Design of online classifier for surface defect detection and classification of cold rolled steel coil

Abstract

The target to be achieved through this project was primarily aimed at detecting the surface defects belonging to different classes in cold rolled steel coils. This was achieved through grabbing the images from the camera, here line scan camera is used which grabs 20 frames per second. Carrying out defect detection on these images and later classifying them. We present a method to automatically detect and localize defects occurring on the surface. Defect regions are segmented from background images using their distinguishing texture characteristics. This method locates candidate defect regions directly in the DCT (Discrete cosine transform) domain using the intensity variation information encoded in the DCT coefficients. More precisely, defect detection employs DCT analysis of each individual non-overlapping region of the image to determine potentially defective blocks, which are further grown and merged to form a defect region on the image. In this thesis a computer vision based, a framework for steel surface defects detection and classification of cold rolled steel strips is implemented. We have designed online classifier for automatic defect detection and classification of defects. In this we measured statistical textural features using gray level co-occurrence matrix presented by Haralick and geometrical features are also calculated. The final decision SVM (Support Vector Machine) handles the problem of classification of the defect types. We also proposed SVM voting strategy for the final decision that handles the problem of multiple outputs of a given input image with a specific defect type. In addition, this approach improves the classification performance. Experimental results demonstrate the effectiveness of the proposed method on steel surface defects detection and classification. In addition, the defect information is encoded in the image. An image viewer application is designed for decoding the defect information.