Hierarchical Deep Feature Fusion GAN for SAR and Visible Images via Cross-Modal Enhancement

This research proposes a Hierarchical Deep Feature Fusion GAN based on cross-modal enhancement for Synthetic Aperture Radar (SAR) and visible image fusion. Through deep learning techniques, the network addresses the challenging fusion task between visible and SAR images caused by their distinct imaging mechanisms, achieving high-quality fusion with rich detail preservation. Initially, the network leverages the bidirectional generative capabilities of CycleGAN for modal enhancement to mitigate issues related to training data scarcity and quality. In the feature extraction phase, the ResNet-50 serves as the backbone network to extract hierarchically refined deep features from both SAR and visible images. During the feature fusion phase, lightweight attention mechanism modules are employed to fuse deep features at corresponding hierarchical levels. Finally, we designed a hierarchical nested connection architecture to reconstruct the fused image from multi-level fused deep features. Experimental results demonstrate that our method exhibits superior image fusion performance in both visual perception and objective metrics. Keywords- synthetic aperture radar (SAR); visible images; crossmodal enhancement; hierarchically refined deep feature