Machine-vision-oriented evaluation of low-light image enhancement algorithms

Abstract

All-weather visual perception is indispensable for autonomous driving, intelligent transportation, smart cities, mobile robotics, and security monitoring. However, images captured at night or under weak illumination usually contain low brightness, weak contrast, color shift, sensor noise, and missing texture, which reduce both human readability and downstream machine-vision performance. This paper evaluates low-light image enhancement algorithms from a machine-vision-oriented perspective rather than treating enhancement as only a visual restoration problem. Four representative baselines are selected: Zero-DCE, SCI, Retinexformer, and LLFlow. The study combines paired image-quality evaluation on LOL eval15 with downstream object-detection evaluation on a stratified ExDARK subset. PSNR and SSIM are used to measure visual fidelity, while YOLOv5s-based mAP and Recall are used to test whether enhancement improves object detection. GFLOPs and FPS are further measured to analyze deployment feasibility. The experimental results show that stronger visual enhancement does not automatically improve detection. Retinexformer obtains the best PSNR and LLFlow obtains the best SSIM, but the original low-light input achieves the highest mAP@0.5 and Recall on the selected ExDARK subset. SCI provides the strongest efficiency profile because of its extremely low computational cost and high FPS. These results indicate that low-light enhancement for machine vision should be selected according to task-level evidence, not only according to image appearance or pixel-level metrics.