基于SegFormer的钛板缺陷涡流C扫描检测图像分割  

Image Segmentation for Eddy Current C-scan Detection of the Titanium Plate Defects Based on SegFormer

作　　者：李肇源 叶波 邹杨坤[2,3] 包俊 曹弘贵[1,2] LI Zhao-yuan;YE Bo;ZOU Yang-kun;Bao Jun;CAO Hong-gui(Faculty of Information Engineering and Automation,Kunming University of Science and Technology;Key Laboratory of Artificial Intelligence of Yunnan Province,Kunming University of Science and Technology;Faculty of Civil Aviation and Aeronautics,Kunming University of Science and Technology)

机构地区：[1]昆明理工大学信息工程与自动化学院 [2]昆明理工大学云南省人工智能重点实验室 [3]昆明理工大学民航与航空学院

出　　处：《化工自动化及仪表》2024年第2期181-191,共11页Control and Instruments in Chemical Industry

基　　金：云南省基础研究计划(批准号:202301AS070052)资助的课题;云南省中青年学术和技术带头人后备人才基金(批准号:202305AC160062)资助的课题。

摘　　要：为了获得涡流检测图像中缺陷的形状与长度信息,对检测图像进行图像分割是其中一种重要的方法。由于边缘效应的影响,TA2钛板涡流C扫描图像中缺陷区域边缘模糊、对比度低,导致通过图像分割后估计的缺陷长度与实际值差别过大,难以通过图像分割方法对缺陷长度进行准确估计。针对此问题,提出一种基于SegFormer的钛板缺陷涡流检测图像分割方法。首先,利用涡流C扫描成像获得TA2钛板表面裂纹缺陷检测图像数据集;根据SegFormer网络的结构组成,设计基于SegFormer钛板缺陷涡流检测图像的分割框架和分割流程,并进行网络参数的设置。随后,利用钛板表面裂纹缺陷检测图像数据集对4个分割模型分别进行训练和测试,并利用平均交并比、平均精度和训练时间对其分割效果进行评价。实验表明:相比于Deeplabv3+、Swim Transformer和OCRNet,SegFormer具有更好的分割效果,更快的训练速度。可视化和定量化结果表明:与非深度学习方法相比,该方法具有更小的缺陷长度估计误差。For purpose of obtaining both shape and length information of the defects in eddy current inspection images,image segmentation becomes an important method.Due to the influence of edge effects,the edge fog and low contrast ratio of the defect area in eddy current C-scan image of TA2 titanium plate causes a significant difference between the estimated defect length after image segmentation and the actual value and it's difficult to accurately estimate defect length through image segmentation methods.In response to this issue,a SegFormer-based image segmentation method for eddy current testing of the titanium plate defects was proposed.Firstly,having the eddy current C-scan imaging used to obtain a dataset of surface crack defect detection images for TA2 titanium plates;and then,having the structural composition of the SegFormer network based to design a segmentation framework and process for defect eddy current detection images of titanium plates and setting network parameters;and finally,having the titanium plate surface crack defect's detection image dataset employed to train and test four segmentation models,including having average intersection and union ratio,average accuracy and training time adopted to evaluate their segmentation effects.Experimental results show that,as compared to Deeplabv3 +,Swim Transformer and OCRNet,the SegFormer has better segmentation performance and faster training speed.The visualization and quantification results indicate that,compared with non-deep learning methods,this scheme has smaller defect length estimation error.

关 键 词：深度学习 SegFormer 钛板缺陷 电涡流检测 图像分割 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]