the masks for both the query and donor images for a pair of query(probe) image and potential donor image if a region of the donor image was spliced into the probe. The former Deep Matching and Validation Network(DMVN) addresses the problem with a novel end-to-end learning based solution. Inheriting the deep dense matching layer, we propose Feature Pyramid Deep Matching and Localization Network(FPLN), whose contributions are three folds. Firstly, instead of using just one feature map as in DMVN, FPLN utilizes a pyramid of feature maps with different resolutions w.r.t. the input image to achieve better localization performance, especially for small objects. Secondly, we add a fusion layer that fuses together all the features after deep dense matching layer, which not only takes full advantage of the correlation information between those features, but is also able to integrate two pathways in DMVN into just one simple pathway, simplifying the subsequent architecture. Lastly, we employ focal loss to address the imbalance problem, as the foreground area is usually much smaller than the background area. The experiments demonstrate the superior performance of our proposed method in detection accuracy and in localizing small tempered regions.