In complex visual scenarios such as autonomous driving and video surveillance, intense light sources (e.g., vehicle headlights, direct sunlight) trigger local overexposure, loss of image details and reduced contrast, which severely impede subsequent image analysis and recognition. Hence, strong light suppression technology is indispensable.

At present, mainstream solutions fall into two primary categories:

These approaches improve visual quality by optimizing image contrast and dynamic range. Representative techniques include improved adaptive histogram equalization, homomorphic filtering, and Retinex theory (especially the multi-scale Retinex algorithm). The algorithm decomposes an image into an illumination component and a reflectance component; it suppresses over-bright illumination components to restore details in dark areas.

This category delivers more targeted suppression. First, threshold segmentation or halo detection locates strong light regions, which are treated as interfering superimposed light layers. Next, image inpainting techniques (neighborhood-based interpolation, deep learning generation, etc.) or inverse illumination models are adopted to estimate and subtract the interfering light layer. Background information is then used for image reconstruction, so that strong light can be suppressed while original image information is preserved to the maximum extent.

Traditional algorithms prioritize fast processing in luminance space or Lab color space. By contrast, modern deep learning methods (such as convolutional neural networks built on encoder-decoder architectures) perform end-to-end mapping learning from overexposed strong-light images to clear outputs, delivering superior visual results yet requiring massive datasets for training.

For real-world deployment, engineers need to align algorithms with the lighting characteristics of specific application scenarios, striking an optimal trade-off among suppression intensity, detail retention and real-time processing performance.