Enhanced Image Manipulation Detection with TPB-Net: Integrating Triple-Path Backbone and Dual-Path Compressed Sensing Attention

Tri-Path Backbone Network (TPB-Net) is introduced, trained end-to-end for effective detection of various image manipulations. Addressing the challenge of localizing image manipulations, which stems from the difficulty in extracting diverse forgery features, a Triple-path Interconnected Backbone (TIB) is employed for robust feature detection. The development of the Dual-path Compressed Sensing Attention (DCSA) module, featuring a dual-path attention mechanism, marks a significant advancement. This module efficiently compresses channels and spatial information, thereby enhancing learning efficiency, representation effectiveness, and model robustness. TPB-Net, an end-to-end framework with trainable modules, promotes joint optimization for optimal performance. Extensive experiments on four standard image manipulation datasets affirm TPB-Net's superior performance over existing state-of-the-art methods.