주조 제품 이미지의 대조 유사도 기반 비지도 결함 분류

Abstract

Purpose: While supervised learning models require a large number of high-quality labeled images, acquiring such data is often impractical due to time and cost constraints. This study aims to classify casting defects using unlabeled images by leveraging contrastive learning techniques. Methods: The proposed model applies image pre-processing and augmentation to increase data diversity with minimal computation. Then, contrastive learning is used to maximize the similarity between augmented images, allowing the model to efficiently learn meaningful features. Results: The proposed method demonstrated improved performance across various evaluation metrics in classifying casting product images. Compared to previous supervised learning-based approaches, the unsupervised model achieved competitive or improved results without requiring labeled data. Conclusion: The results suggest that the proposed contrastive learning-based classification model can serve as an alternative to supervised methods in scenarios where labeled data are scarce or even unavailable. This approach offers a practical and scalable solution for defect detection in casting product quality control. * 본 논문은 산업통상자원부(MOTIE)와 한국에너지기술평가원(KETEP)로부터 연구비를 지원받아 작성하였다. †교신저자 sjbae@hanyang.ac.kr 2025년 4월 11일 접수; 2025년 7월 29일 수정본 접수; 2025년 7월 30일 게재 확정.