Cross-scale condition aggregation and iterative refinement for copy-move forgery detection.

Copy-move forgery detection poses a significant challenge because the brightness and contrast of forged and real regions are highly consistent, and the forensic clues are weakened by various affine transformations and post-processing operations. Existing deep learning models ignore the hierarchical dependencies of the features and pay less attention to information purification. To solve these problems, we propose a novel two-stage model to detect copy-move forgery, dubbed by CCAIR-Net. First, in the coarse localization stage, we propose a cross-scale condition aggregation module to integrate multi-level features from global to local. This aggregation structure can eliminate semantic disparities and capture multi-scale hierarchical dependencies of features. In particular, the condition aggregation block enables feature purification and filtering of interfering information. Second, in the refinement stage, the designed weighted fusion mechanism can guide the model to remove falsely detected regions and supplement the miss-detected regions by adaptively weighted fusion of coarse-grained and fine-grained features. Furthermore, the stage-wise training strategy takes advantage of different losses to train the network to detect tampered regions at various scales. Extensive experiments demonstrate that the proposed CCAIR-Net performs better than state-of-the-art methods. It can detect and segment forged and real areas more accurately, even for affine transformations and post-processing attack images. [ABSTRACT FROM AUTHOR]