基于MCL-UNet网络的激光熔覆层表面平整度识别  

Surface smoothness identification of laser cladding layer morphology based on MCL-UNet network

作　　者：郭士锐[1] 刘银银 崔陆军[1] 陈永骞 郑博 GUO Shi-rui;LIU Yin-yin;CUI Lu-jun;CHEN Yong-qian;ZHENG Bo(School of Mechanical and Electrical Engineering,Zhongyuan University of Technology,Zhengzhou 450007,China)

机构地区：[1]中原工学院机电学院,河南郑州450007

出　　处：《激光与红外》2023年第5期685-692,共8页Laser & Infrared

基　　金：河南省研究生教育改革与质量提升工程项目(No.YJS2022AL057);中原工学院优势学科实力提升计划资助“学科骨干教师支持计划”项目(No.GG202220)与“骨干学科发展计划”项目(No.FZ202204);中原工学院研究生校企联合课程专项经费资助建设项目(No.LH202301);河南省重点研发与推广专项(科技攻关)项目(No.232102220051)资助。

摘　　要：表面平整度是衡量多道搭接熔覆层表面质量的重要指标之一,为改进人工标注获取表面平整度耗时费力的问题,本文利用图像识别和语义分割神经网络方法提出了自动识别熔覆层表面平整度。针对搭接熔覆层特征,基于改进的U-Net与注意力机制(CBAM)提出一种用于熔覆层形貌的自动分割网络MCL-UNet,优化改进U-Net模型,依据CBAM模块从通道维度和空间维度调整特征图层的权重信息,将CBAM模块以优化输入和输出的原则部署在网络上。在搭接熔覆层数据集上对改进的网络进行评估对比,实验结果表明,本文提出的MCL-UNet网络模型,其熔覆层分割效果在验证集上的平均IoU准确率为93.76%,相比原始U-Net的IoU准确率提高了5.81%,在测试集上MCL-UNet模型输出的表面平整度的平均相对误差为3.2%,说明该模型可有效提高搭接熔覆层横截面形貌的分割精度,并能较准确输出表面平整度。Surface smoothness is one of the most important indicators of the surface quality of multi-pass clad layer.In order to improve the time-consuming and laborious problem of manual annotation to obtain surface smoothness,an automatic identification method of surface smoothness of clad layer using image recognition and semantic segmentation neural network is proposed in this paper.An automatic segmentation network MCL-UNet for multi-pass clad layer morphology based on convolutional neural network U-Net and CBAM is presented.The weight information of the feature layer from channel dimension and spatial dimension is adjusted based on the CBAM module,the CBAM is deployed on the U-Net network with the principle of optimizing the input and output,and the evaluation of improved networks compared on a clad layer dataset is performed.The experimental results show that the MCL-UNet network achieves the average IoU accuracy of 93.76%on the validation dataset,which is 5.81%higher than the original U-Net.The average relative error of the surface smoothness by the MCL-UNet model on the test dataset is 3.2%,indicating that the model can effectively improve the segmentation accuracy of the cross-sectional morphology of the multi-pass clad layer,and can accurately output the surface smoothness.

关 键 词：图像处理 表面平整度 语义分割 MCL-UNet网络 

分 类 号：TP302.1[自动化与计算机技术—计算机系统结构] TN249[自动化与计算机技术—计算机科学与技术]