Chinese Journal of Lasers, Volume. 52, Issue 8, 0802105(2025)
Real‑Time Weld Defect Detection Algorithm Based on YOLO‐DEFW
Figures & Tables(15)
Fig. 6. Image enhancement. (a) Elastic transform; (b) flip; (c) histogram equalization; (d) perspective transformation; (e) rotation; (f) sharpening; (g) affine transformation; (h) angular distortion; (i) color space transformation
Fig. 7. Comparison of confidence in detecting small cracks. (a) YOLOv8; (b) YOLOv8-DSConv-EMA
Fig. 8. Comparison of confidence in detecting porosities and spatters. (a) YOLOv8; (b) YOLO-DEFW
Fig. 9. Comparison in precision, recall, and mean average precision for different models
Table 1. Structure of YOLO-DEFW model
View tableTable 1. Structure of YOLO-DEFW model
Layer No. From n Parameters Module Arguments 0 -1 1 464 DSConv2D ［3， 16， 3， 2］ 1 -1 1 4672 DSConv2D ［16， 32， 3， 2］ 2 -1 1 5312 C2f_DSConv ［32， 32， 1， True］ 3 -1 1 18560 DSConv2D ［32， 64， 3， 2］ 4 -1 2 33280 C2f_DSConv ［64， 64， 2， True］ 5 -1 1 73984 DSConv2D ［64， 128， 3， 2］ 6 -1 2 132096 C2f_DSConv ［128， 128， 2， True］ 7 -1 1 295424 DSConv2D ［128， 256， 3， 2］ 8 -1 1 329216 C2f_DSConv ［256， 256， 1， True］ 9 -1 1 164608 SPPF ［256， 256， 5］ 10 -1 1 0 Upsample ［None， 2， 'nearest'］ 11 ［ 1 ，6 ］1 0 Concat ［ 1 ］12 -1 1 115456 C2f_DSConv ［384， 128， 1］ 13 -1 1 0 Upsample ［None， 2， 'nearest'］ 14 ［ 1 ，4 ］1 0 Concat ［ 1 ］15 -1 1 29056 C2f_DSConv ［192， 64， 1］ 16 -1 1 672 EMA ［64， 8］ 17 -1 1 36992 DSConv2D ［64， 64， 3， 2］ 18 ［ 1 ，12 ］1 0 Concat ［ 1 ］19 -1 1 90880 C2f_DSConv ［192， 128， 1］ 20 -1 1 2624 EMA ［128， 8］ 21 -1 1 147712 DSConv2D ［128， 128， 3， 2］ 22 ［ 1 ，9 ］1 0 Concat ［ 1 ］23 -1 1 361984 C2f_DSConv ［384， 256， 1］ 24 -1 1 10368 EMA ［256， 8］ 25 ［ 16 ，20 ，24 ］1 752482 Detect ［ 6 ， 64， 128， 256］
Table 2. Experimental parameter settings
View tableTable 2. Experimental parameter settings
Parameter name Definition Value batch Number of images per batch 48 workers Number of worker threads 8 lrf Final learning rate 0.01 box Box loss gain 8.0 cls Class loss gain 0.7 IoU Intersection over union （IoU） threshold 0.65 epochs Total number of training epochs 400 batch Number of images per batch 48
Table 3. Convolution performance experimental results
View tableTable 3. Convolution performance experimental results
Convolution name Output size Parameters GFLOPs Conv2D size （［1， 16， 64， 64］） 480 0.0020 Depth-Conv2D size （［1， 16， 64， 64］） 288 0.0013 Ghost-Conv2D size （［1， 16， 64， 64］） 336 0.0014 GSConv2D size （［1， 16， 64， 64］） 464 0.0020 DSConv2D size （［1， 16， 64， 64］） 32 0.0003 DCNv3 size （［1， 8， 64， 64］） 368 0.0015
Table 4. Attention mechanism performance experimental results
View tableTable 4. Attention mechanism performance experimental results
Model Precision Recall mAP YOLOv8 0.524 0.539 0.368 YOLOv8-GAM 0.507 0.548 0.391 YOLOv8-CBAM 0.515 0.587 0.394 YOLOv8-EMA 0.501 0.544 0.413
Table 5. Ablation experimental results of improved YOLOv8 models
View tableTable 5. Ablation experimental results of improved YOLOv8 models
Model Parameters Precision Recall mAP Epochs GFLOPs Faster R-CNN 25600000 0.415 0.649 0.559 400 45.5 YOLOv5 2222064 0.530 0.575 0.423 400 6.1 YOLOv8 3006818 0.531 0.604 0.468 400 8.1 YOLOv8-DSConv 2600642 0.518 0.610 0.440 400 7.0 YOLOv8-DSConv-EMA 2600642 0.614 0.695 0.562 400 7.3 YOLOv8-DSConv-EMA-FWCE Loss-WCE （YOLO-DEFW） 2600642 0.602 0.707 0.584 400 7.3
Table 6. Comparison of experimental results between YOLOv8 and YOLO- DEFW for detecting various types of defects
View tableTable 6. Comparison of experimental results between YOLOv8 and YOLO- DEFW for detecting various types of defects
Defect YOLOv8 YOLO- DEFW P R mAP P Growth ratio of P R Growth ratio of R mAP Growth ratio of mAP All 0.531 0.604 0.468 0.602 13% 0.707 17% 0.584 25% Bad welding 0.652 0.582 0.483 0.735 13% 0.698 20% 0.601 24% Crack 0.284 0.583 0.282 0.331 17% 0.706 21% 0.384 36% Incomplete fusion 0.644 0.588 0.442 0.715 11% 0.650 11% 0.594 34% Good welding 0.666 0.590 0.494 0.745 12% 0.750 27% 0.636 29% Porosity 0.315 0.750 0.678 0.370 17% 0.857 14% 0.762 12% Spatters 0.605 0.480 0.350 0.720 19% 0.583 21% 0.524 50%
Tools
Get Citation
Copy Citation Text
Jianfeng Yue, Weiming Li, Lihua Ning, Xingyu Gao, Yu Li, Wenlong Wang, Baiqing Yang, Yani Liu. Real‑Time Weld Defect Detection Algorithm Based on YOLO‐DEFW[J]. Chinese Journal of Lasers, 2025, 52(8): 0802105
Paper Information
Category: Laser Forming Manufacturing
Received: Sep. 20, 2024
Accepted: Nov. 25, 2024
Published Online: Mar. 17, 2025
The Author Email: Weiming Li (liaifan@buaa.edu.cn), Lihua Ning (nlhua123@sina.com.cn)
CSTR:32183.14.CJL241223
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20