Publication: Deep learning-based cmp ring defect detection system
Date
2021-07-01
Authors
Mani Segaran, Sobini
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
With the increasing production of chemical mechanical polishing (CMP) rings, the detection efficiency and accuracy has become more important. Manual vision inspection is unstable and inefficient due to the randomness and uniqueness of defects. Therefore, conventional machine vision methods are commonly used to assist humans in generating good detection results. However, these methods are outperformed by state-of-the-art deep learning techniques as it results in a high inspection accuracy and robustness. This thesis proposed a deep learning-based defect detection model based on a pre-trained model of Faster R-CNN Inception v2 and YOLOv5s where the efficiency between the two algorithms are compared. Three types of defects are detected by the proposed models trained with custom dataset which includes top scratches, side scratches and burr. The Faster R-CNN model achieved an average precision of 60.53%, 72.05% and 51.50% for the detection of burr, side scratches and top scratches, respectively. YOLOv5s model achieved an average precision of 45%, 50.6% and 22.4% for the detection of burr, side scratches and top scratches, respectively. Faster R-CNN model achieved a mean average precision of 71.02% while YOLOv5s model achieved 54.4% in detecting the defects on CMP ring. The minimum score threshold is set to 0.5 when evaluating the performance of both the models. In conclusion, the proposed Faster R-CNN model is suitable to be used as the CMP ring defect detector model in this project.