Image Manipulation Detection via Multiscale Region and Cross-Image Edge Contrastive Learning

Pixel-level image manipulation localization requires separating weak forensic traces from image content while recovering manipulated regions under post-processing and domain shift. Existing methods do not jointly impose structured discrimination on multiscale region features and edge features drawn from different images. We therefore propose a dual contrastive learning framework built on a Mesorch-based convolutional-Transformer detector. Multiscale Region Contrastive Learning (MRCL) applies intra-image directional contrast to authentic and manipulated samples from four hierarchically fused feature scales. Cross-Image Edge Contrastive Learning (CECL) uses Canny-derived authentic edges and mask-derived manipulated boundary bands to construct within-image contrast, while separate historical queues provide manipulated positives and authentic negatives for cross-image contrast. Both objectives regularize training without changing the active prediction path at inference. Under Protocol-MVSS, models are trained on CASIAv2 and evaluated on COVERAGE, Columbia, NIST16, CASIAv1, and IMD2020. Across seeds 42, 43, and 44, the complete model improves mean F1 and IoU over the Mesorch-based binary-cross-entropy baseline on all five test sets. In the descriptive comparison with representative methods, its three-seed mean is highest for both metrics on NIST16 and IMD2020 and for F1 on CASIAv1, with the second-highest IoU on CASIAv1.