Convolutional neural network for sapphire ingots defect detection and classification
Euphrem Mugisha Rwagasore,Xiong Zhang,Kaifang Gao,Zuoxuan Gao,Zhitao Zan,Xiaohu Lui,Mengtong Wang,Yuhang Mi,Hongjian Chen,Wenbo Yan
DOI: https://doi.org/10.1016/j.optmat.2021.111292
IF: 3.754
2021-09-01
Optical Materials
Abstract:<p>Optical homogeneity detection can be brought into play for quality control of sapphire substrate, sapphire mobile phone screen, and sapphire window material. By detecting the optical homogeneity of the sapphire crystal, the crystal lattice distortion inside the crystal can be evaluated non-destructively and intuitively. However, due to the small birefringence coefficient of sapphire crystals, it is difficult to detect the optical homogeneity of sapphire ingots through the conventional setup. In this work, we detected the optical homogeneity of the sapphire ingot using the Focused Polarization Interference (FPI) method also known as Conoscopic Birefringence Interferometry. The main principle of this technique is that the stress induced by the lattice distortion may cause the variation of the crystal birefringence which leads to the formation of the polarization interference patterns. To quantitatively evaluate the pattern deformation, we used pattern recognition technique. However, the main issue in pattern recognition and notably images classification problems is image feature extraction and image encoding. To solve the above problem, here we showed and compared different approaches: Bag of features, ResNet-50, AlexNet, VGG16, and GoogleNet pre-trained convolutional neural network (CNN) models. The reliability, practicality, and verification of the proposed models, in this work, is according to the detection measures use accuracy, confusion matrix, and Area Under the Curve (AUC). The Bag of features model achieved the classification accuracy of 94.3%, while the ResNet-50 pre-trained convolutional neural network achieved 97.6%. The findings show that ResNet-50 is capable of correctly classifying the samples from the datasets with high true fraction value and low false positive followed by AlexNet as compare to other models like VGG16 and GoogleNet. Finally, the proposed models have been in contrast with the well-known chemical etching method. The results showed that the time required to etch one sample using chemical etching is about 10 h, while by using the proposed models, we can detect and classify the whole dataset within 80 min and 27 s. The proposed pre-trained convolutional neural network models based on sapphire crystal defects classification would potentially serve as a new tool for predicting the quality inspection-based on crystal growth technology.</p>
materials science, multidisciplinary,optics