Algorithm for Automatic Pattern Classification Designed for Real Metallographic Images

Abstract

The problem of automatic pattern classification in real metallographic images from the steel plant ArcelorMittal Ostrava is addressed. The goal is to monitor the process quality in the steel plant. In the images of metal, there are dark dots that are produced by imperfections along the central axis of each plate. It is necessary to determine automatically the number and sizes of these dots. The number and sizes of the dots is a measure of how imperfect each plate is. The process is presented that segments the area of plates that contains segregation, identifies those rows of pixels along which the dots lie, and counts the pixels that are marked as dots by evaluating all the vertical columns of pixels that intersect the rows that contain the dots. The threshold value is set to be 95% of the mean value of grey scale for each column of pixels and makes the dots white. White dots that are most likely noise are removed to identify dots that are smaller than 4 connected pixels across. The explanations related to the obtained results are firmly related to the information provided by human experts.