Image Splicing Detection Based on Hybrid CNN-Transformer Network

Image splicing is widely recognized as the predominant technique in image manipulation. Most current splicing detection methods have low accuracy and poor localization of small-sized splicing areas. Therefore, we design a hybrid CNN-Transformer network for image splicing detection. Firstly, a hybrid structure encoder is constructed, which first uses CNN to extract the input images' local features, and then inputs it to Transformer to calculate long-distance dependencies to obtain global feature representation. Then, a deformable attention enhancement module (DAEM) is designed to make the CNN local features at different scales at each stage more refined and improve the deformation adaptability of the network. Finally, features from the hybrid structured encoder are added to the decoder via skip connections, to make the decoder perform better feature reconstruction. Results on three datasets show that the network can more accurately locate tampered areas and has superior performance than other methods.