A Deep Learning-Based Surface Defect Inspection System for Smartphone Glass

Abstract

In recent years, convolutional neural network has become a solution to many image processing problems due to high performance. It is particularly useful for applications in automated optical inspection systems related to industrial applications. This paper proposes a system that combines the defect information, which is meta data, with the defect image by modeling. Our model for classification consists of a separate model for embedding location information in order to utilize the defective locations classified as defective candidates and ensemble with the model for classification to enhance the overall system performance. The proposed system incorporates class activation map for preprocessing and augmentation for image acquisition and classification through optical system, and feedback of classification performance by constructing a system for defect detection. Experiment with real-world dataset shows that the proposed system achieved 97.4% accuracy and through various other experiments, we verified that our system is applicable.