Self-Attention Progressive Network for Infrared and Visible Image Fusion

Abstract

Visible and infrared image fusion is a strategy that effectively extracts and fuses information from different sources. However, most existing methods largely neglect the issue of lighting imbalance, which makes the same fusion models inapplicable to different scenes. Several methods obtain low-level features from visible and infrared images at an early stage of input or shallow feature extraction. However, these methods do not explore how low-level features provide a foundation for recognizing and utilizing the complementarity and common information between the two types of images. As a result, the complementarity and common information between the images is not fully analyzed and discussed. To address these issues, we propose a Self-Attention Progressive Network for the fusion of infrared and visible images in this paper. Firstly, we construct a Lighting-Aware Sub-Network to analyze lighting distribution, and introduce intensity loss to measure the probability of scene illumination. This approach enhances the model’s adaptability to lighting conditions. Secondly, we introduce self-attention learning to design a multi-state joint feature extraction module (MSJFEM) that fully utilizes the contextual information among input keys. It guides the learning of a dynamic attention matrix to strengthen the capacity for visual representation. Finally, we design a Difference-Aware Propagation Module (DAPM) to extract and integrate edge details from the source images while supplementing differential information. The experiments across three benchmark datasets reveal that the proposed approach exhibits satisfactory performance compared to existing methods.