YOLO Based Deep Learning Network for Metal Surface Inspection System

Abstract

Surface defects on metal are complex and it becomes one of the most common problems in the industrial environment. Different production lines can cause surface defects with various properties. Defect detection of the metal surface is necessary for the quality control of industrial products. Traditional programming approach, machine learning, or deep learning are applied for vision-based metal surface defect inspection system. Among them, deep learning becomes popular to improve the quality control in industrial applications. In this paper, You Only Look Once (YOLO) object detection network is used for the development of metal surface defect inspection system. The purpose of this paper is to present the metal surface inspection system that can differentiate the types of defects and locate the defect. YOLO deep learning network for metal surface inspection system is implemented on the Jetson Nano embedded board. The Northeastern University (NEU) surface defect dataset is used. YOLO network can predict six types of defect on the metal surface and performance of 71% mAP is achieved. The total processing time for each defect image is 0.034 s.