HFDA-Net: Utilizing High-Frequency Feature and Dual-Attention to Enhance Image Manipulation Detection and Localization

In this paper, we propose HFDA-Net, a novel approach for image manipulation detection and localization (IMDL) tasks. Unlike existing methods that only extract high-frequency features from the input image, HFDA-Net further extracts high-frequency features from RGB feature maps, capturing richer manipulation traces. In addition, HFDA-Net introduces a new module that efficiently calculates and combines position and channel attention, improving the accuracy and efficiency of manipulated region localization. Moreover, HFDA-Net supports feature extraction and aggregation at multiple scales and employs a coarse-to-fine pattern to predict manipulated regions, demonstrating remarkable generalizability. Thanks to its lightweight architecture, HFDA-Net achieves a processing speed of 65+ FPS when handling 1080P images. Extensive experiments on four image forensics benchmarks demonstrate that HFDA-Net generally outperforms existing advanced methods in manipulation detection by 1% to 15% and in manipulation localization by 1.5% to 5.4% under AUC. Furthermore, HFDA-Net exhibits good robustness compared to existing methods.