面向小目标与模糊目标的轨道缺陷检测框架

黄腾飞,火久元,曹芳,武泽宇,孔垂乐

Framework for railway defect detection oriented to small and blurry object

Tengfei HUANG,Jiuyuan HUO,Fang CAO,Zeyu WU,Chuile KONG

表 4 动态多尺度与纹理感知框架在不同基准算法上的性能对比

Tab.4 Performance comparison of dynamic multi-scale and texture-aware framework on different baseline algorithms

模型 P/% R/% F1/% mAP@0.5/% mAP@0.5:0.95/% np/106 rf/(帧·s−1) YOLOv8 85.6 72.3 78.4 85.1 46.1 3.16 77.50 DMT-YOLOv8 91.1 81.2 85.9 89.1 50.2 3.19 76.20 Faster R-CNN (ResNet50+FPN) 51.9 62.7 56.8 61.0 28.8 41.34 22.30 DMT-Faster R-CNN (ResNet50+FPN) 64.8 75.6 69.8 71.1 36.5 51.37 17.10 YOLOv5n 82.1 71.6 81.3 81.8 45.8 1.77 59.72 DMT-YOLOv5n 87.0 81.2 84.0 85.9 48.5 1.81 55.26