A Frequency Attention-Based Dual-Stream Network for Image Inpainting Forensics

The rapid development of digital image inpainting technology is causing serious hidden danger to the security multimedia information. In this paper, a deep network called frequency attention-based dual-stream (FADS-Net) proposed for locating region. FADS-Net established by encoder and an blue-associative decoder. includes two feature extraction streams, raw input stream (RIS) recalibration (FRS). RIS directly captures maps from input, while FRS performs after recalibrating via learning in domain. addition, module based on dense connection designed ensure efficient full fusion features. associative decoder consists main branch decoders. up-sampling fine-tuning fused features using attention mechanisms skip connections, ultimately generates predicted mask inpainted image. Then, decoders are utilized further supervise training ensuring that they both work effectively. A joint loss function entire streams optimal forensic performance. Extensive experimental results demonstrate achieves superior localization accuracy robustness multiple datasets compared state-of-the-art forensics methods.