Injection-molded products frequently exhibit localized surface defects such as weld lines, flow marks, scratches, bubbles, and burn marks due to variations in material flow, mold temperature, and cooling conditions. Conventional visual inspection is highly dependent on operator experience, while rule-based machine vision methods are limited under variations in lighting and surface texture. This study proposes a deep learning–based defect detection model using YOLOv8 combined with a novel Defect-Aware Augmentation technique designed to enhance robustness for small, local defect regions. The proposed augmentation pipeline includes geometric transformations, optical perturbations, local defect patch synthesis, and diffusion-based synthetic defect generation. Experiments were conducted on a custom dataset of 5,000 images (3,000 normal and 2,000 defective). Results show that the proposed model achieves significant improvements over baseline models, obtaining 95% precision, 90% recall, and 0.96 mAP@0.5, outperforming the default YOLOv8 model by 7%p in mAP. Ablation studies verify that defect-aware augmentation is the dominant factor contributing to the performance gain. The proposed system demonstrates high applicability for automated quality inspection in injectionmolding production lines.

Abstract
1. 서 론
2. 관련 연구
    2.1 사출품 표면 불량 유형
    2.2 딥러닝 기반 불량 검출 연구
3. 데이터셋 구축 및 전처리
    3.1 데이터 구성
    3.2 데이터 분할
    3.3 전처리 과정
4. 데이터 증강 기법
5. 실험 결과
    5.1 비교 모델 성능 평가
    5.2 혼동 행렬 분석
    5.3 정밀도-재현율 곡선
    5.4 정확도 평균값 및 손실 수렴 분석
    5.5 구성요소 분석 실험 결과
6. 결 론
References

• 박창규(Dept. of Mechanical & Automotive Engineering, Gwangju University) | Chang-Kyu Park Corresponding author