Multi-Scale Supervised Spatio-Channel Aggregation Network for Image Manipulation Detection and Localization

Recent advancements in image editing technology have made image content more susceptible to manipulation, designing a universal detection method for forged images is imperative. This study introduces the Multi-Scale Supervised Spatio-Channel Aggregation Network (MSSA-Net). This network not only assimilates the semantic information of the original image but also exploits manipulation artifacts at the edges of tampered image to learn agnostic semantic features. By supervising the network at multiple scales (pixel, edge, and image), it is trained to acquire more generalized features.