低光照目标检测:增强、分辨率与模型规模评测OA

Object detection under low-light conditions is a critical problem in computer vision and intelligent robotics.In applications such as nighttime surveillance,autonomous driving,and military reconnaissance,insufficient illumination leads to loss of detail,amplified noise,and reduced contrast,which significantly impairs detector performance.We present a systematic evaluation of five representative low-light enhancement methods like Zero-DCE,Zero-DCE++,RetinexFormer,EnlightenGAN,and KinD++on the public ExDark dataset.Using the original(unenhanced)images as the baseline,we assess their impact on the YOLOv5 detector with two model scales(v5s,v5x)and two input resolutions(640×640,960×960).It is showed that most enhancements improve mean Average Precision(mAP)and Recall,with Zero-DCE++and KinD++offering the most balanced gains.Crucially,scaling the detector to the larger YOLOv5x yields more pronounced improvements and markedly reduces reliance on enhancement,generally outperforming any enhancement applied to the smaller YOLOv5s.Moreover,increasing input resolution provides limited benefit for YOLOv5s while incurring substantial computational overhead.These findings reveal a potential mismatch between perceptual enhancement and task-driven detection,and they offer practical guidance for algorithm design and system implementation in robotic vision and nighttime surveil-lance.