Research On Detection And Classification Method Of Surface Quality Defects In Strip Steel

As one of the important products in the steel industry,strip steel has been widely used in the fields of electrical and electronic,mechanical manufacturing,aerospace industry,etc.Therefore,the quality of the surface of the strip is a necessary indicator to determine the quality of its products.However,in the production process,it will be affected by many factors such as equipment on the production line,specific processing technology and the entire production space environment.In the production process,the surface of the strip will have defects caused by various problems such as side waves,inclusions,holes,scratches and stains,and these defects will seriously affect the quality of the strip products.Therefore,how to effectively and accurately detect the defects appearing on the surface of the steel strip has become a problem that needs to be solved to improve the quality of the steel strip,and has important theoretical research value.In this context,the pretreatment techniques,feature extraction and feature selection techniques and final classification methods used to deal with images of surface quality defects of strip steel are studied.main tasks as follows:(1)In this paper,the following three aspects of the strip surface defect image preprocessing technology are studied: Firstly,in order to avoid the noise interference in the image,by comparing the four commonly used denoising methods,the median filtering with better defect retention effect is selected to remove image denoising.Secondly,in order to make the defect target and the background better distinguish in the defect image,the image enhancement method is adopted,and the method of homogenizing the histogram which makes the pixel distribution more uniform is selected.Finally,in order to separate the defect area from the image,the image segmentation method is used to compare the four commonly used segmentation methods,and the Canny operator with the most obvious segmentation effect is selected for image segmentation.(2)In order to transform image visual information into data information,this paper adopts the feature extraction method,and extracts 154-dimensional features including geometric features,gray features and texture features from the pre-processed images.In order to avoid over-fitting and to improve the accuracy of classification,the method of principal component analysis,the method of Relief F algorithm and the method of Relief F algorithm combined with PCA are adopted.These three methods perform feature dimension reduction and feature selection processing on the feature dataset.The processed features are 25-dimensional,25-dimensional and 22-dimensional,respectively,which reduces the subsequent calculation.(3)In this paper,the three kinds of data results after dimension reduction and selection are classified by support vector machine,K-nearest neighbor algorithm,random forest algorithm,BP neural network and RBF neural network.By comparing the accuracy of the classification and the length of time,it is concluded that the accuracy of PCA combined with SVM is the highest,and the accuracy is 96.25%.However,in terms of comprehensive accuracy and time,random forests not only have higher classification accuracy but also shorter time.