GRA‐Net: Geometry‐ and Response‐Aware Convolutional Network for Nighttime Deflaring

Nighttime image quality in autonomous driving is often degraded by flare, which arises from reflections of intense streetlights and headlights on lenses. Existing methods for flare removal usually exhibit structural distortion and inconsistent illumination in restored images, due to insufficient directional priors and poor light modulation. To overcome these drawbacks, we propose GRA‐Net, a novel geometry‐ and response‐aware convolutional network that achieves unified spatio‐photometric flare rectification by leveraging geometric priors and adaptive illumination responses. To suppress flare artifacts and restore structural integrity, we introduce a geometry‐aware residual module (GARM), which captures anisotropic propagation cues via orientation‐sensitive convolutional pathways. To address illumination inconsistency and prevent artifacts, we design a response‐aware illumination compensation (RAIC) module, which adaptively modulates illumination correction based on spatially varying flare severity. Extensive experiments on the Flare7K++ dataset demonstrate that GRA‐Net outperforms existing lightweight baselines in both quantitative accuracy and perceptual quality.