BICFusion: An unsupervised infrared and visible image fusion framework for beyond illumination constraints

Infrared and visible image fusion is aimed at merging features from both modalities in order to produce a more information-rich fused image. However, the majority of existing methods have overlooked the specific requirements and challenges inherent in fusion tasks under low-light conditions. In such scenes, texture degradation due to poor illumination is common, and furthermore, local overexposure may result in significant information loss. To tackle these challenges, a novel framework named BICFusion is introduced, which addresses these issues through reflectance separation, cross-modal feature compensation, and dual enhancement of texture and contrast. The Retinex theory is employed to design a network that extracts reflectance representing the intrinsic structure and details of the scene from the visible image, thereby providing the fusion result with rich structural information under minimal illumination constraints. The cross-modal feature guidance weighting module (CFGW) is developed to compensate for missing details by leveraging the infrared image when the visible image lacks sufficient texture information due to adverse lighting conditions such as low light or overexposure. Subsequently, the texture enhancement fusion module (TEFM) and the global-local contrast enhancement loss function are proposed to jointly enhance the fusion quality in terms of texture and contrast. Experiments conducted with twelve state-of-the-art methods on three publicly available datasets validate the superior performance of BICFusion in preserving fine details under low-light and overexposed conditions.