A Robust and Fast Deep Learning-Based Method for Defect Classification in Steel Surfaces

The final product quality control is critical for any manufacturing process. In the case of steel products, there are different inspection methods that are able to classify the defects, but they usually require human intervention. In this context, a deep learning-based automatic defect classifier method for steel surfaces is proposed. The method combines some traditional Machine Learning techniques with a Convolutional Neural Network (CNN). Different experiments were carried out in order to obtain the best classifier parameter setup. To verily the robustness of the classifier some additional experiments were done, obtaining high classification rate against some sources of noise such as illumination changes or occlusions. The proposed method achieves a classification rate of 99.95% taking 0.019 seconds to classify a single image. The method is compared with seventeen related methods and outperforms them on a publicly available dataset, with six types of defects and 300 samples for each class. The source code of the proposed method is publicly available.