Research On Detection And Classification Of Steel Plate Surface Defects

Iron and steel industry plays an important role in China's economic industry.It can be called the mainstay of the manufacturing industry,and it is an intensive industry gathering huge capital and energy.Although our country produces a lot of steel every year,there are some problems in the quality of all kinds of steel products in our country.Due to various reasons,hotrolled steel plate in the production process will form a variety of defects on the surface,these defects will reduce the quality of products,affect the mechanical properties of products,and affect the appearance.In the quality of iron and steel products,the importance of surface quality is self-evident,but its improvement is difficult.It is of great significance to improve the surface quality of iron and steel products for improving the market competitiveness.It is an important prerequisite for improving the quality of steel plate to accurately detect and identify the surface defects of steel plate and establish an evaluation system for the surface defects of steel plate.However,it is difficult to detect the surface of hot-rolled steel plate.On the one hand,the environment of the hot-rolled production line is very bad,and the detection equipment is not easy to install and protect.On the other hand,there will be a lot of moisture on the hot-rolled surface,and the light is uneven,so the detection is difficult.At present,there are many researches on the surface detection of steel plate at home and abroad,as well as many detection methods of steel plate surface defects.Among them,the image detection method of steel plate surface defects based on computer technology and machine vision technology is becoming popular.At present,a popular detection mode is to use various image preprocessing technologies to preprocess the image,enhance the quality of the image,extract the relevant feature information from the image,reduce the dimension of the extracted features,and finally select the appropriate machine learning algorithm for training and testing.In this context,the key technology of steel plate surface defect detection is studied in this paper.This paper makes full use of the image data of steel plate surface defects and introduces the method of artificial intelligence to realize the classification recognition and target detection of steel plate surface defects.The specific research contents are as follows:(1)Consulting and learning a lot of literature and theoretical knowledge,investigating the actual site,finding out the problems existing in the process of steel plate surface defect detection,combining with the research status of steel plate surface detection technology at home and abroad,this paper puts forward a project of using machine learning and deep learning methods to achieve steel plate surface defect classification and target detection.(2)The feature extraction methods of local binary mode and local phase quantization are analyzed.Local binary mode can extract spatial domain information,local phase quantization can extract frequency domain information,and the fusion of them makes the feature information more comprehensive.After fusion,the statistical histogram is calculated,because of the selection of penalty parameters and kernel function parameters of support vector machine has a great influence on the classification results,so the optimization algorithm can be introduced to optimize the parameters,while the ant colony algorithm has a strong global optimization ability,and the particle swarm algorithm has a fast iterative speed,so the two are mixed to optimize the penalty parameters and kernel function parameters of the support vector machine,and then carry out the classification of steel plate surface defects,with an average recognition rate of up to about 94%.(3)Master the working principle of deep learning algorithm YOLO v3.With the increase of YOLO v3 accretion layer,the gradient will disappear.In order to reduce this phenomenon,remove the two layers of accretion layer in front of three yolo layers,and then detect the defect targets.The mAP value of each defect can reach 82.63%.In order to make the smaller targets detected by YOLO v3,three scales of YOLO v3 are increased to four scales for detection.The new 104 × 104 detection layer divides the grid into smaller ones.In order to extract more location information of the target in the shallow layer of the network,a residual unit is added to the first and second residual blocks respectively.In order to reduce the amount of calculation and parameter,a depthwise separable convolutions is introduced instead of the convolution structure of V3,and then detect the defect targets,the mAP value of each defect can reach84.04%.