DETECTION AND CLASSIFICATION OF SUBMICRON SURFACE DEFECTS BASED ON THEIR INTERFERENCE IMAGES USING DEEP LEARNING

Abstract

An automated method for detecting and classifying submicron surface defects on mirrors used in high-precision optical systems has been developed, utilizing interferometric image analysis and deep learning. This approach replaces manual inspection with a neural network, delivering faster and more objective defect diagnostics, eliminating human bias, and accelerating quality control in serial production of optical components. A synthetic dataset of interferometric images was generated to train the model, simulating scratch-type defects on mirror surfaces through specialized software based on the Linnik interferometer model. The dataset encompasses three surface classes: flat surfaces, single scratches, and multiple scratches. The neural network is built upon MobileNetV2, pre-trained on ImageNet, with fine-tuning of its final blocks to adapt to the task’s specifics. The architecture incorporates GlobalAveragePooling2D for feature compression, Dense layers with ReLU activation and BatchNormalization, Dropout to mitigate overfitting, and a Softmax output layer for classifying the three categories. Data augmentation and soft voting techniques were employed to enhance the model’s generalization ability. Classification accuracy, assessed using the accuracy metric, achieves 96% on the synthetic validation set and 82.7% on real images acquired from a Linnik interferometer. The highest accuracy is observed for flat surfaces, while the lowest occurs for multiple scratches, highlighting challenges posed by real-world conditions such as noise and artifacts. The method proves its practical value for automated diagnostics, with future enhancements tied to improving synthetic data realism-potentially by incorporating modeled noise-and extending the model’s adaptability to additional defect types like indentations or protrusions, thereby broadening its applicability in optical system manufacturing