Fig. 1. Architecture of PCBNet
Fig. 2. Dilation and extrusion convolution
Fig. 3. DeConv receptive field
Fig. 4. Architecture of DeCSPlayer
Fig. 5. SPD-Conv
Fig. 6. Subtle feature enhancement module
Fig. 7. Schematic diagram of feature extraction in the PE module
Fig. 8. Schematic diagram of PCB surface soldering defects
Fig. 9. Image acquisition and detection system
Fig. 10. Visualization of detection results for each algorithm
Fig. 11. Visualization of detection results for each algorithm
| Number | Dilation | Layer | s | Output size |
|---|
| 0 | - | Focus | - | 320×320×32 | | 1 | - | SPD-Conv | - | 160×160×64 | | 2 | 1 | DeCSPlayer | - | 160×160×64 | | 3 | - | SPD-Conv | 2 | 80×80×128 | | 4 | 2 | DeCSPlayer | - | 80×80×128 | | 5 | - | SPD-Conv | 2 | 40×40×256 | | 6 | 3 | DeCSPlayer | - | 40×40×256 | | 7 | - | SPD-Conv | 2 | 20×20×512 | | 8 | - | SPP | - | 20×20×512 | | 9 | 5 | DeCSPlayer | - | 20×20×512 |
|
Table 1. Dilatation rate configuration table in the backbone
| 项目 | 参数 |
|---|
| 相机型号 | MV-CS060-10UC-PRO | | 光圈 | F4 | | 焦距 | 8 mm | | 光源类型 | Ring light | | 曝光时间 | 50 000 μs | | 分辨率 | 3 072×2 048 | | 物距 | 100 mm | | 视野 | 50 mm×35 mm |
|
Table 2. Camera parameters during data collection
| 缺陷类别 | 缺陷数目 |
|---|
| Aiguille | 2 487 | | Dissymmetry | 1 667 | | Holes | 1 923 | | Interconnect pad | 4 069 |
|
Table 3. Number of defects in various categories in the dataset
| 方法 | 图像尺寸 | mAP0.5/% | mAP0.5：0.95/% | Params/Mb | FPS |
|---|
| YOLOX-S | 640 | 92.5 | 60.8 | 8.94 | 106 | | YOLOX-S+SFEM | 640 | 93.6 | 64.7 | 9.73 | 89 | | YOLOX-S+DeConv | 640 | 94.5 | 64.6 | 9.36 | 90 | | YOLOX-S+SPD-Conv | 640 | 94.3 | 64.9 | 13.64 | 100 | | YOLOX-S+SPD-Conv+SFEM | 640 | 94.8 | 64.7 | 14.43 | 90 | | YOLOX-S+SPD-Conv+SFEM+DeConv | 640 | 95.3 | 65.0 | 14.85 | 86 |
|
Table 4. Results of the ablation study
| 方法 | mAP0.5/% | mAP0.5：0.95/% | Params /Mb | FPS |
|---|
| r=2 | 94.6 | 64.7 | 15.92 | 85 | | r=3 | 95.3 | 65.0 | 14.85 | 86 | | r=4 | 95.1 | 64.9 | 14.54 | 87 | | r=5 | 94.7 | 65.1 | 14.41 | 87 |
|
Table 5. Impact of different r values in DeConv on the optimal model accuracy
| 空洞率 | mAP0.5 /% | mAP0.5：0.95 /% | Params /Mb | FPS |
|---|
| 2，2，2，2 | 95.0 | 65.4 | 14.85 | 87 | | 1，2，1，2 | 95.2 | 65.3 | 14.85 | 88 | | 1，6，12，18 | 95.3 | 65.0 | 14.85 | 86 | | 1，2，5，7 | 94.9 | 65.0 | 14.85 | 86 | | 1，2，3，5 | 95.4 | 65.5 | 14.85 | 83 | | 1，2，5，1 | 94.9 | 65.0 | 14.85 | 87 | | 1，2，9，12 | 95.1 | 65.2 | 14.85 | 86 |
|
Table 6. Impact of different dilation values in the backbone of PCBNet on the optimal model accuracy
| 样本数 | mAP0.5 /% | mAP0.5：0.95 /% | Params /Mb | FPS |
|---|
| 80% | 95.4 | 65.5 | 14.85 | 87 | | 70% | 95.2 | 65.3 | 14.85 | 87 | | 60% | 95.0 | 65.2 | 14.85 | 87 |
