Image recognition of interference fringes in polishing by convolutional neural network with data augmentation by deep convolutional generative adversarial network

Abstract

With the growing need for high specification requirements for the latest manufacturing processes and optical designs of glass lenses, the technical requirements for lens polishing have increased. The manufacturing parameters must be adjusted in a timely manner to meet the required specifications. We use a Fizeau interferometer to classify and analyze interference fringes measured in the actual polishing process of glass lenses. Given the low incidence of interference fringes in practice, the training dataset contained a disproportionate ratio of data for each data type. To reduce the manufacturing cost and data collection time, this study focused on three common types of interference fringes in the manufacturing processes and integrated a deep convolutional generative adversarial network with convolutional neural networks (CNNs) for fringe type classification. The deep convolutional generative adversarial network was used to establish a data augmentation generator, and Jensen–Shannon divergence was employed to identify the epoch number that yielded distributions of interference fringe numbers closest to the real distributions; this approach could achieve the diversity of interference fringes in the generated images. Finally, the generated data were used to train the CNN models, and the accuracy of image recognition reached above 86%.