Unsupervised defect classification using casting product images based contrastive similarity

Abstract

Quality control is a very important factor in the manufacturing process. In particular, quality control in the casting product industry uses recently captured product images to classify them with supervised learning-based automated defect model. However, high-quality image data with perfect labeling is essential for good performance. Moreover, acquiring high-quality labeling data has temporal and financial difficulties. In this study, we proposed unsupervised learning defect classification model using unlabeled images. Starting with preprocessing to minimize computation, we have easily increased image qualities with simple data augmentation to extract a variety of defect locations and data points. After extracting features with CNN to learn the representation of augmented views, the dimensions are reduced with a projection layer. Then, we calculated the optimal weights and hyper-parameters that maximize cosine similarity-based contrastive learning. Finally, the features were clustered with K-Means to confirm the performance difference from the previous approach. In addition, to compare with the supervised learning study of casting product images, we confirmed the performance through fine-tuning which classifies the same images with real label. The proposed method achieved high performance compared to the existing supervised and unsupervised learning in various indicators. In other words, considering the practical problem of data labeling, we would like to propose a defect classification model for quality control through unsupervised learning based on contrast similarity using features and representations of images without labels.