MITD-Net: Multi-scale iterative tamper detection network for image manipulation localization

Abstract

Advancements in image tampering technology have escalated social risks and security challenges associated with tampered images. Accurately identifying tampered regions remains a significant challenge. To address the issue effectively, a multi-scale iterative tampering detection network is proposed, which decomposes the task into two phases, the Global-Local Feature Synchronization Phase and the Tampered Area Refinement Phase. Initially, dual-branch preprocessing and parallel feature extraction are employed. The Enhanced Downsampling Attention Block (EDAB) is introduced, and an Edge Enhancement Module (EEM) is implemented to extract edge features. Subsequently, Multi-Scale Dilated Convolution (MSDC) is utilized in the tampered area refinement phase, along with the proposed Pixel-level Feature Clustering Module (PFCM) and an iterative mechanism. Finally, incorporating a cross-scale fusion mechanism enables the synthesis of feature information from various scales, thereby enhancing the model's ability to detect subtle alterations in tampered regions. MITD-Net was validated on publicly available datasets, including CASIA, Columbia, COVERAGE, NIST 16, and IMD 20, achieving AUC scores of 84.5%, 98.5%, 85.8%, 86.6%, and 82.6%, respectively. Numerous experiments indicate that MITD-Net achieves higher detection accuracy, exhibits superior robustness, and can adapt to various tampering types and styles.