ESRNet: Efficient Search and Recognition Network for Image Manipulation Detection

With the widespread use of smartphones and the rise of intelligent software, we can manipulate captured photos anytime and anywhere, so the fake photos finally obtained look “Real.” If these intelligent operation methods are maliciously applied to our daily life, then fake news, fake photos, rumors, slander, fraud, threats, and other information security issues around us can happen all the time. Today’s intelligent retouching software can make various modifications to photos, some of which do not change the content that the photos themselves want to express, such as retouching, contrast improvement, and so on. In this article, we mainly study the three operation modes of changing the authenticity of photo contents, which are Copy-move, Splicing, and Removal. Few scholars have done relevant research due to the lack of a corresponding dataset. To address this issue, we elaborately collect a novel dataset, called the multi-realistic scene manipulation dataset (MSM30K), which consists of 30,000 images, including three types of tampering methods, and covering 32 different tampering scenes in life. In addition, we propose a unified detection network: the efficient search and recognition network (ESRNet) for three tampering methods. It mainly includes four main modules: Efficient feature pyramid network (EFPN), Residual receptive field block with attention (RFBA), Hierarchical decoding identification (HDI), and Cascaded group-reversal attention (GRA) blocks. On these three datasets, ESRNet can reach 0.81 on the S-measure, 0.72 on the F-measure, and 0.85 on the E-measure. The inference speed is ~53 fps on a single GPU without I/O time. ESRNet outperforms various state-of-the-art manipulation detection baselines on three image manipulation datasets.