|
Table 7. Performance statistics of the model with different numbers of training samples
| 模型 | 主干 | AP0.5（%） | mAP0.5 /% | mAP0.75 /% | mAP0.5∶0.95 /% | FPS | Params /Mb |
|---|
| A | D | H | I-P |
|---|
| Faster-RCNN | resnet50 | 61.8 | 63.8 | 66.7 | 85.3 | 69.4 | 15.3 | 28.5 | 30 | 136.75 | | CenterNet | resnet50 | 88.7 | 91.7 | 95.3 | 92.7 | 92.1 | 61.0 | 55.2 | 101 | 32.67 | | RetinaNet | resnet50 | 83.8 | 87.9 | 91.8 | 92.9 | 89.1 | 61.0 | 54.9 | 53 | 36.39 | | EfficientDetv1 | EfficientNet-B0 | 50.8 | 68.8 | 71.0 | 86.2 | 69.2 | 25.4 | 32.7 | 33 | 3.83 | | YOLOV4-Tiny | | 63.7 | 84.8 | 89.6 | 89.1 | 81.8 | 28.3 | 37.9 | 192 | 5.88 | | YOLOV5-s | C3 | 80.8 | 80.9 | 92.0 | 91.6 | 86.3 | 56.0 | 51.0 | 92 | 7.07 | | YOLOV8-s | C2f | 90.9 | 93.9 | 94.9 | 94.4 | 93.5 | 76.8 | 63.4 | 114 | 11.14 | | CFPNet | - | 87.7 | 90.6 | 96.1 | 96.5 | 92.7 | 71.9 | 60.8 | 89 | 13.14 | | YOLOX-s | - | 86.7 | 92.1 | 95.6 | 95.5 | 92.5 | 72.3 | 60.8 | 106 | 8.94 | | YOLOX-m | - | 90.8 | 92.1 | 97.3 | 96.4 | 93.8 | 75.5 | 63.0 | 74 | 25.28 | | YOLO-PCB | - | 54.0 | 80.0 | 89.5 | 64.2 | 71.9 | 65.8 | 52.6 | 95 | 8.9 | | MDVI | - | 75.8 | 87.7 | 94.0 | 93.9 | 87.8 | 43.3 | 46.0 | 161 | 12.22 | | RTDD-LCD | - | 68.8 | 86.8 | 93.9 | 92.8 | 85.6 | 34.9 | 41.6 | 73 | 25.08 | | PCBNet | - | 91.2 | 94.8 | 97.4 | 98.2 | 95.4 | 79.0 | 65.5 | 83 | 14.85 |
|
Table 8. Results compared to other methods（For statistical convenience， aiguille is simplified as A， disymmetry is simplified as D， holes is simplified as H， and Interconnect pad is simplified as I-P）
| 任务 | 图像尺寸 | 数目 |
|---|
| 训练 | 600×600×3 | 8 460 | | 验证 | 600×600×3 | 961 | | 测试 | 600×600×3 | 1 067 |
|
Table 9. Dataset splitting
| 模型 | 模型版本 | 主干 | mAP0.5/% | mAP0.75/% | mAP0.5∶0.95/% | FPS | Params/Mb |
|---|
| Faster-RCNN | - | resnet50 | 72.6 | 11.2 | 28.4 | 30 | 136.75 | | CenterNet | - | resnet50 | 97.7 | 46.3 | 52.1 | 101 | 32.67 | | RetinaNet | - | resnet50 | 92.7 | 43.4 | 49.3 | 53 | 36.39 | | EfficientDetv1 | - | EfficientNet-B0 | 85.6 | 38.6 | 43.4 | 33 | 3.83 | | YOLOV4-Tiny | - | | 97.2 | 32.6 | 46.6 | 192 | 5.88 | | YOLOV5 | s | C3 | 98.3 | 60.0 | 56.6 | 92 | 7.07 | | YOLOV8 | s | C2f | 99.2 | 78.4 | 65.2 | 114 | 11.14 | | CFPNet | s | - | 98.8 | 65.3 | 59.3 | 89 | 13.14 | | YOLOx | s | - | 98.2 | 59.6 | 57.0 | 106 | 8.94 | | YOLOX | m | - | 99.1 | 86.7 | 76.5 | 74 | 25.28 | | YOLO-PCB | - | - | 96.7 | 71.5 | 61.3 | 95 | 8.9 | | MDVI | - | - | 98.2 | 41.4 | 50.5 | 161 | 12.22 | | RTDD-LCD | - | - | 97.6 | 39.0 | 49.1 | 73 | 25.08 | | PCBNet | - | - | 99.5 | 88.7 | 73.2 | 83 | 14.85 |
|
Table 10. Results compared to other methods
