改进RCF算法的电缆绝缘层边缘检测  

Edge detection of cable insulation based on improved RCF algorithm

作　　者：翁玉尚 肖金球[1,2] 汪俞成 焦文开 WENG Yushang;XIAO Jinqiu;WANG Yucheng;JIAO Wenkai(College of Electronics and Information Engineering,Suzhou Univers让y of Science and Technology,Suzhou 215009,China;Intelligent Measurement and Control Engineering Technology Research Center,Suzhou University of Science and Technology,Suzhou 215009,China)

机构地区：[1]苏州科技大学电子与信息工程学院,江苏苏州215009 [2]苏州市智能测控工程技术研究中心,江苏苏州215009

出　　处：《光学技术》2022年第1期86-92,共7页Optical Technique

基　　金：江苏省产学研前瞻性联合项目基金(BY2011132);江苏省研究生创新与教改项目(09150001);苏州科技大学研究生创新工程基金(SKCK17_025)。

摘　　要：目前电缆绝缘层厚度检测算法主要采用图像处理技术提取出绝缘层的边缘轮廓,此类算法存在绝缘层边缘过宽和边缘不连续等问题,影响了后续的检测精度。为提高绝缘层测量精度,新算法基于RCF算法进行改进,在模型的4、5阶段采用空洞卷积,增大模型的感受野;并在侧输出网络加入尺度增强模块(SEM模块)和由浅到深的级联网络,增加侧输出图像的细节信息。通过自制的电缆绝缘层数据集对模型进行训练,结果表明改进后的模型在数据集最优尺度(ODS)和单张图片最优尺度(OIS)分别为0.821和0.842,平均精度为0.799,算法相较于RCF模型ODS和OIS分别提高了0.008和0.01,检测精度提升了0.021。并在伯克利大学数据集(BSD500)数据集上对模型的性能进一步验证,其中ODS和OIS分别为0.810和0.825,所提算法相较于RCF模型ODS和OIS分别提高了0.009和0.006。The current cable insulation layer thickness detection algorithm mainly uses image processing technology to extract the edge contour of the insulation layer.Such algorithms have problems such as excessively wide insulation layer edges and discontinuous edges,which affect the subsequent detection accuracy.In order to improve the measurement accuracy of the insulation layer,new algorithm is based on the RCF(Richer Convolutional Features) algorithm to improve,in the 4 th and 5 th stages of the model,the cavity convolution is used to increase the receptive field of the model;and the scale enhancement module(SEM module) is added to the side output network.And the cascade network from shallow to deep to increase the detailed information of the side output image.The model was trained through the self-made cable insulation data set.The results show that the improved model has 0.821 and 0.842 in the optimal scale of the data set(ODS) and the optimal scale of a single picture(OIS),respectively,and the average accuracy is 0.799.Compared withthe RCF model ODS and OIS,the algorithm is improved by 0.008 and 0.01 respectively,and the detection accuracy is improved by 0.021.The performance of the model is further verified on the Berkeley University Data Set(BSD500)data set,where ODS and OIS are 0.810 and 0.825,respectively.Compared with the RCF model,the ODS and OIS of this algorithm are improved by 0.009 and 0.006,respectively.

关 键 词：电缆绝缘层边缘检测 深度学习 空洞卷积 多尺度模块 级联网络 

分 类 号：TP394.1[自动化与计算机技术—计算机应用技术]