Nighttime Image Preprocessing in Traffic Scenes: When Low-light Enhancement Meets Flare Removal

Detecting whether motorcycle drivers are wearing helmets is one of the critical tasks in object detection. However, most existing helmet detection methods are optimized for well-lit daytime conditions and struggle to perform in low-visibility nighttime environments. Night scenes contain poorer visibility and wider blind spots, which pose significant challenges for detection tasks. Additionally, this paper highlights that in nighttime scenarios, the factors affecting detection tasks are not limited to low-light conditions but also include reflective flares from vehicle headlights or streetlights. In response to the two main challenges mentioned above, this paper verifies a widely observed optical center symmetry prior information in cameras, where light sources and their resulting reflective flares are always symmetric about the lens center. This paper also proposes a data preprocessing model, DCRE. This model uses layer decomposition techniques to generate attention maps that guide the enhancement of low-light areas in images while leveraging optical center symmetry prior information to eliminate reflective flares, thereby improving detection accuracy in nighttime conditions. Extensive experiments validate the effectiveness of the DCRE. The detection accuracy increased from 62.8% to 76.0% after preprocessing with the DCRE model.