Degradation-Resistant Infrared-Visible Image Fusion With Auto-Generated Textual Objectives and Embedded Contrastive Learning

Infrared-visible image fusion aims to combine multi-modal image information to generate informative and robust scene representations, thereby enhancing perception capabilities and reliability in intelligent transportation systems. However, captured images often suffer from complex degradation issues, leading to low-quality source data. Existing methods are deficient in adapting to multiple degradation conditions, which limits their fusion performance. In this paper, we aim to develop a degradation-resistant image fusion method that automatically adapts to various degradations. For this purpose, we first construct an auto-generation prompt pipeline based on cascaded multi-modal and language models. It utilizes the vision-language understanding capabilities of large models to comprehensively detect degradation, then produces degradation prompts and corresponding text-based fusion objectives for each image. To resist degradations and produce the fusion results as described by fusion objectives, we next propose an embedded contrastive learning method within CLIP space to supervise the model training. This method ensures that the image fusion process is free from degradation and better aligned with the fusion objectives, which enhances the fusion model’s anti-degradation capability. Extensive experiments on public datasets validate the superiority and generalization ability of our method, and its robust degradation-adaptive capability makes it particularly suitable for complex scenes.