Defect Detection And Analysis System For Steel Bloom And Billet

• Main Authors: Wen-cheng Hsu, 許紋誠
• Other Authors: Yung-Nien Sun
• Format: Others
• Language: zh-TW
• Published: 2008
• Online Access: http://ndltd.ncl.edu.tw/handle/41441379640218080512

碩士 === 國立成功大學 === 資訊工程學系碩博士班 === 96 === In this thesis we present an automatic defect detection and analysis system for steel billet. In practice, the system analyzes the sequentially acquired steel billet images to locate defects on the steel surface and recognizes the types of detected defects. The proposed system is consisted of three modules: (1) Image processing and defect detection, (2) Feature extraction and selection, (3) Incremental learning classifier. For the purpose of on-line detection, we use low complexity algorithms in the first module. We also adopt some strategies to speed up the computation. Before classification, it is important to find out the features that have better ability to distinguish defect classes. Therefore, we design a feature selection module which combines Tabu Search with k-nearest neighbor classifier to obtain the best set of features for classification. In practice, defect samples will ceaselessly arrive and thus increase the cost of training time to update classifier and more memory space to store these samples. In this study, we utilize the Learn++ classifier to overcome the problem mentioned above. In the experiment, some expert detection results were used to judge the correctness of the proposed defect detection. The results show that the proposed system not only detects defect correctly and rapidly, but also correctly recognizes classes of steel defect. In the classification module, training the classifier with incremental learning algorithm saves a lot of time than that with the conventional BPN skill.