Using Computer Vision to Recognize Defects on the Surface of Hot-rolled Steel

The problem of recognizing defects on the surface of hot-rolled steel is quite old, but technology has only recently reached a sufficient level to enable automation of this process. One of the most suitable methods is applying convolutional neural networks (CNN). We selected the Northeastern University surface defect database, which is a dataset of the most identified cases of hot rolled defects, as the qualitative dataset for network training. This article presents CNN models to recognize 6 defects with an accuracy of 93.59% and 6 defects and images of a clean surface with an accuracy of 92.31%. The recognition time was 0.001384±5% seconds for all samples. As well, program give recommendations based on the most common defects of a particular type.