ECA注意力机制增强的轻量级网络焊接熔池轮廓提取方法  

Enhanced Lightweight Network for Melt Pool Contour Extraction with ECA Attention Mechanism

作　　者：张昆 范东阳 袁飞 黄勇[3] 李晓鹏 ZHANG Kun;FAN Dongyang;YUAN Fei;HUANG Yong;LI Xiaopeng(The first military Representative Office of Naval Armament Department in Shanghai,Shanghai 201913,China;Jiangnan Shipyard(Group)Co.,Ltd.,Shanghai 201913,China;School of Materials Science and Engineering,Nanjing University of Science and Technology,Nanjing 210094,China)

机构地区：[1]海军装备部驻上海地区第一军事代表室,上海201913 [2]江南造船(集团)有限责任公司,上海201913 [3]南京理工大学材料科学与工程学院,江苏南京210094

出　　处：《电焊机》2024年第12期28-34,共7页Electric Welding Machine

摘　　要：熔池形貌特征在线检测是实现机器人焊接智能化与自动化的重要途径之一。然而,目前的熔池轮廓提取方法多使用基于阈值分割的传统处理算法,这类算法在弧光干扰的情况下,熔池轮廓准确度低、鲁棒性差;而基于深度学习对熔池轮廓进行提取的报道还较少,并且在实时性方面存在一定的挑战。针对这一问题,本文提出一种改进的DeepLabV3+的高效语义分割模型。该模型采用轻量级的MobileNetv2主干网络,降低模型复杂度,提高推理速度;并在ASPP特征提取模块后添加ECA注意力机制,增强网络对熔池特征的关注,提升分割精度;针对熔池图像数据集中前景和背景像素数量不均衡的问题,采用Dice Loss和交叉熵损失函数的线性组合作为新的损失函数,改善模型训练效果。实验结果表明,该模型在熔池轮廓提取任务中性能优异,平均交并比(mIoU)达到96.08%,类别平均像素准确率(mPA)为97.85%,推理时间由60.72 ms缩短到23.11 ms,满足了熔池轮廓提取准确性与实时性的要求。Online detection of arc additive molten pool morphology features is an important way to achieve system intelli‐gence and automation.However,current methods for extracting molten pool contours mostly use traditional processing algo‐rithms based on threshold segmentation.These algorithms have low accuracy and poor robustness in the presence of arc in‐terference.Reports on using deep learning for molten pool contour extraction are relatively rare,and there are certain chal‐lenges in terms of real-time performance.To address this issue,this paper proposes an improved efficient semantic segmenta‐tion model based on DeepLabv3+for molten pool contour extraction.Experimental results show that the proposed algorithm achieves the mean Intersection over Union(mIoU)of 96.08%and the mean pixel accuracy(mPA)of 97.85%,while reduc‐ing inference time from 60.72 ms to 23.11 ms,thus meeting the requirements for accurate and real-time molten pool contour extraction.

关 键 词：语义分割 熔池轮廓提取 轻量级网格 ECA注意力机制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TG409[自动化与计算机技术—计算机科学与技术]