Defect detection in manufacturing processes is a critical requirement for ensuring product reliability and maintaining production stability. As smart manufacturing environments continue to advance, the need for precise and robust vision-based inspection methods has become increasingly significant. This study proposes a hybrid defect analysis framework that integrates YOLOv5-based defect candidate detection with an Attention U-Net–based segmentation module. Experiments conducted on chromate-coated industrial images demonstrate that the proposed framework achieves an accuracy of 0.97, precision of 0.91, recall of 0.89, F1-score of 0.93, and IoU of 0.88, exhibiting stable performance even for small defects and irregular boundaries. The combination of region- of-interest extraction and attention-enhanced pixel-level segmentation improves both computational efficiency and boundary reconstruction quality. The findings extend the applicability of attention-based segmentation to industrial defect inspection and provide practical insights for deploying deep learning–based quality monitoring systems in automated manufacturing environments.

1. 서 론
2. 관련 연구
    2.1 딥러닝 기반 제조 결함 탐지연구
    2.2 Transformer 기반 모델의 비교 연구
    2.3 본 연구의 제안과 차별성
3. 본 론
    3.1 방법론
    3.2 제안된 통합 프레임워크
4. 데이터셋 및 실험환경
5. 실 험
    5.1 YOLO와 Attention U-Net 결합
    5.2 Attention U-Net 학습
    5.3 결함 검출 결과
6. 결 론
References

• Dong-Gun Jang(Department of Industrial Engineering, Dankook University) | 장동건 (단국대학교 일반대학원 산업공학과)
• Young Chul Chang(Department of Industrial Engineering, Dankook University) | 장영철 (단국대학교 일반대학원 산업공학과) Corresponding author
• Jae Hyung Cho(Department of Industrial Engineering, Dankook University) | 조재형 (단국대학교 일반대학원 산업공학과)