CMFD-CC: image copy-move forgery detection via context clustering

In the context of increasingly complex image forgery techniques, image copy-move forgery detection faces huge challenges. The discrepancies of features between the source region and the target region are extremely subtle, and there is no difference in optical aberrations, so it’s a tough job to effectively discern the forged region. This paper puts forward a technique for the detection of copy-move forgeries within images based on context clustering (CMFD-CC). It utilizes a structure with an encoder and a decoder combined. The encoder includes down-sampling, dilated embedding, and feature clustering modules. At the same time, semantic illumination information is used as auxiliary input to extract the illumination consistency between image patches. Through self-correlation calculation and dilated embedding, the similarity of image blocks is enhanced, and the forged source and target features are clustered and matched. The decoder decodes by layer-by-layer up-sampling. During training, the loss function is determined according to the discrepancies between the predicted mask and the ground truth mask. Experiments are conducted on benchmark datasets such as USC-ISI, CASIA v2.0, and DEFACTO. The comparison experiments demonstrate that CMFD-CC performs better in F1 value and accuracy. On the USC-ISI dataset, the F1 value is significantly improved compared with some strategies. Moreover, the ablation experiments also authenticates the viability of each module of the model. The CMFD-CC method achieve excellent performance in the detection of copy-move forgeries within images and provides a new and effective solution.