Earticle

초록

영어

In this study, we propose and analyze data augmentation techniques tailored for injection-molded product surface defect detection using deep learning–based vision inspection systems. To alleviate the scarcity and imbalance of defect samples in real production lines, a data augmentation strategy consisting of four categories—geometric transformation, photometric transformation, defect-aware local augmentation, and synthetic defect generation—is designed. The effectiveness of the proposed strategy is evaluated using a ResNet-50–based classification model and a YOLOv8-S–based defect detection model, trained on real images acquired from an industrial injection molding process. Experimental results show that, compared to the baseline model trained without augmentation, the proposed strategy improves the F1-score of defect classes by approximately 10–15 percentage points and enhances mAP@0.5 by about 8–12 percentage points. In particular, the recall of rare defect types such as micro-cracks and weld lines is significantly improved, indicating that data augmentation plays a crucial role in increasing the practicality of automatic inspection in injection molding processes. The findings of this work provide useful guidelines for implementing deep learning–based visual inspection systems in real industrial environments with limited defect data.

목차

Abstract
1. 서론
2. 관련 연구 및 이론적 배경
2.1 사출품 표면 불량 유형
2.2 딥러닝 기반 불량검출
2.3 제조 비전 검사 분야 데이터 증강
3. 데이터 증강 기법
3.1 데이터 셋 구성
3.2 전처리
3.3 제안 데이터 증강 기법
4. 실험 구성 및 결과
4.1 실험 환경 및 모 델 설정
4.2 비교 실험 시나리 오
4.3 평가 지표
4.4 분류 성능 비교
4.5 검출 성능 비교
4.6 조명 조건 변화에 대한 강인성 평가
5. 결론
후기
References

키워드

저자

• 박창규 [ Chang-Kyu Park | Member, Dept. of Mechanical & Automotive Engineering, Gwangju University ] Corresponding Author

참고문헌

발행기관

간행물

표지보기 관심저널 등록

• 간행물명

  한국기계항공기술학회지(구 한국기계기술학회지) [Journal of the Korean Society of Mechanical and Aviation Technology ]

• 간기

  격월간

• pISSN

  1229-604X

• eISSN

  2508-3805

• 수록기간

  1999~2026

• 등재여부

  KCI 등재

• 십진분류

  KDC 550 DDC 620