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 1 所提方法与经典算法、最新优化算法在ExDark数据集上的检测精度与速度对比

Tab.1 Comparison of detection accuracy and speed of proposed method, classical algorithms and latest optimized algorithms on ExDark dataset

方法 AP/% mAP/% FPS/ (帧·s−1) 自行车 船 瓶子 公交车 汽车 猫 椅子 杯子 狗 摩托车 行人 桌子 Faster R-CNN[8] 83.0 72.3 74.6 85.0 82.6 78.6 77.2 81.6 81.0 82.7 81.3 72.0 79.3 71.1 Mask R-CNN[9] 87.4 74.4 78.1 87.9 83.1 80.2 80.6 81.0 78.3 76.6 83.2 70.6 80.2 68.5 RetinaNet [10] 84.5 73.2 72.7 86.5 80.8 76.8 76.8 75.9 73.4 78.3 76.6 67.1 76.9 74.6 YOLOv10[31] 85.2 76.4 82.7 87.4 80.9 75.9 75.3 79.2 82.5 82.3 80.6 74.6 79.8 93.7 YOLOv11[32] 87.6 76.9 82.4 87.5 81.2 76.3 77.0 80.9 81.9 79.6 81.0 74.4 80.5 92.7 YOLOv12[33] 88.7 76.8 82.0 88.1 81.9 76.0 77.3 81.2 82.4 82.9 81.4 74.7 81.1 91.5 RT-DETR[7] 85.1 76.6 81.4 87.0 81.2 75.6 81.3 81.8 82.2 82.8 83.7 70.9 80.8 84.0 DENet[16] 85.6 75.2 77.8 84.4 83.5 77.9 78.7 79.5 80.6 83.5 82.3 74.0 80.3 94.1 IAT[34] 86.5 75.6 77.4 88.7 83.2 79.6 81.1 80.5 77.6 83.1 80.3 76.4 80.9 88.8 PE-YOLO[18] 88.7 75.4 79.8 90.6 83.9 77.8 82.5 82.4 78.7 82.5 80.8 73.4 81.4 91.2 WSA-YOLO[17] 88.0 78.8 81.3 92.6 84.6 78.5 80.3 80.9 80.7 84.3 81.9 77.1 82.4 87.2 本研究方法 89.3 81.5 82.6 94.2 86.1 77.6 79.6 82.0 82.9 84.5 83.1 74.2 83.4 85.0