Enhanced polarization image fusion via dense blocks and multi-scale attention mechanism

Polarization image fusion combines information from different polarizations to enhance image quality. We propose DMAFusion, a unique deep network architecture designed specifically for fusing the degree of linear polarization and intensity (S0) images. DMAFusion uses residual dense blocks (RDBlocks) to retain more information while reducing network parameters. In addition, residual channel attention fusion modules are integrated between the encoder and decoder at multiple scales, enabling adaptive parameter adjustment across different channels. Our approach significantly enriches polarization information and adjusts brightness levels, resulting in improved visual perception. Compared with state-of-the-art methods, DMAFusion demonstrates superior performance in qualitative and quantitative evaluations, enhancing the richness of fused images by 6.67% in entropy and 39.11% in standard deviation. These results validate the effectiveness of our proposed method for comprehensive polarization image fusion.