DPMSN: A Dual-Pathway Multiscale Network for Image Forgery Detection
Nianyin Zeng, Peishu Wu, Yuqing Zhang, Han Li, Jingfeng Mao, Zidong Wang
Abstract
Multimedia images have become an important way for the sharing of digital information. However, advanced editing tools provide easy methods for malicious content modification, resulting in less viable information. Therefore, there is an urgent need to design an algorithm for image tampering detection and localization, which is capable of locating the authenticity region of the received image, thus delivering the accurate information and assisting in correct decision making in industries and other fields. In this article, a novel dual-pathway multiscale network (DPMSN) is proposed for the image forgery detection, which mainly focuses on extracting the edge information. In particular, a dual-pathway structure is deployed to align visual features in red, green and blue (RGB) space and edge information in LAB space, where a coarse prediction mask is generated to promote accurate localization of the forged regions. By applying the variation convolution operators, comprehensive attention can be paid to various forged regions in multiple sizes. Moreover, in the multiscale fusion module, features at different stages and other low-level information are sufficiently fused to realize a robust presentation of the forged regions. Experimental results show the effectiveness of DPMSN as compared with other state-of-the-art image forgery detection models and the great robustness when facing image attacks, which means DPMSN is a trustworthy forgery detection approach in the industrial field.