Robust Deep Learning Framework untuk Deteksi Cacat Permukaan pada Material Logam dalam Kondisi Pencitraan yang Terganggu Noise dan Blur

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Surface defect detection plays a crucial role in industrial quality control, especially in metal manufacturing, where defects can impact structural integrity and product reliability. Although deep learning-based inspection systems have shown promising results, most approaches assume ideal imaging conditions and experience significant performance degradation when faced with disturbances such as noise and blur. This study proposes a robust deep learning framework for surface defect detection under degraded imaging conditions. The framework integrates controlled image degradation simulation and a segmentation-based model using U-Net to enable precise pixel-level localization of defects. Robustness is evaluated through the addition of Gaussian noise, salt-and-pepper noise, and motion blur. Experimental results on the Severstal Steel Defect Dataset demonstrate that the proposed method outperforms the baseline segmentation model YOLOv8n-seg. Under clean conditions, the model achieves a precision of 0.6328 and an F1-score of 0.3199. In degraded conditions, the model remains stable despite performance drops, whereas other methods show significant declines. Additionally, the segmentation output allows for quantitative estimation of defect areas, providing valuable information for industrial applications. The findings indicate that segmentation-based approaches offer greater robustness and more reliable defect localization in challenging imaging environments.