Evaluation of the Transfer Learning Models in Wafer Defects Classification

In a semiconductor industry, wafer defect detection has becoming ubiquitous. Various machine learning algorithms had been adopted to be the “brain” behind the machine for reliable, fast defect detection. Transfer Learning is one of the common methods. Various algorithms under Transfer Learning had b...

全面介紹

Saved in:
書目詳細資料
Main Authors: Jessnor Arif, Mat Jizat, Anwar, P. P. Abdul Majeed, Ahmad Fakhri, Ab. Nasir, Zahari, Taha, Yuen, Edmund, Lim, Shi Xuen
格式: Conference or Workshop Item
語言:English
出版: Springer Nature 2022
主題:
在線閱讀:http://umpir.ump.edu.my/id/eprint/36763/1/Evaluation%20of%20the%20Transfer%20Learning%20Models%20in%20Wafer%20Defects%20Classification%20%281%29.pdf
http://umpir.ump.edu.my/id/eprint/36763/
https://doi.org/10.1007/978-981-33-4597-3_78
標簽: 添加標簽
沒有標簽, 成為第一個標記此記錄!
機構: Universiti Malaysia Pahang Al-Sultan Abdullah
語言: English
實物特徵
總結:In a semiconductor industry, wafer defect detection has becoming ubiquitous. Various machine learning algorithms had been adopted to be the “brain” behind the machine for reliable, fast defect detection. Transfer Learning is one of the common methods. Various algorithms under Transfer Learning had been developed for different applications. In this paper, an evaluation for these transfer learning to be applied in wafer defect detection. The objective is to establish the best transfer learning algorithms with a known baseline parameter for Wafer Defect Detection. Five algorithms were evaluated namely VGG16, VGG19, InceptionV3, DeepLoc and Squeezenet. All the algorithms were pretrained from ImageNet data-base before training with the wafer defect images. Three defects categories and one non-defect were chosen for this evaluation. The key metrics for the evaluation are classification accuracy, classification precision and classification recall. 855 images were used to train and test the algorithms. Each image went through the embedding process by the evaluated algorithms. This enhanced image data numbers then went through Logistic Regression as a classifier. A 20-fold cross-validation was used to validate the score metrics. Almost all the algorithms score 85% and above in terms of accuracy, precision and recall