一种基于卷积神经网络的管道焊缝图像识别算法  

A Pipeline Weld Image Recognition Algorithm Based on CNN

作　　者：杨中雨 李鹏 YANG Zhongyu;LI Peng(College Of Optical And Electronical Information Changchun University of Science and Technology,Changchun,Electronic Engineering Branch,Changchun 130000,China)

机构地区：[1]长春理工大学光电信息学院电子工程分院,长春130000

出　　处：《激光杂志》2021年第4期64-67,共4页Laser Journal

基　　金：国家自然科学基金(No.61703056);吉林省优秀青年人才基金项目(No.20190103154JH)。

摘　　要：为了提高管道焊缝图像缺陷检测的识别能力,提出了一种改进型卷积神经网络识别算法。该算法采用灰度拉伸与中值排序的方式完成图像预处理,在卷积神经网络中通过自适应矩估计的方式避免算法陷入局部最优解。实验针对3000张焊缝缺陷图片进行学习和训练,并与焊接异常图像识别的两种常用算法进行对比,结果显示,本算法对夹渣、裂纹、烧穿、气孔及未熔合五种常见缺陷的整体缺陷平均识别概率达90.4%,识别概率得到显著提升。在整体测试数据中,误检率、召回率及平均识别率均优于两种传统方法。验证了算法的可行性,具有更好的识别效果和应用前景。An improved convolutional neural network recognition algorithm is proposed in order to improve the recognition ability of pipeline weld seam image defect-detection.Grayscale stretching and median sorting methods are used for image preprocessing in this algorithm.The algorithm is prevented from falling into the optimal local solution through adaptive moment estimation in the convolutional neural network.In the convolutional neural network,the algorithm is prevented from falling into the optimal local solution by means of adaptive moment estimation.Three thousand weld seam defect pictures were learned and trained in the experiment compared with the two commonly used algorithms currently used for welding abnormal image recognition.Experimental results show that the algorithm has an average recognition probability of 90.4%for the five common defects,including slag,cracks,burn-through,pores,and unfused defects.The recognition probability has been significantly improved.In the overall test data,the false detection rate,recall rate,and average recognition rate are better than the two traditional methods.It verifies the feasibility of the algorithm,which has a better recognition effect and application prospect.

关 键 词：管道焊缝 缺陷识别 卷积神经网络 最优解 

分 类 号：TN212[电子电信—物理电子学]