•  
  •  
 

Abstract

Detection of defects on low-contrast steel surfaces, especially crazing and rolled-in-scale, remains a major challenge due to their visual similarity to background patterns. Although state-of-the-art methods have achieved high accuracy through complex architectural adjustments, the contribution of preprocessing techniques has not been thoroughly investigated. This study investigates pre-processing-based improvements to Faster R-CNN by combining Bilateral Filtering to reduce noise, CLAHE to enhance local contrast, CIoU Loss for more effective bounding box regression, and customized anchor settings for irregular defect configurations. Evaluated using the NEU-DET dataset, our BF-CIoU Faster R-CNN model achieved a mAP@50 score of 72.32%, with an AP of 43.74% for crazing and 53.04% for rolled-in-scale. Although these results fall short of the performance of state-of-the-art architectures that utilize feature fusion and attention mechanisms (80.2% mAP), our approach demonstrates that preprocessing improvements alone can yield competitive baseline performance without additional architectural complexity. This study confirms the effectiveness of Bilateral Filtering and CLAHE in removing defective signals, while highlighting the need for more advanced feature-extraction modules to achieve higher accuracy. Further research will examine hybrid approaches that combine preprocessing with attention-based architectures for steel inspection systems in industry.

First Page

133

Last Page

144

Download

Share

COinS