Attention-guided polarization image fusion based on guided filtering

To address the limitations of existing polarization image fusion methods that often overlook noise interference in Degree of Linear Polarization (DoLP) images and indiscriminately fuse all source image information, we propose a novel image fusion method. Firstly, we created a polarization dataset, using DoLP images as guide and input images. For guided filtering, the filter window radius and regularization parameter were set to 2 and 0.05, respectively, to produce the output image. Subsequently, we developed an unsupervised polarization intensity image fusion network incorporating the Convolutional Block Attention Module (CBAM). This network comprises three sub-modules: feature extraction, fusion, and reconstruction. The feature extraction module employs an Attention-based Dense Block (ADB) to extract prominent features from the source images. We validated our approach using both the created dataset and a publicly available dataset, comparing it against seven advanced image fusion methods using seven quantitative evaluation metrics. Additionally, we conducted an ablation study to assess the impact of the attentional mechanism. Qualitative and quantitative results demonstrate that our proposed method effectively merges linear polarization information while suppressing noise and distortion from DoLP, preserving the significant features of the target more effectively than the other methods.