Abhishek Goyal 98010 BTP Guide

Feature Based Classification using Radial Based Functions Neural Networks is being used to determine defects in steel. Other Strategies to solve the problem are Multi layered Networks which require designing of various layers in hidden layer of the network but such techniques are infeasible here due to lack of adequate training data and high probability of inaccuracy due to improper design of network. Moreover such techniques take more time to give results so they are difficult for such real time requirement. Radial Based Functions technique used here is giving 71 % correctness in result and fast enough to be extended for parallel systems to produce a real time solution for this problem. The Issues in Surface Defect Inspection At present there are commercially available products which can detect the presence or absence of surface defects at reasonable costs. The following discuss these issues in further details. A. High Data Throughput: A typical CRGO sheet is 1 to 3 m. wide with a thickness of 1 to 5 mm. These steel sheets of endless length move continuously on the conveyer belt at a very high speed of about 20 m. per second. Thus, it is a tough job for the inspection system to acquire and effectively process a high amount of data in a short amount of time. B. Inter-class Similarity and Intra-class Diversity: A single class of defect may vary widely in appearance and structure. More so, the members of one class of defect may closely resemble to that of the other class. C. Large Number of Classes: A typical defect identification scheme should deal with a large number of defect classes. It is not unusual to deal with a few dozen of defects. D. Non availability of adequate imperfection imagery: Another very significant problem encountered during the design and development of the inspection system is the non-availability of adequate imperfect imagery for feature extraction and machine learning. The hazardous environment of the steel industry hampers the collection of imagery data. This problem is more acute due to the fact that the process of machine learning requires a very large amount of training data for proper identification of the high number of defect classes. E. Dynamic Defect Populations: Little changes and alterations in the manufacturing process may create an entirely new set of defect classes or may add up new types of defects to the existing ones.