Research On Visual Detection Of Surface Defect Of Steel Rope

As a load-bearing components,steel ropes are widely used in metallurgy,mining,construction,aerospace and other fields.However,the accidents of steel rope rupture suddenly occurs from time to time,which endangered the production safety seriously.This paper studied the machine vision online inspection technology for steel rope surface defects,which has very important engineering application value,the main studies are as follows:(1)Bacause of the shaking or swinging,the images of steel rope will produce blurring,this paper adopts DeblurGAN(Deblur Generative Adversarial Network)to deblur the steel rope images collected in real time,which improves the quality of the steel rope images effectively.(2)Based on this,the paper proposes two detection schemes according to the structural characteristics of steel ropes and the type of damage.The first method is use deep learning target detection.Considering that most of the steel rope surface defects are small targets and in order to improve the speed of image recognition,this paper combines the lightweight Mobilenet network with the feature pyramid of YOLOV3 network,deletes the 13x13 feature map used for large target detection in the pyramid structure to form a target detection model applicable to steel rope defects.(3)Another detection method is to use the obvious texture features on the surface of steel ropes,this paper raises a machine learning recognition model for steel rope defects with feature fusion+heterogeneous integrated learning method.The gradient direction histogram features and the uniform local binarization pattern are fused,and then the features of steel rope pictures via principal component analysis dimension reduction,when judging the final result based on the results of three different learning models:K-nearest neighbor,support vector machine and decision tree,use the voting method to determine the final detection result.The above two methods were tested,and the speed of deep learning target detection model reached 0.05s with 93.3%detection accuracy;the detection time of heterogeneous integrated learning model was 0.18s with 93.6%correct rate.