自适应采样与重影多尺度特征融合的轻量化焊缝缺陷检测

doi:10.16182/j.issn1004731x.joss.24-0244

摘要/Abstract

摘要：

为提升焊接缺陷识别的准确率和速度，并实现模型的轻量化，提出了一种基于YOLOv8的轻量化焊缝缺陷检测网络LAW-YOLO(light adaptive-weight sampling-YOLO)。设计了一种轻量级自适应权重采样LAWS模块，通过学习感受野区域内交互的特征来构建自适应权重注意力特征图。采用优化的高效加权双向特征金字塔网络作为LAW-YOLO中的特征提取网络，设计重影多尺度采样模块并引用了混合注意力机制，以增强对小目标缺陷的检测能力。实验结果表明：该方法在SteelTube数据集中mAP0.5达到97.6%，处理数据速度可达91帧/s，比基线模型提高了5.5%的平均精度及4.6%的处理速度，在保持高效性能的同时减少了25.3%的计算量和50%的模型大小，更便于部署在边缘设备上进行场景作业。

关键词: 缺陷检测, YOLOv8, 重影多尺度卷积, 感受野空间特征, 混合注意力机制

Abstract:

To improve the accuracy and speed of welding defect detection and achieve lightweight models, a lightweight weld defect detection network based on YOLOv8, named light adaptive-weight sampling-YOLO (LAW-YOLO), was proposed. A lightweight adaptive weight sampling LAWS module was designed. It constructed an adaptive weight attention feature map by learning the interacting features within the receptive field.An optimized efficient weighted bidirectional feature pyramid network was adopted as the feature extraction backbone in LAW-YOLO. Furthermore, a ghost multi-scale sampling module was designed, and a hybrid attention mechanism was introduced to enhance the detection capability for small-scale defect targets. Experimental results demonstrate that the average precision (mAP0.5) of the proposed approach on the SteelTube dataset reaches 97.6%, with a data processing speed of 91 frames per second. The approach achieves a 5.5% increase in average defect recognition accuracy and a 4.6% improvement in processing speed compared to the original model and maintains high efficiency while reducing computation by 25.3% and model size by 50%, facilitating deployment on edge devices for operational tasks.

Key words: defect detection, YOLOv8, ghost multi-scale convolution, spatial feature of receptive field, hybrid attention mechanism

中图分类号:

TP391.41

鲁斌,杨烜,杨振宇等 . 自适应采样与重影多尺度特征融合的轻量化焊缝缺陷检测[J]. 系统仿真学报, 2025, 37(8): 1978-1990.

Lu Bin,Yang Xuan,Yang Zhenyu,et al . Adaptive Sampling and Ghost Multi-scale Fusion for Lightweight Weld Defect Detection[J]. Journal of System Simulation, 2025, 37(8): 1978-1990.

