SURFACE DEFECTS CLASSIFICATION IN STEEL PRODUCTS: A COMPARISON BETWEEN DIFFERENT ARTIFICIAL INTELLIGENCE-BASED APPROACHES

摘要

In many steelmakers, the quality control of the flat steel products is, nowadays, performed by an automatic surface inspection system. This system analyzes the images of the steel surface and classifies the detected defects depending on their types. The reliability of the ASIS classification is limited by the huge amount of the taken images and by the time constraints. In the present paper the correct classification of a particular kind of surface defect, named Large Population of Inclusion, is sought. With this aim several algorithms have been developed and compared: an adaptive neuro-fuzzy inference system, a multi-layer perceptron, a decision tree, a support vector machine and a learning vector quantization. The tests that have been developed show that the more suitable algorithm for this task is an adaptive neuro-fuzzy inference system, which has been developed also by exploiting the human experience in the defect recognition.