Steel Defect Classification Using Machine Learning

Abstract

Ensuring the quality of industrial production for the steel industry is very crucial. For complete defect detection, it is important to know the exact location and class of defect, due to which it becomes difficult to apply this method and attain accuracy in both location and classification. Methods used for the detection of steel defects include the YOLO detection network, acoustic emission method, end-to-end steel surface defect classification, and detection of defects by magneto-optical imaging and neural networks. But as data is too large, deployment and training of these systems become expensive and time-consuming, therefore the algorithms used for the detection of defects should have good generalization. With the growth in computer vision and deep learning automation it is possible to classify images with maximum accuracy. We used machine learning algorithms KNN and transfer learning using VGG-16 for this task to help in quality improvement, quick detection, and classification. KNN used for the classification of defects provided fairly improved results with a significant gain in the accuracy. Detection of defects done by transfer learning via VGG-16 provided promising results. The model trained using VGG-16 achieved high accuracy of 97.54%. These techniques provide an optimal solution for both the classification and detection of defects.