基于多尺度特征相似性匹配的低照度目标检测

于鑫淼,夏楠,江佳鸿,郝子莹,把云胜

Low-light target detection based on multi-scale feature similarity matching

Xinmiao YU,Nan XIA,Jiahong JIANG,Ziying HAO,Yunsheng BA

表 2 所提方法与最新优化算法在自建数据集上的检测精度与速度对比

Tab.2 Comparison of detection accuracy and speed of proposed method and latest optimized algorithms on self-built dataset

方法 AP/% mAP/% FPS/ (帧·s−1) 自行车 公交车 汽车 货车 摩托车 行人 RetinaNet [10] 76.1 77.6 66.5 64.2 68.8 68.2 70.2 66.0 YOLOv10[31] 76.6 79.2 69.1 68.9 73.7 72.9 73.4 82.6 YOLOv11[32] 77.3 81.4 69.7 69.3 73.6 73.0 74.0 81.5 YOLOv12[33] 79.0 83.1 76.3 75.7 72.0 77.4 77.1 80.3 RT-DETR[7] 78.9 80.3 71.9 72.6 71.2 69.9 74.1 76.2 DENet[16] 78.7 80.7 74.9 74.3 72.7 74.8 76.0 83.6 IAT[34] 81.2 82.6 77.8 76.6 73.9 74.9 77.8 79.4 PE-YOLO[18] 80.7 82.4 77.9 79.2 76.4 79.5 79.3 79.9 WSA-YOLO[17] 81.6 83.4 79.2 79.3 76.5 78.9 79.8 78.4 本研究方法 82.3 84.0 79.4 77.9 78.6 80.2 80.4 77.6