생성적 적대 신경망(GAN)을 이용한 광학렌즈 결함판정 모델 개발

Abstract

Currently, the defect determination of the optical lens is performed by a method through an inspection personnel. This not only consumes a lot of time and money during the inspection process, but also adversely affects the production quality of products due to continuous detection errors. To solve this problem, we tried to develop a defect detection model based on convolutional neural networks, but the performance of the deep learning model could not be secured due to the data imbalance problem.Therefore, in this paper, a defect determination model of optical lenses was developed using generative adversarial neural networks and convolutional neural networks. Three types of upper and lower images of optical lenses were used for development, and polar coordinate conversion, Canny edge, Sobel filter, and Laplacian filter were combined and used as image preprocessing methods. The convolutional neural network model conducted transfer learning using vggNet, ResNet, and DenseNet structures, and performance evaluation was conducted for each model, image conversion, and abnormal image ratio. The semi-supervised Anomaly detection models used AnoGan, Ganormaly, and skip-Ganormaly, and the optimal model was selected through performance evaluation of each model. After that, the performance was evaluated by implementing a weight-based ensemble model of the two models.