DSBU-Net: Dual-stream branch union network for Image manipulation localization

Image manipulation detection and localization is a hot topic in the field of computer forensics, but with the development of image editing technology, it is easier to tamper with images. Although previous work has attempted to capture features from the noise views to explore the noise inconsistency between authentic and forgery regions with some success, more efficient fusion methods are still being explored. To address this problem, a dual-stream branch union network (DSBU-Net) for image forgery localization is proposed. Firstly, an end-to-end DSBU-Net network structure is designed to jointly capture the features of both RGB view and noise view. Secondly, in order to make the network model focus on the inconsistency between the authentic and forgery regions, as well as to suppress the impact of semantic information on localization, the extracted noise is encoded by the noise branch encoder to obtain noise features. Next, the RGB features processed by the RGB branch encoder are input into the designed dual-stream fusion module based on the selfattention mechanism to obtain the fused features. Finally, the forgery regions are localized by inputting through the skip connection into the RGB decoder. Experimental results show that the proposed method outperforms current popular methods on datasets such as CASIA, COVERAGE, and NIST16. At the same time, the effectiveness of the dual-stream fusion module has been proven through ablation experiments.