Multi-classification and object detection in intelligent manufacturing

Abstract

Defect detection in industries is typically conducted manually. While there are state-of-the-art machine vision techniques for automated inspection systems, there is still a gap between research advancement and practical applications, especially for manufactures with high volume and low margin. The thesis aims to develop a computer vision system for automated galvanized steel tube defect detection. Based on images collected from a Japanese steel tube producer, multiple methods were explored and tested. Firstly, inception v4 was used as an image classification model. Its performance was first tested on an online dataset, then on our own cleaned dataset. In the next step, since classification only labels a whole image, object detection algorithms were then used for indicating locations as well as the defect class. Several object detection algorithms were adopted and compared: Faster R-CNN, YOLO v4, and YOLO v5. They achieved mAP@0.5 of 94.31%, 95.22%, 75.5% respectively, and recall rates of 67%, 89%, 73.5% respectively, which demonstrated promising results for applications on the production line. However, the results were primarily limited by the quantity and quality of images. Future work could focus on advanced data augmentation, further cleaning on collected data, and improvement in raw image quality. Furthermore, the algorithms need to be validated with real-time inspection speed and on more classes of defects.