基于上下文信息融合与动态采样的主板缺陷检测方法

鞠文博,董华军

Motherboard defect detection method based on context information fusion and dynamic sampling

Wenbo JU,Huajun DONG

表 7 GC10-DET数据集上本研究算法与其他算法的精度对比

Tab.7 Comparison of accuracy between proposed algorithm and other algorithms on GC10-DET dataset

模型 P/% mAP/% 冲孔 焊缝 月牙 水斑 油斑 丝斑 异物 压痕 折痕 腰折 Faster RCNN 94.4 75.1 86.0 59.9 65.5 50.1 24.5 24.4 55.0 88.3 62.3 Cascade RCNN 98.4 89.8 94.8 73.0 72.3 60.2 18.5 34.6 16.7 66.6 62.5 TridenNet[32] 94.5 66.4 95.8 76.7 72.9 67.2 38.7 41.7 38.5 82.1 67.5 RetinaNet 97.6 76.0 93.6 65.2 53.7 59.2 9.9 21.8 3.2 82.6 56.3 SSD 91.6 74.4 92.8 55.2 61.2 68.9 16.8 15.5 52.7 87.2 61.6 Swin transformer 97.6 86.5 94.6 70.0 64.8 61.6 14.1 28.0 8.3 88.2 61.4 YOLOv7tiny 95.2 79.1 96.1 81.5 60.1 58.2 34.9 23.3 53.1 78.9 66.0 YOLOv8s 96.0 88.3 96.9 84.6 57.9 61.3 22.6 29.9 51.1 81.8 67.1 Yolov9 95.6 87.0 97.0 83.5 56.6 62.0 27.1 21.4 62.8 87.1 68.0 BC-YOLO（本研究算法） 98.5 90.6 98.0 85.9 61.2 58.7 29.0 39.8 64.6 88.2 71.5