IML-ViT-RSD: Enhancing Image Manipulation Localization with Multi-Scale Feature Extraction and Contextual Fusion

With the continuous advancement of digital image processing, image tampering has become increasingly prevalent and sophisticated, particularly in complex backgrounds and highrisk scenarios. To address these challenges, we propose an enhanced deep learning model, IML-ViT-RSD, based on the IMLViT architecture, designed to improve Image Manipulation Localization (IML) in such environments. First, we introduce an SIAF mechanism, which captures fine-grained relationships among high-level semantic features, enabling more precise IML traces. Next, a DSFPN is integrated to generate multi-scale semantic features, enhancing the model’s ability to handle diverse image contexts. To further improve generalization, depthwise separable convolutions are employed to reduce the parameter count. Finally, the RCM is introduced to refine contextual information modeling, strengthening the model’s understanding of multi-scale objects. Experimental results on the CASIA dataset demonstrate that the proposed model achieves an F1 score of 71.65, reflecting its robust generalization and resilience in IML.