MDFusion:A Multi-Scale Feature Disentangling Framework for Multi-Modality Image Fusion

• 1、School of Comupter Science,Beijing University of Posts and Telecommunications，beijing 100876

Abstract：In the field of computer vision research, cross-modal image information integration technology is attracting widespread attention. To address the limitations of existing algorithms in feature detail extraction and cross-domain adaptability, this study innovatively constructs the MDFusion multi-scale feature separation architecture. The core of this architecture lies in proposing a novel feature interaction decoupling mechanism, which significantly reduces mutual interference between features and effectively enhances image detail preservation through a multi-dimensional cross-modal information extraction strategy. In terms of technical implementation, this study optimizes and improves the UNETR network architecture by introducing a multi-channel information processing mechanism, thereby enhancing the model's ability to fuse multi-scale features of source images and achieving more comprehensive image information capture. This innovative architectural design not only improves the model's feature learning capability but also achieves breakthrough progress in cross-domain adaptability. Experimental validation demonstrates that this framework can achieve direct transfer application to cross-domain datasets without the need for parameter adjustment or additional training, reaching industry-leading levels on multiple evaluation metrics. This research provides new technical insights and solutions for cross-modal image processing in the field of computer vision.

Keywords： Computer application technologyey Multi modal image fusion Computer vision Deep learning Multi scale feature disentangling