摘要
针对金属化陶瓷环缺陷面积小、可利用信息少的特点和缺陷检测精度低的问题,提出一种目标检测与图像分类网络融合的金属化陶瓷环缺陷检测方法。首先,使用针对小面积目标检测特点改进的Faster-RCNN目标检测网络实现对缺陷的初步识别与定位。接着,使用插值方法将定位到的缺陷区域放大,利用图像相邻像素之间的信息关联,增加缺陷检测的特征信息量。然后,使用ResNet图像分类网络对放大后的区域进行缺陷类别判断。最后,融合目标检测网络和图像分类网络的结果,获得最终的缺陷检测结果。实验结果表明,所提方法能在保障缺陷检测查全率的同时有效提升查准率,且能准确定位缺陷区域。
Aiming at the characteristics of small defect areas and less available information of metalized-ceramic rings,and the problem of low defect detection accuracy,a defect detection method of metalized-ceramic rings based on the fusion of target detection and image classification networks is proposed.First,an improved Faster-RCNN target detection network for small-area target detection is used to realize the preliminary identification and location of the defects.Then,the interpolation method is used to enlarge the located defect area,and the information association between the adjacent pixels of the image increases the feature information of the defect detection.Moreover,the ResNet image classification network is used to judge the defect category of the enlarged area.Finally,the final defect detection results were obtained using the target detection and image classification network results.The experimental results show that the proposed method can effectively improve the precision while ensuring defect detection recall and accurately locate the defect area.
作者
满英杰
王宪
孙冬悦
邓宁道
吴士旭
Man Yingjie;Wang Xian;Sun Dongyue;Deng Ningdao;Wu Shixu(School of Mechanical Engineering,Hunan University of Science and Technology,Xiangtan 411201,Hunan,China;Changsha Shilang Technology Co.,Ltd.,Changsha 410006,Hunan,China)
出处
《激光与光电子学进展》
CSCD
北大核心
2023年第20期159-167,共9页
Laser & Optoelectronics Progress
基金
国家重点研发计划(2018YFB1308200)
国家自然科学基金(51405154)
湖南省自然科学基金(2021JJ30251,2018JJ3167)
国家留学基金委公派访问学者项目(202008430103)。
