DS‐Net: Dual supervision neural network for image manipulation localization

Abstract With the rapid development of image editing technology, tampering with images has become easier. Maliciously tampered lead to serious security problems (e.g., when used as evidence). The current mainstream methods are divided into three types which copy‐move, splicing and removal. Many detection can only detect one type tampering. Additionally, some learn features by suppressing content, result in false positives identifying areas. In this paper, authors propose a novel framework named dual supervision neural network (DS‐Net) localize regions mentioned above. First, extract richer multiscale information, add skip connections atrous spatial pyramid pooling (ASPP) module. Second, channel attention mechanism is introduced dynamically weigh results generated ASPP. Finally, build additional supervised branches for high‐level further enhance extraction these before fusing them low‐level features. conduct experiments on various standard datasets. Through extensive experiments, show that AUC scores reach 86.4% , 95.3% 99.6% CASIA, COVERAGE NIST16 datasets, respectively, F1 56.0% 73.4% 82.7% respectively. demonstrate authors’ method accurately locate achieve better performance datasets than other same type.