Scale-Adaptive Deep Matching Network for Constrained Image Splicing Detection and Localization

Constrained image splicing detection and localization (CISDL) is a newly formulated forensics task that aims at detecting localizing the source forged regions from series of input suspected pairs. In this work, we propose novel Scale-Adaptive Deep Matching (SADM) network for CISDL, consisting feature extractor, scale-adaptive correlation module mask generator. The extractor built on VGG, which has been reconstructed with atrous convolution. computation module, squeeze-and-excitation (SE) blocks truncation operations are integrated to process arbitrary-sized images. generator, an attention-based separable convolutional block designed reconstruct richer spatial information generate more accurate results less parameters burden. Last but not least, design pyramid framework SADM capture multiscale details, can increase accuracy boundaries. Extensive experiments demonstrate effectiveness framework.