Research on Defect Detection and Region Segmentation of Welding Forming based on Improved Mask R-CNN

LIU Jian; DUAN Ruibin; CUI Wanting; HE Yazhang; WANG Kekuan; WU Rongyao; GAO Jialin; XIA Weisheng

doi:10.7512/j.issn.1001-2303.2025.09.09

Welding Technology | Views : 0 下载量: 265 CSCD: 0

PDF
Export
Share
Collection
Album

Research on Defect Detection and Region Segmentation of Welding Forming based on Improved Mask R-CNN
Vol. 55, Issue 9, Pages: 98-101(2025)
Received：09 September 2024，

Revised：2024-10-24，

Published：20 September 2025
DOI： 10.7512/j.issn.1001-2303.2025.09.09
稿件说明：

移动端阅览

刘剑，段瑞彬，崔琬婷，等.基于改进Mask R-CNN的焊接成形缺陷检测与区域分割研究［J］.电焊机，2025，55（9）：98-101. DOI： 10.7512/j.issn.1001-2303.2025.09.09.

LIU Jian, DUAN Ruibin, CUI Wanting, et al.Research on Defect Detection and Region Segmentation of Welding Forming based on Improved Mask R-CNN[J].Electric Welding Machine, 2025, 55(9): 98-101. DOI： 10.7512/j.issn.1001-2303.2025.09.09.

摘要

针对焊接成形缺陷人工检测效率低、传统机器学习方法鲁棒性差，以及现有深度学习模型难以实现缺陷精确分割的问题，提出一种改进Mask R-CNN算法的焊接成形缺陷检测模型。首先，构建包含638组样本（驼峰304组、弧坑100组、飞溅234组）的焊接缺陷数据集，通过Labelme进行像素级掩码标注，并采用数据增强优化样本多样性。在Mask R-CNN骨干网络（ResNet）中嵌入CBAM混合注意力机制模块，通过通道与空间双维度自适应特征增强，提升模型对小目标缺陷（如飞溅）的敏感度，然后采用改进的Mask R-CNN算法进行训练，并进行算法性能验证。结果表明，与改进前Mask R-CNN模型相比，采用注意力机制改进之后的Mask R-CNN模型的

mAP

@0.5提高了3.34%，

mAP

@0.75提高了9.78%，

mAP

@0.5∶0.95提高了4.96%，能够更有效用于焊接成形缺陷识别与区域分割。

Abstract

To address the inefficiency of manual inspection

poor robustness of traditional machine learning methods

and the lack of precise segmentation in existing deep learning models for welding forming defects

this study proposes an improved Mask R-CNN-based detection model. A dataset of 638 samples (304 humps

100 craters

234 spatters) was constructed with pixel-level mask annotations using Labelme

enhanced through data augmentation. The CBAM hybrid attention module was embedded into Mask R-CNN's ResNet backbone

enabling dual channel-spatial attention mechanisms to improve sensitivity to small-target defects (e.g.

spatters). Experimental results demonstrate that the attention-enhanced model achieves 91.52%

mAP

@0.5 (+3.34%)

79.93%

mAP

@0.75 (+9.78%)

and 63.17%

mAP

@0.5:0.95 (+4.96%) compared to the baseline

confirming superior effectiveness for welding defect identification and region segmentation.

关键词

Keywords

references

肖润泉，许燕玲，陈善本，等 . 焊接机器人关键技术及应用发展现状［J］. 金属加工（热加工）， 2020 （ 10 ）： 24 - 31 .

XIAO R Q ， XU Y L ， CHEN S B ， et al . The key technology and application development of welding robot ［J］. MW Metal Forming ， 2020 （ 10 ）： 24 - 31 .

Yang L ， Wang H ， Huo B ， et al . An automatic welding defect location algorithm based on deep learning ［J］. NDT & E International ， 2021 ， 120 ： 102435 - 102445 .

刘凯 . 基于Faster R-CNN模型的焊缝缺陷目标检测的研究与实现［D］. 北京：北京邮电大学， 2020 .

LIU K . The Research and Implementation of Welding Defect Objectdetection based on Faster R-CNN Model ［D］. Beijing ： Beijing University of Posts and Telecom munications ， 2020 .

王睿，胡云雷，李海涛，等 . X射线焊缝图像缺陷实例分割算法［J］. 哈尔滨工业大学学报， 2022 ， 54 （ 05 ）： 140 - 145 .

WANG R ， HU Y L ， LI H T ， et al . Defect segmentation algorithm for X-ray weld images ［J］. Journal of Harbin Institute of Technology ， 2022 ， 54 （ 05 ）： 140 - 145 .

叶定剑，华学明，吕艳丽，等 . 高速焊接驼峰形成机理综述［J］. 热加工工艺， 2011 ， 40 （ 11 ）： 148 - 151+154 .

YE D J ， HUA X M ， LV Y L ， et al . Formation Mechanism of Humping at High Speed Welding ［J］. Hot Working Technology ， 2011 ， 40 （ 11 ）： 148 - 151+154 .

陈英红，杜明坤 . 基于Mask-RCNN与结构激光的纵焊缝5个参数检测方法［J］. 自动化与信息工程， 2018 ， 39 （ 05 ）： 18 - 23 .

CHEN Y H ， DU M K . A Mask-RCNN & Structured Light Based on Five Shape Parameters Detection Method for Longitudinal Welds ［J］. Automation & Information Engineering ， 2018 ， 39 （ 05 ）： 18 - 23 .

He K ， Gkioxari G ， Dollár P ， et al . Mask R-CNN ［J］. IEEE Transactions on Pattern Analysis & Machine Intelligence ， 2017 .

谢世朋，李博，张冬 . 基于改进Mask-RCNN的配电线路语义分割［J］. 南京邮电大学学报（自然科学版）， 2021 ， 41 （ 06 ）： 41 - 46 .

XIE S P ， LI B ， ZHANG D . Improved Mask⁃RCNN based semantic segmentation for distribution power lines ［J］. Journal of Nanjing University of Posts and Telecommunications：Natural Science Edition ， 2021 ， 41 （ 06 ）： 41 - 46 .

Woo S ， Park J ， Lee J Y ， et al . CBAM：Convolutional Block Attention Module ［C］// Proceedings of the European Conference on Computer Vision （ECCV）， 2018 ： 3 - 19 .

Ajmi C ， Zapata J ， Martínez-Álvarez J J ， et al . Using Deep Learning for Defect Classification on a Small Weld X-ray Image Dataset ［J］. Journal of Nondestructive Evaluation ， 2020 ， 39 （ 3 ）： 68 - 81 .

Alert me when the article has been cited

提交

Identification of MIG Welding Burn-through Defects in Stainless Steel Sheet Based on Machine Learning

Weld location and defect identification based on deep learning

Laser Welding of HR-2 Hydrogen-Resistant High-Strength Steel Pipes: Process Parameters and Mechanical Properties

Effects of Welding Thermal Cycle and Post-Weld Aging on the Microstructure and Mechanical Properties of GH2070P Fe-Ni-Based Superalloy Welded Joints

Research Status of Microstructure and Properties of HSLA Steel Welded Joints