Better utilization of illumination prior via KANs for nighttime flare removal

Strong light sources in nighttime photography often induce flares, which are inherently coupled with the scene illumination distribution. However, existing methods neglect this coupling, resulting in artifacts such as color distortion and edge blurring. Therefore, we propose an illumination-guided frequency adaptive network that innovatively integrates Kolmogorov-Arnold Networks (KANs) to leverage illumination priors fully. By employing adaptive basis functions, KANs effectively model nonlinear variations in illumination, guiding the network to perceive the frequency characteristics of flares and enabling efficient and precise flare removal. Specifically, we design a novel illumination prior learning paradigm that models the complex relationships of the illumination distribution by constructing smooth nonlinear activation mappings, thereby accurately capturing its deep variation patterns. These priors are subsequently utilized to dynamically adjust the frequency distribution of the input features, enabling the model to adaptively enhance or suppress specific frequency components on the basis of the illumination intensity. Additionally, we propose a frequency-guided data synthesis strategy, Swap frequency bands, to mitigate the potential bias introduced by relying on synthetic data for training. Through swapping different frequency bands between the background and synthetic flare images, we provide proper guidance to the model, achieving physically interpretable supervision. Extensive experiments validate the effectiveness of the proposed method. The code is available at https://github.com/AXNing/Illumination-Guided-Frequency-Adaptive-Network.