图/表 15

图1

图2

图3

图4

图5

图6

图7

表1

图8

图9

图10

表2

图11

表3

表4

参考文献 29

[1]	Sun Hongbin, Ramuhalli P, Jacob R E. Machine Learning for Ultrasonic Nondestructive Examination of Welding Defects: A Systematic Review[J]. Ultrasonics, 2023, 127: 106854.
[2]	王庆, 姚俊, 谭文禄, 等. 基于Faster R-CNN的排水管道缺陷检测研究[J]. 软件导刊, 2019, 18(10): 40-44, 49.
	Wang Qing, Yao Jun, Tan Wenlu, et al. Research on Defect Detection of Drainage Pipeline Based on Faster R-CNN[J]. Software Guide, 2019, 18(10): 40-44, 49.
[3]	孙志刚, 赵毅, 刘传水, 等. 基于深度学习的金属焊接管道内壁缺陷检测方法研究[J]. 焊管, 2020, 43(7): 1-7.
	Sun Zhigang, Zhao Yi, Liu Chuanshui, et al. Research on Inner Wall Defect Detection Method of Metal Welded Pipe Based on Deep Learning[J]. Welded Pipe and Tube, 2020, 43(7): 1-7.
[4]	Ren Shaoqing, He Kaiming, Girshick R, et al. Faster R-CNN: Towards Real-time Object Detection with Region Proposal Networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(6): 1137-1149.
[5]	李砚峰, 刘翠荣, 吴志生, 等. 基于深度学习的化工金属材料焊接小目标缺陷识别定位研究[J]. 材料保护, 2020, 53(8): 186.
[6]	胡宏伟, 张婕, 彭刚, 等. 基于LBP-KPCA特征提取的焊缝超声检测缺陷分类方法[J]. 焊接学报, 2019, 40(6): 34-39.
	Hu Hongwei, Zhang Jie, Peng Gang, et al. Defect Classification for Ultrasonic Inspection in Weld Seam Based on LBP-KPCA Feature Extraction[J]. Transactions of the China Welding Institution, 2019, 40(6): 34-39.
[7]	Bao Yanqi, Song Kechen, Liu Jie, et al. Triplet-graph Reasoning Network for Few-shot Metal Generic Surface Defect Segmentation[J]. IEEE Transactions on Instrumentation and Measurement, 2021, 70: 1-11.
[8]	Xu H, Yan Z H, Ji B W, et al. Defect Detection in Welding Radiographic Images Based on Semantic Segmentation Methods[J]. Measurement, 2022, 188: 110569.
[9]	Lin T Y, Dollár Piotr, Girshick R, et al. Feature Pyramid Networks for Object Detection[C]//2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2017: 936-944.
[10]	He Kaiming, Zhang Xiangyu, Ren Shaoqing, et al. Deep Residual Learning for Image Recognition[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2016: 770-778.
[11]	Redmon J, Farhadi A. Yolov3: An Incremental Improvement[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2018: 1-6.
[12]	Ultralytics. Ultralytics Yolov5[EB/OL].[2023-12-16]..
[13]	Ge Zheng, Liu Songtao, Wang Feng, et al. YOLOX: Exceeding YOLO Series in 2021[EB/OL]. (2021-08-06) [2023-10-16]. .
[14]	Wang C Y, Bochkovskiy A, Liao Hongyuan. YOLOv7: Trainable Bag-of-freebies Sets New State-of-the-art for Real-time Object Detectors[C]//2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2023: 7464-7475.
[15]	Reis D, Kupec J, Hong J, et al. Real-time Flying Object Detection with YOLOv8[EB/OL]. (2023-05-17) [2023-10-17]. .
[16]	Wan Dahang, Lu Rongsheng, Shen Siyuan, et al. Mixed Local Channel Attention for Object Detection[J]. Engineering Applications of Artificial Intelligence, 2023, 123, Part C: 106442.
[17]	Han Kai, Wang Yunhe, Tian Qi, et al. GhostNet: More Features from Cheap Operations[C]//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2020: 1577-1586.
[18]	Tan Mingxing, Pang Ruoming, Le Q V. EfficientDet: Scalable and Efficient Object Detection[C]//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2020: 10778-10787.
[19]	Kupyn Orest, Budzan Volodymyr, Mykhailych Mykola, et al. DeblurGAN: Blind Motion Deblurring Using Conditional Adversarial Networks[C]//2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2018: 8183-8192.
[20]	Hou Qibin, Zhou Daquan, Feng Jiashi. Coordinate Attention for Efficient Mobile Network Design[C]//2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2021: 13708-13717.
[21]	Zhang Xin, Liu Chen, Yang Degang, et al. RFAConv: Innovating Spatital Attention and Standard Convolutional Operation[EB/OL]. (2023-04-06) [2023-10-21]. .
[22]	Srinivas A, Lin T Y, Parmar N, et al. Bottleneck Transformers for Visual Recognition[C]//2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2021: 16514-16524.
[23]	Liu Shu, Qi Lu, Qin Haifang, et al. Path Aggregation Network for Instance Segmentation[C]//2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2018: 8759-8768.
[24]	Li Xiang, Chengqi Lü, Wang Wenhai, et al. Generalized Focal Loss: Towards Efficient Representation Learning for Dense Object Detection[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023, 45(3): 3139-3153.
[25]	Zhang Hao, Xu Cong, Zhang Shuaijie. Inner-IoU: More Effective Intersection over Union Loss with Auxiliary Bounding Box[EB/OL]. (2023-11-14) [2023-12-27]. .
[26]	Yang Dingming, Cui Yanrong, Yu Zeyu, et al. Deep Learning Based Steel Pipe Weld Defect Detection[J]. Applied Artificial Intelligence, 2021, 35(15): 1237-1249.
[27]	Fu Jianlong, Zheng Heliang, Mei Tao. Look Closer to See Better: Recurrent Attention Convolutional Neural Network for Fine-grained Image Recognition[C]//2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2017: 4476-4484.
[28]	Wang Qilong, Wu Banggu, Zhu Pengfei, et al. ECA-net: Efficient Channel Attention for Deep Convolutional Neural Networks[C]//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway: IEEE, 2020: 11531-11539.
[29]	Selvaraju R R, Cogswell M, Das A, et al. Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization[C]//2017 IEEE International Conference on Computer Vision (ICCV). Piscataway: IEEE, 2017: 618-626.

