MMFA-Net: Multi-Modal Feature Learning and Multi-Head Fuse Attention for Image Manipulation Detection and Localization

With the advent of various advanced image processing techniques and the increasingly low manipulation threshold, Image Manipulation Detection and Localization (IMDL) has become a complex and challenging task. Previous studies have shown the importance of manipulation artifacts in the IMDL problem. However, existing methods often struggle to capture effective artifacts when addressing with refined forgery (post-processed forgery). In this paper, we introduce MMFA-Net, a novel approach for IMDL. We extract RGB features, Laplacian high-frequency features, Bayar statistical anomaly features, and manipulation imprint from the image through the Multi-modal Feature Extractor module, concatenating them as multi-modal feature inputs to capture manipulation edge artifacts, which are increasingly difficult to detect in image manipulation. In addition, we propose an innovative Multi-head Fuse Attention (MFA) module. Building on the backbone network's ability to model inconsistencies between original and tampered regions at coarse-to-fine granularities, the MFA module can more effectively capture contextual dependencies in the spatial dimensions and inter-channel correlations for more precise localization. The performance on several public datasets demonstrates that the proposed network outperforms existing SOTA models and remains robust against refined forgery.