Application of bandelet transform to surface defect recognition of hot rolled steel plates

Surface defects are important factors to surface quality of steel plates. The detection and recognition of surface defects can provide effective information for production optimization. There are several types of surface defects on hot rolled steel plates which are covered by lots of scales. The purpose of this paper is to recognize eight kinds of typical surface defects from scales. Bandelet transform is applied to extraction of geometrical features. Firstly, each sample image is decomposed into multiple directional subbands at several scales by bandelet transform. Then, some statistical values of bandelet coefficients are calculated and combined into a feature vector from all subbands. In this process, several important parameters of bandelet transform are discussed and determined through experience and experiments. Finally, the feature matrices of training set and testing set are inputted into Support Vector Machine for classification. Experiments with sample images from a real production line of hot rolled steel plates show that bandelet transform is superior to curvelet transform and contourlet transform. Most of surface defects can be effectively recognized and the highest recognition rate of testing set is up to 96.07%.