类别名	Label	原样本数	增强后样本数
气孔	Air-hole	1 339	11 018
咬边	Bite-edge	35	625
断弧	Broken-arc	531	6 459
裂纹	Crack	119	1 601
重叠	Overlap	219	4 869
夹渣	Slag-inclusion	136	1 480
未熔合	Unfused	416	3 614
空心珠	Hollew-bead	613	1 510

注意力机制	位置	精确率/%	召回率/%	mAP0.5/%	参数量/10⁶	FLOPs/10⁹
YOLOv8n		92.2	89.7	92.1	3.151 904	8.7
YOLOv8n+RA	Backbone	92.9	90.4	91.8	3.328 976	8.0
YOLOv8n+RA	Neck	93.1	88.2	92.4	3.319 040	8.3
YOLOv8n+CA	Backbone	92.8	91.1	91.8	3.285 792	9.0
YOLOv8n+CA	Neck	91.6	90.2	89.7	3.279 776	9.0
YOLOv8n+ECA	Backbone	92.8	90.8	92.4	3.219 680	8.8
YOLOv8n+ECA	Neck	93.1	91.3	92.6	3.219 680	8.8
YOLOv8n+MLCA	Backbone	92.9	90.8	92.7	3.151 938	8.7
YOLOv8n+MLCA	Neck	93.6	92.8	93.2	3.151 932	8.7

LAWS	GMSC	BIFPN	MLCA	Inner-CIoU	精确率/%	召回率/%	mAP0.5/%	FLOPs/10⁹	参数量/10⁶
					92.2	89.7	92.1	8.7	3.151 904
√					93.8	92.2	92.9	7.9	2.675 848
	√				93.6	90.2	92.7	7.6	2.862 974
		√			94.2	92.7	93.1	6.8	2.031 124
			√		93.6	91.8	92.2	8.7	3.151 932
√	√				94.4	92.6	93.8	7.5	2.387 336
√	√	√			96.1	93.5	96.4	6.5	1.616 692
√	√	√	√		96.4	94.1	97.1	6.5	1.616 795
√	√	√	√	√	96.8	95.3	97.6	6.5	1.616 795

算法名称	mAP0.5/%	精确率/%	召回率/%	参数量/10⁶	FLOPs/10⁹	帧率/(帧/s)
Faster RCNN	85.6	83.1	82.8	41.2	198.5	45
YOLOv5n	86.7	86.9	84.3	1.9	4.5	98
YOLOv5s	88.2	87.1	83.8	7.2	16.5	70
YOLOXs	91.3	89.1	86.9	9.0	26.8	60
YOLOv7-tiny	91.8	91.2	87.3	6.2	13.7	68
YOLOv8n	92.1	92.2	89.7	3.2	8.7	87
LAW-YOLO	97.6	96.8	95.3	1.6	6.5	91