Steel Plate Defect Detection And Classification Based On Deep Learning

There is a huge demand for steel in modern society,especially for China,which is in rapid development.We need a lot of steel to build infrastructure.The surface defects of steel plate will seriously affect the strength and properties of steel plate.In order to improve the quality of steel plate and product efficiently,it is necessary to build an efficient defect detection system.This paper investigates the development history of steel plate surface defect detection at home and abroad,and finds that using machine vision method for defect detection and classification is the best solution.On this basis,according to the collected steel plate surface defect image,the relevant detection algorithm is studied.The image processing algorithms include:defect image screening,defect image segmentation,defect image classification and defect detection using SSD(single shot detection).For the defect image has more high-frequency information,a defect screening algorithm based on image gradient is proposed,which can accurately distinguish the defect image from the normal image.The algorithm runs fast,which is of great significance for real-time detection system.The local entropy of the image is used to segment the defect image,and the histogram back projection of the image is used to improve the algorithm,so that the algorithm can locate the defect more accurately.To solve the time-consuming problem of local histogram algorithm,the histogram is updated line by line to reduce the time complexity of the algorithm from(9)~2)to(9)).The convolution neural network is applied to the classification of steel plate defect image,which saves the complex feature extraction process,and the classification effect is better than the traditional classification algorithm.SSD algorithm is used to detect the surface defects of steel plate.Aiming at the problems of image quality collected in actual production,seven data enhancement methods are proposed,which can increase the diversity of data sets and avoid over fitting of models.