GP-LGFusion: Deep Fusion of Infrared-Visible Driving Images Based on Gradient Preservation and Local Guidance

Infrared and visible driving image fusion represents a pivotal technology in multi-source perception for automated driving. The objective of this technology is to generate fused images that exhibit significant targets and comprehensive road information in complex traffic scenes. However, the existing image fusion algorithms demonstrate inconsistent capacity to complement information in diverse environments. Additionally, there are limitations in their ability to extract features, such as the detailed texture of traffic targets under complex lighting conditions, including low-light scenes and multi-exposure scenes. To overcome these limitations, we propose a novel gradient-preserving and locally guided fusion method (GP-LGFusion). Our primary contribution is a Multi-scale Gradient Residual Block (MGRRB), an encoder module specifically designed to capture and retain both strong and weak texture features across different scales, a capability lacking in conventional approaches. Second, we introduce a Local Feature Guide Block (LFGB) in the decoder, which uses local context to guide feature aggregation, significantly improving the reconstruction quality of object contours. Lastly, we design a tailored loss function that prioritizes features relevant to driving perception. Experimental findings demonstrate the efficacy of this algorithm in preserving the target contour and enhancing road texture details in standard driving scenarios. Its performance is shown to exceed that of existing methods.