Research On Surface Defect Detection Of Metal Strip Based On Image Processing And Capsule Network

Strip is one of the main products of the metal industry and is widely used in aerospace,automobile manufacturing,electronic and electrical,construction and other product fields.With the rapid development of China’s economy,its demand is increasing.However,in the production process of metal strips,due to factors such as the production environment,equipment,and technology,it will cause defects that are difficult to ignore on the surface of the finished product.Not only affects the aesthetics of the product,but also the performance of the product,which causes serious accidents in subsequent use.Therefore,how to accurately and efficiently detect defects on the surface of the metal strip has become a problem that needs attention in the metal rolling process.The machine vision-based surface defect detection method uses computer simulation of human vision,extracts useful information from objective images,and then performs image processing and pattern recognition to achieve defect detection.The key technology in the inspection system is the intelligent algorithm of image processing and defect recognition,which determines the performance of the entire inspection system.Therefore,this paper studies the image processing,defect detection and defect classification algorithms in the surface defect detection problem of metal strip.The research contents of this article are as follows:(1)Research on Image Processing of Surface Defects of Metal Strips.Based on the study of traditional image enhancement methods,based on the characteristics of non-uniform areas prone to surface defect images of metal strips,an image enhancement method of metal strip surface defects was proposed.This method can reduce the influence of noise caused by the reflection of the material surface on image analysis,and improve the grayscale contrast between the defect area and the background area.(2)Research on defect detection of metal strip surface defect image.Based on the processed gray image,several threshold segmentation methods are compared,and the maximum correlation criterion segmentation method is used for experimental comparison.Finally,the defect area is marked and segmented by the minimum outer rectangle method.(3)Classification study of surface defects on metal strips.By analyzing the existing defect classification technologies,using deep learning-based methods for defect identification,and focusing on the capsule network algorithm.Aiming at the shortcomings of the capsule network,the training parameters,network structure,and network convergence speed were studied.Finally,a metal strip defect recognition classifier based on the improved capsule network was designed.The experimental results on copper strip and steel strip images show that the proposed method for the enhancement and segmentation of defect images of metallic strips not only improves the contrast and entropy of the defect images,but also obtains a binary image with the optimal error segmentation rate.The classification method proposed in this paper makes the accuracy of defect classification reach more than 97%.