Research On Recognition Of Steel Strip Surface Defects Based On Machine Vision

With the rapid development of society, people have a higher requirement in product line of manufacturing. It is necessary to increase productivity but also guarantee quality. Due to steel strips are the original materials in automobile manufacturing, aviation, daily necessities, its quality requriments are more and more strict. However, in cicil iron and steel company, the economic losses are still high in production line of steel strips. The main reason is that problems during manufacturing band steels are not controlled in real time.In this paper, we put forward an approach to monitor the production line using machine vision, replacing the method depends on workers to supervise quality. Through the camera to capture an image, after rapid detection and pretreatment, using defect detection technologies to identify defect areas in the image. Using machine learning methods to train a classification model, and then distinguish the defect areas. This method can improve productivity and ensure quality greatly. At the same time, it can reduce the labor intensity.This paper mainly uses two machine learning methods: back propagation(BP) neural network and support vector machine(SVM), to finish the defect recognition on steel strip surface. Before feature extraction step, the methods sets two kinds of threshold according to the mean of gray scale for different defects. When extracting features, combining with target image and binary image to extract three types features that are geometrical characteristic, gray characteristic and shap characteristic. Using three layers network to train a classifer in BP neural network and the number of neurons in hidden layer are confirmed by several experiment results. During training SVM, it utilizing Gauss radial basis function as the Kernel function, determining model parameters C andγby cross-validation, employing"one against one"method for mutilclass classification.In this paper, we use two classification models to study the five definition defects. Most of the experiments are finished by the OpenCV library. Experiment results show that SVM model has a higher classification accuracy rate while BP neural network is better than SVM model in average recognition time. In a word, these two methods are suitable for system's real-time requirements.