Chinese Journal of Lasers, Volume. 49, Issue 21, 2104005(2022)
Improved Lightweight X-Ray Aluminum Alloy Weld Defects Detection Algorithm Based on YOLOv5
Figures & Tables(17)
Fig. 7. Data analysis. (a) Distribution of center point of labeling frame; (b) size distribution of labeling frame
Fig. 8. Change curves of precision rate and recall rate. (a) Precision; (b) recall
Fig. 10. Detection effect diagrams of different models. (a) YOLOv5s; (b) YOLOv5-Tiny-DIoU; (c) YOLOv5-Tiny-CIoU
Table 1. YOLOv5-Tiny structure diagram
View tableTable 1. YOLOv5-Tiny structure diagram
Serial number From Parameters Module Arguments 0 －1 3520 Focus [3，32，3] 1 －1 704 DWConv [32，64，3，2] 2 －1 35200 C3 [64，64，1] 3 －1 1408 DWConv [64，128，3，2] 4 －1 222464 C3 [128，128，3] 5 －1 2048 SELayer [128，16] 6 －1 2816 DWConv [128，256，3，2] 7 －1 887296 C3 [256，256，3] 8 －1 8192 SELayer [256，16] 9 －1 5632 DWConv [256，512，3，2] 10 －1 2231296 C3 [512，512，1] 11 －1 656896 SPP [512，512，[5，9，13]] 12 －1 1024 DWConv [512，256，1，1] 13 －1 0 Upsample [None，2] 14 [－1，8] 0 Concat [1] 15 －1 408192 GhostBottleneck [512，512] 16 －1 768 DWConv [512，128，1，1] 17 －1 0 Upsample [None，2] 18 [－1，5] 0 Concat [1] 19 －1 105792 GhostBottleneck [256，256] 20 －1 2560 DWConv [256，128，3，2] 21 [－1，16] 0 Concat [1] 22 －1 105792 GhostBottleneck [256，256] 23 [19，22] 12336 Detect
Table 2. Hardware facilities
View tableTable 2. Hardware facilities
Hardware Configure CPU I5-11400 GPU RTX 3060 Memory /GB 16
Table 3. Training parameters before thawing
View tableTable 3. Training parameters before thawing
Parameter Content Size /(pixel×pixel) 416×416 Batch_size 16 Learning_rate 10－3 Optimizer Adam Epoch 0－80
Table 4. Training parameters after thawing
View tableTable 4. Training parameters after thawing
Parameter Content Size /(pixel×pixel) 416×416 Batch_size 8 Learning_rate 10－4 Optimizer Adam Epoch 80－200
Table 5. Precision and recall rate of different models on three defects
View tableTable 5. Precision and recall rate of different models on three defects
Model Precision Recall Pore Slag inclusion Incomplete penetration Pore Slag inclusion Incomplete penetration YOLOv5-Tiny-DIoU 85.3 88.2 84.6 89.2 91.2 70.8 YOLOv5-Tiny-CIoU 85.4 88.7 84.4 88.4 90.3 69.4 YOLOv5s 86.3 88.5 84.2 84.2 88.9 64.7 YOLOv4 85.4 86.9 87.9 46.5 65.9 21.1 YOLOv3 87.9 90.8 88.4 58.2 69.3 63.3 YOLOv4-Tiny 82.3 89.1 74.8 52.4 67.1 26.6
Table 6. Performance comparison of different models
View tableTable 6. Performance comparison of different models
Model F1 Parameters Weight /MB Detection speed /(frame·s－1) Pore Slag inclusion Incomplete penetration YOLOv5-Tiny-DIoU 0.87 0.90 0.77 4693936 9.2 67 YOLOv5-Tiny-CIoU 0.87 0.89 0.76 4693936 9.2 67 YOLOv5s 0.85 0.89 0.73 7068936 13.7 57 YOLOv4 0.60 0.75 0.34 64040001 244.4 42 YOLOv3 0.70 0.79 0.74 61949149 235.1 51 YOLOv4-Tiny 0.64 0.77 0.39 5961014 22.5 155
Table 7. Average precision (AP) and mean average precision (mAP) of different models
View tableTable 7. Average precision (AP) and mean average precision (mAP) of different models
Model AP mAP@0.5 Pore Slag inclusion Incomplete penetration YOLOv5-Tiny-DIoU 89.9 91 81.4 87.4 YOLOv5-Tiny-CIoU 89.8 91.2 80.6 87.2 YOLOv5s 88.2 91.1 77.1 85.5 YOLOv4 67.3 79.3 54.4 67.0 YOLOv3 75.2 86.1 75.3 78.8 YOLOv4-Tiny 63.9 77.6 45.6 62.3
Tools
Get Citation
Copy Citation Text
Song Cheng, Honggang Yang, Xueqian Xu, Min Li, Yunxia Chen. Improved Lightweight X-Ray Aluminum Alloy Weld Defects Detection Algorithm Based on YOLOv5[J]. Chinese Journal of Lasers, 2022, 49(21): 2104005
Paper Information
Category: Measurement and metrology
Received: Jan. 28, 2022
Accepted: Mar. 9, 2022
Published Online: Nov. 2, 2022
The Author Email: Chen Yunxia (chenyx@sdju.edu.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20