SDANet: symmetric differential attention network for polarization and visible image fusion

Abstract

Polarization and visible image fusion aims to generate high-quality fused images that combine detailed polarization texture features and comprehensive scene information from the two modalities. However, most existing image-fusion algorithms fail to effectively balance fine polarization details and complete scene information, restricting their fusion capability. Therefore, we propose a symmetric differential attention network for polarization and visible image fusion (SDANet). First, we design a dual-path parallel network architecture that organically combines independent and shared feature extraction-fusion-reconstruction branches. This structural design enables the network to effectively mine and utilize discriminative cross-modal features. Furthermore, a differential attention module is embedded in the feature fusion stage. This module implements differential feature encoding for polarization and visible images, which accurately captures the unique feature representations of each modality and efficiently preserves cross-modal complementary information. Finally, a differential loss function is developed to improve the model’s ability to model discrepant information between source images. Extensive qualitative and quantitative experiments on public datasets demonstrate that the proposed SDANet achieves better performance than several representative mainstream fusion methods, which validates the effectiveness of our network.