Development of image recognition system for steel defects detection

Chen, Wai Yang (2022) Development of image recognition system for steel defects detection. Final Year Project, UTAR.

Preview

PDF Download (1931Kb) | Preview

Abstract

Hot rolled steels are among the highest-demand steels in the construction and manufacturing industry. The manufactured steel inevitably comes with some defects from the production line. Hence, it is essential to conduct a quality control process to ensure the produced hot rolled steels meet the customer’s requirements. Currently, most industries rely on human visual inspection systems for quality control. However, this inspection is not efficient and time-consuming. Furthermore, the quality of inspection may differ because different inspectors may have their own judgement on the quality. An image recognition system can improve the quality of hot roll steels and work efficiency. In this project, an image recognition system for steel defects detection has been developed to detect three types of hot rolled steel defects: rusting, edge, and loose wrap. For the rusting detection algorithm, a deep learning model, Single Shot Detector (SSD), was trained to detect and crop the hot rolled steel from the input image for colour detection. The colour detection was implemented to determine the rusting area on the hot rolled steel based on the orange-brown colour that appeared on the hot rolled steel. The system can decide whether to release or hold the hot rolled steel based on the percentage of the rusting area on the hot rolled steel. Meanwhile, the system carries out the model inference by utilizing the trained SSD model to find and crop the Region of Interest (ROI) from the input image regarding edge defects and loose wrap detection. Then, the system conducts Canny Edge Detection to find out the irregular edge lines caused by the defects. The system can determine whether to release or hold the hot rolled steel based on the generated edge lines that indicate its severity. Based on the experimental result, the rusting detection has more than 90% accuracy with less than 50 ms processing time. Besides, the edge defects detection has an average of 69% accuracy with an average 63 ms processing time. Last but not least, the loose wrap detection achieved an average of 84.9 % accuracy with 51.3ms inference time. The detection errors are due to the variety of input images in terms of angle and brightness.

Actions (login required)