To solve the problem that the Copy-Move Forgery Detection(CMFD)network is difficult to detect the falsified regions of different scales at the same time,a U-shaped network BCSU-Net is proposed,which integrates BiFormer and cross-scale correlation computing module.Different from the existing CMFD network,which uses convolutional backbone network to extract local features,BCSU-Net uses BiFormer to capture the long-distance dependency between pixels to better extract the highly correlated features in the feature map.In addition,a cross-scale correlation computing module is proposed to calculate the similarity between features at different scales,which helps the model locate the tamper region in the Copy-Move forged image more accurately.Compared with existing methods,BCSU-Net shows better performance on COVERAGE and CoMoFod datasets.
维普
万方数据