箱梁结构件焊缝表面缺陷特征提取及分类研究  

Feature extraction and classification of weld surface defects in box girder structures

作　　者：葛为民 申耀华[1] 王肖锋 Ge Weimin;Shen Yaohua;Wang Xiaofeng(Tianjin Key Laboratory of the Design and Intelligent Control of the Advanced Electromechanical System,Tianjin 300384,China;National Experimental Teaching Demonstration Center of Electromechanical Engineering of Tianjin University of Technology,Tianjin 300384,China)

机构地区：[1]天津市先进机电系统设计与智能控制重点实验室,天津300384 [2]天津理工大学机电工程国家级实验教学示范中心,天津300384

出　　处：《仪器仪表学报》2018年第12期207-215,共9页Chinese Journal of Scientific Instrument

基　　金：国家重点研发计划(2017YFB1303304);天津市科技计划(17ZXZNGX00110);天津市自然科学基金重点项目(16JCZDJC30400)资助.

摘　　要：针对箱梁结构件焊缝表面缺陷分类识别过程中的在线特征提取及实时性等问题,在二维主成分分析(2DPCA)的基础上提出了一种基于泛化的增量式2DPCA(GI2DPCA)的特征提取算法。首先,对获取的箱梁结构件焊缝表面图像进行图像增强、图像平滑和阈值分割等预处理,实现了焊缝缺陷区域的有效信息提取,建立了焊缝缺陷数据库;然后,利用所提出的GI2DPCA算法对数据库中的焊缝图像提取其特征主元,并通过BP神经网络进行缺陷分类识别;最后,在建立的焊缝缺陷数据库上进行了算法性能的对比实验,并在ORL人脸数据库上对所提算法的通用性进行了验证。结果表明, GI2DPCA+BP神经网络算法的提取及分类速度可达36 fps,识别率达到97%,能够满足箱梁结构件焊缝表面缺陷检测的工程应用及实时性处理需求。Aiming at the problems that the online feature extraction and real-time in the process of classification and identification of weld surface defects in box girder structural parts, based on two-dimensional principal component analysis(2DPCA), a generalized incremental 2DPCA(GI2DPCA) algorithm is presented. Firstly, the image of the weld surface of the obtained box girder structure is preprocessed by image enhancement, image smoothing and threshold segmentation, and the effective information extraction of the weld defect area is realized. Thus, a database of weld defects was established. Then, using the proposed GI2 DPCA algorithm to extract the feature principals of the weld image in the database, and identify the defects by BP neural network. Finally, the comparison experiments of algorithm performance were carried out on the established weld defect database, and the versatility of the proposed algorithm was verified on the ORL face database. Experimental results show that the speed of feature extraction and classification based on GI2DPCA+BP can reach 36 frames per second, and the recognition rate can reach 97%, which can meet the engineering application and real-time processing requirements of surface defect detection of box girder structural parts.

关 键 词：焊缝表面缺陷 特征提取 二维主成分分析 BP神经网络 分类识别 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TG441.7[自动化与计算机技术—计算机科学与技术]