A Study on Vision-based Surface Inspection Method for Highly Noised Steel Production

DoctorAn automated vision-based surface inspection for highly noised steel production is proposed in this thesis.The ``highly noised'' means the surface of the steel product is covered with a lot of scales, which make the inspection difficult task. This thesis focuses the inspection method for steel wire rod. The optimal illumination structure is proposed to detect defects with depth by using double dark fields.The inspection methods for the most critical defects, roller-mark (periodic defect) and scratch (non-periodic defect) of wire rod are presented. The two kinds of defects are caused by defective rollers and occurs in a large quantity. It is essential for operators to receive early warnings of such defects and take appropriate action to avoid a large quantity of defective products.This thesis presents an efficient and real-time method for detecting defects in continuous steel wire rods. The detection algorithm is based on the undecimated discrete wavelet transform (UDWT). To overcome the noisy background of surface image, the algorithm exploits the translation invariant property of the UDWT, which improve the signal-to-noise ratio (SNR) of the image. The horizontal and vertical detail coefficients are binarized by double thresholding technique followed by the appropriate size filtering. The final defect candidates are obtained by morphological reconstructing operation.To identify periodic defects, the frequency spectrum of defect positions is analyzed. The spatial periodicity of detected defects is estimated from the inversely related periodic information of frequency spectrum.To get 1D frequency spectrum of defects, the algorithm converts 2D positional information of defects into 1D defect histogram. A robust algorithm to detect repeating impulse-like points in the spectrum uses a median and local maximum filters. The algorithm was proved to operate even when the false positive defects are over 15 times more than true positive defects, or 70 % of true positive defects are missed.To classify non-periodic defects, the classification using support vector machine (SVM) is proposed. The 14 features are extracted from binary and gray information of detected objects. The best model SVM classifier with radial basis function (RBF) kernel is obtained by 5-fold cross validation. The grid search for finding the two parameters for RBF kernel is performed. The support vectors are generated from train data and the final confusion matrix is obtained from test data, which show 85 % and 94 % of true positive and negative rate, respectively.The inspection system using the proposed method was applied to a real production line and showed the effectiveness of the method