Image Tampering Detection Method Based on Hybrid Attention Mechanism

Aiming at the problems of current image tampering detection methods, such as inaccurate localization or poor robustness. We propose a novel network model structure leveraging a hybrid attention mechanism. The model incorporates two parallel branches: the main branch is dedicated to extracting features from RGB images, emphasizing the identification of visual artifacts like unnatural tampering boundaries and strong contrast differences, while the secondary branch, employing constrained convolution and the spatial rich model (SRM) filter, is focused on extracting features associated with noise. To enhance image representation, we introduce a hybrid attention mechanism module within the dual stream. This module includes a positional attention mechanism and a window-based self-attention mechanism. Additionally, we employ atrous spatial pyramid pooling to effectively fuse the features from the dual streams. The experimental results showcase the efficiency of the suggested approach, outclassing several advanced techniques in both detection and localization assignments.