Deep learning-based methods for detecting surface defects in steel plates

Abstract

Steel is one of the most essential construction materials in today's world due to its many unique advantages. However, the quality of steel can be adversely affected by common surface defects like pitting corrosion, pores, scabs, blisters, crazing, etc. Manual and automatic inspections are the two ways one can go for the identification of these defects. Automating the inspection procedure would speed up the production of steel sheets. The objective of this study is to examine the Severstal Steel Defect dataset and to explore efficient methods to classify and label the faulty areas pixel by pixel. For multilabel classification, a hypertuned single and multioutput convolutional neural network is trained, and its performance is compared with a pretrained Xception architecture. The findings reveal that the new method's deep learning network model has good detection performance, with a mean dice coefficient of 0.6888. © 2022 Elsevier Inc. All rights reserved.