EM-Net: A Deep Learning Model Using Enhanced Attention and Multi-Scale Feature Fusion for Text Image Forgery Detection

In this paper, we propose an EM-Net for detecting the tiny regions of subtle features, which includes three stages. First, a Multi-scale Feature Extraction module (MSFE-module) is used to extract multi-scale and multi-layer feature information. A Feature Enhanced (FE-module) catches the key information in different scales and refines the results by assigning more weights to forgery areas. Therefore valuable features can be enhanced while irrelevant information is suppressed. As the network is constantly downsampling the image, the resolution becomes small. To restore the size, in the Feature Fusion decoder module (FFU-module), the multi-level information by integrating the global features and location details, so as to improve the performance of locating the forgery areas. To demonstrate the validity of our model, we compare the performance of the proposed EM-Net with state-of-art methods. The results show that our performance was excellent on the testing dataset.