Efficient Nighttime Infrared and Visible Image Fusion for Autonomous Driving Systems via Pixel-Iterative Weighting Estimation

Infrared and visible image fusion plays a critical role in enabling precise environmental perception for autonomous driving systems. However, unlike normal lighting conditions, maintaining high processing speed while achieving excellent visual perception in nighttime environments remains a significant challenge. To overcome above challenges, this study proposes an efficient nighttime infrared and visible image fusion method based on pixel-iterative weight estimation, named PIWE-Fusion. Firstly, we establish a specific iterative fusion model, which can retain the prominent infrared target and adaptively restore the visible dark region by fine-grained adjusting the dynamic range of each pixel in the iterative process. Subsequently, a lightweight weight estimation network is developed to accurately infer the weight parameters of the model. Its accuracy is guaranteed by a set of specially formulated loss functions that can invisibly measure the quality of color consistency, light balance, texture preservation, spatial smoothness and target saliency of the fused image. Extensive experiments on various datasets demonstrate that our method outperforms nine state-of-the-art(SOTA) methods in both qualitative and quantitative evaluations. Specifically, our method achieves improvements of 0.17%, 26.19%, 1.85%, and 0.60% in terms of entropy (EN), standard deviation (SD), spatial frequency (SF), and visual information fidelity (VIF), respectively. Moreover, the excellent operational efficiency enables our method to be effortlessly deployed in autonomous driving systems, making it a practical solution for real-time nighttime environmental perception.