Improving Nighttime Flare Removal with Subspace Basic Projection

Flare is an imaging issue caused by optical refraction in the lens. The imaging hazards caused by flares are more pronounced at night due to the large difference in intensity between the light source and the background. Since it is difficult to infer a flare-free scene from a flare-corrupted image, recovering textures under nighttime flare remains a challenging problem. In this paper, we propose a novel nighttime flare removal framework based on subspace projection and a jet map mask-based crossattention mechanism to remove nighttime flare and recover the scene under the flare, termed FBNet. On one hand, we propose subspace projection as a function of the flare basis network to better separate the flare from the scene. On the other hand, we found that the binarized threshold segmentation result obtained by using the jet map as a threshold segmentation condition is remarkably effective in extracting the flare as an input feature to the cross-attention mechanism. Therefore, we came out with a jet map-based transformer cross-attention block that uses a binary mask as a priori to achieve optimal recovery in the flare region Extensive evaluations on synthetic benchmarks and the real-world datasets Flare7K and Flare7pp highlight the superiority of our approach compared to the previous state-of-the-art methods for flare removal. All proposed methods improve the nighttime flare removal performance.