Multi-scale attention context-aware network for detection and localization of image splicing Efficient and robust identification network

Nowadays, advanced image editing tools and techniques produce more realistic tampered images, especially the addition of intelligent retouching technology makes the threshold of tampered operations lower and lower, which can easily evade image forensic systems and make it more difficult to verify the authenticity of images. The field of forensics has failed to achieve effective development due to the lack of high-quality splicing datasets. In this paper, we present the SMI20K-the first benchmark dataset for image splicing operations under intelligent tampered techniques, which contains a total of 20,000 splicing tampered images. By combining Seamless Cloning and image similarity search techniques, the tampered images have more hidden manipulation traces, making it difficult for the naked eye to distinguish the tampered targets. SMI20K brings a new challenge to the field of image forensics. Furthermore, we propose the novel Multi-scale Attention Context-aware Network (MAC-Net) to address the novel challenge of image tampered presently. Specifically, we propose a Multi-scale Multi-level Attention Module (MMAM) that not only effectively resolves feature inconsistencies at different scales, but also automatically adjusts the coefficients of the original input features to maintain detailed features. The fused features are then fed to the proposed Multi-Branch Global Context Module (MGCM), it has three different branches that not only enriches the contextual information but also maintains the detailed features of the target through automatic coefficient adjustment. Extensive experimental results on three public datasets and the proposed dataset show that the proposed model outperforms other state-of-the-art (SOTA) models in image forgery localization.