Publication: Deep learning based machine vision inspection system for defect detection
Loading...
Date
2021-07-01
Authors
Mohamad Nazari, Mohamad Haikal
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Machine vision based product inspection methods have been extensively investigated to improve quality and minimize labor costs. Recent progress in deep learning offers advanced
analytics tools with high accuracy and robustness in inspection. The construction of the deep learning model, however, is typically computationally expensive, which may not match the quick inspection requirements. Therefore, this paper presents a new machine vision inspection method based on deep learning by using two different YOLO model which YOLOv3-spp and YOLOv5x to identify and classify CMP surface defects. To achieve accurate detection and classification, YOLOv3-spp and YOLOv5x are trained and tested on a dataset containing three different types of defects which are top scratches, side scratches and side burr. The network can also obtain the coordinates of the detected bounding boxes, which can be used to determine the size and location of the defects. After that, both YOLOv3-spp and YOLOv5x model are compared based on mean average precision (mAP), precision and recall in order to determine which YOLO model is the best to implement for CMP defect detection system. Based on the results generated in this project, the proposed method for detecting CMP surface defects is feasible. Compared to YOLOv3-spp and YOLOv5x, both models made significant progress even on a small dataset. Using GoogleCollaboratory, the best mean average precision (mAP) of the YOLOv5x model is 6.57 %, and the highest recall of the YOLOv5x model is up to 13.4 %. However, when compared to the YOLOv5x model, the YOLOv3-spp model achieved slightly higher precision by 20.9 %. Furthermore, the YOLOv5x model outperforms the YOLOv3-spp model in terms of robustness and accuracy in distinguishing three different defects.