Faster R-CNN定位后的工业CT图像缺陷分割算法研究  

Research on defect segmentation algorithm of industrial CT image after Faster R-CNN positioning

作　　者：吴晓元 常海涛 苟军年[1,2] Wu Xiaoyuan;Chang Haitao;Gou Junnian(School of Automation and Electrical Engineering,Lanzhou Jiao-tong University,Lanzhou 730070,China;Key Laboratory of Optoelectronic Technology and Intelligent Control Ministry of Education,Lanzhou Jiao-tong University,Lanzhou 730070,China)

机构地区：[1]兰州交通大学自动化与电气工程学院,甘肃兰州730070 [2]兰州交通大学光电技术与智能控制教育部重点实验室,甘肃兰州730070

出　　处：《电子技术应用》2019年第1期76-80,共5页Application of Electronic Technique

基　　金：光电技术与智能控制教育部重点实验室(兰州交通大学)开放课题(KFKT2018-14)

摘　　要：由Faster R-CNN定位的缺陷区域内存在弱边缘,若直接采用常规分割算法对该小区域进行处理,会出现严重的过分割或欠分割现象。在此研究了一种针对Faster R-CNN定位后的工件缺陷的精确阈值分割法。在利用形态学开闭重建算法对定位区域进行重建,并对重建后的图像用Otsu双阈值法做变换处理的基础上,进一步利用最大熵阈值分割法对变换后的图像进行分割,最终对分割出的缺陷进行面积、周长等参数的测量。实验结果表明,所研究算法较常规的算法对工件的缺陷(裂纹、气泡和夹渣)有更好的分割能力。该算法不仅可以准确地分割出包含弱边缘的目标,还可以有效排除轮廓背景对分割的干扰。The defect area located by Faster R-CNN has weak edges.The area would be over-segmented or under-segmented if conventional segmentation algorithm is adopted.This paper made an analysis on precise threshold segmentation algorithm for workpiece defects based on Faster R-CNN location,reconstructing the localization area by morphological opening and closing reconstruction algorithm,processing the reconstructed image by Otsu′s dual threshold method,segmenting transformed images by maximum entropy threshold segmentation method,and finally measuring the area,perimeter and other parameters of the segmented defects.The research shows that the algorithm in this paper has higher segmentation ability regarding workpiece defects(crack,bubble and slag),compared to conventional algorithms.It not only can accurately segment objects with weak edges,but also can effectively remove the interference from the contour background to the segmentation.

关 键 词：FasterR-CNN 缺陷分割 形态学开闭重建算法 Otsu双阈值法 最大熵阈值分割法 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]