Extracting Surface Defect Features From Steel Strip Images

Strip steel is one of the main products in the steel industry, which is widely used in the national economy and daily life. Due to the constraints of various factors, the strip steel surface may have various types of defects. Therefore, timely and accurate detection of the surface defects of the strip steel is of great significance to improve the quality of the strip steel products.In the production line Steel strip runs very fast, defect detection system should be real time to meet time constraints. In the actual production environment, the vast majority of the strip surface images are the images without defects. We use the fast prediction technology to filter out images without defects before the feature extraction. Prediction has two stages, image binarization and defect judgment. After binary processing, the defective parts have been presented, and then whether there are the defects or not are determined by analyzing image gray distribution statistics. In the image analysis, texture is an important kind of feature. Considering most of the strip steel defect images are rich in texture information, the gray level co-occurrence matrix(GLCM) is used for feature extraction. The feature vector is constructed been on generated direction, generated step and gray level. Unlike traditional methods using average value of parameters in the different direction as the final vector elements, the average value of multiple parameters in each direction are used here, this method can significantly improve the classification accuracy of the defects.Histogram of oriented gradients(HOG) feature is another feature extraction that can be extracted from an image. It has a very good invariance to image geometry and optical distortion, which is originally used to detect pedestrians in static images. Because of the strong ability to describe the image features, it is extended to various fields. Aiming at the problem of strip steel defect classification, the parameters affecting the characteristics of HOG are optimized, and the template of gradient calculation is improved. The experi-mental results show that the high dimensional template is more suitable to the classification of strip steel defects.