Label Defect Detection And Content Analysis Of A Low-voltage Circuit Breaker Based On Machine Vision

Low-voltage circuit breaker is the basic equipment of power transmission and distribution engineering,and its quality problems seriously affect the safety and efficiency of the engineering.In the production of enterprises,the traditional way of manually inspecting the appearance and assembly quality of low-voltage circuit breaker has some problems,such as low detection efficiency,poor stability,high labor cost,and so on.In view of this,under the background of the “Industry 4.0”,we take the low-voltage circuit breaker as research object,and use machine vision technology to develop an automatic inspection system for appearance and assembly quality.Among many testing contents,the laser engraving label(nameplate)of low-voltage circuit breaker is marked with important attribute information of the product.Therefore,in order to ensure that the attribute information meets the factory requirements,this paper focuses on the label appearance defect detection and content analysis algorithm.The research works are as follows:(1)Design the detection system and preprocess the image.In the aspect of hardware,a highdefinition imaging platform based on industrial robot technology is designed,and the imaging system is calibrated.In the aspect of software,modular design idea is adopted,visual inspection interface is designed with C++ language,and the task logic of the industrial application is constructed with C# language to complete software development.In the aspect of image preprocessing,a label background removal algorithm based on local mean of the image is proposed,and the perspective correction algorithm of label image with or without prior knowledge is studied in combination with industrial field application and label structure characteristics.(2)Design label appearance defect detection algorithms.According to the detection requirements,the defect detection algorithms of label reprinting,label offset,frame breakage,row tilt,character adhesion,character breakage and character offset are designed.Our main contributions are as follows: We proposed an improved least square line fitting algorithm based on outlier analysis in row tilt detection.In order to solve the problem of pseudo-adhesion of characters,a Mask-Otsu threshold segmentation algorithm is proposed based on the research of traditional Otsu algorithm.In view of the large number of characters in label,large size differences,random fracture morphology and other factors,we construct a character breakage detection network Le Net-CB with the reference of Le Net-5 network and the design idea of VGG network.Experiment shows that the test accuracy of Le Net-CB is 99.65%,and the average time consumption of single character detection is 18.2ms,which meets the requirements of detection indexes.(3)Design label content analysis algorithms.Firstly,we designed a character recognition network Le Net-SE with "attention" mechanism in the feature map channel based on SENet.Experiments show that the average recognition accuracy of Le Net-SE is 99.25%.However,the test shows that it has insufficient ability to classify confusing characters(such as characters "0","O" and "o" in different sizes,etc.).Therefore,this paper finally designed a solution based on TesseractOCR framework,and the experiment shows that the average identification accuracy of label electrical parameter is 99.60%,and the average time consumption of single parameter is 83.7ms,which meets the requirements of detection indexes.Although the Le Net-SE network has not been applied to this research in the end,it provides a reference solution for other confusion-free character recognition tasks(such as wire harness number tube recognition,etc.).