Steel Surface Automatic Detection and Segmentation Based on an Improved Faster R-CNN

作者：Qingyuan Zhu, Tao Wei, Shanbo Xu, Yunlong Gao, Guifang Shao, Zhichao Li · 年份：2024 · DOI：10.23919/ccc63176.2024.10662649 · 被引用次数：3 · 研究领域：Industrial Vision Systems and Defect Detection

Despite the application of various deep learning networks and machine learning algorithms for steel surface defect detection, achieving satisfactory accuracy remains a challenge. To address issues such as dynamic image quality, insignificant features, and varied defect sizes, this paper proposes an enhanced Faster R-CNN method termed IFRC. To mitigate the impact of uneven illumination, normalization techniques and vertical mean removal are employed. Moreover, to augment the significance of defect features, dense connection is integrated into the VGG-16 architecture. Furthermore,an anchor parameter adjustment method based on K-means clustering is introduced to detect long, thin defects effectively. To enhance the segmentation of patches by eliminating holes and edge burrs, the Felzenswalb algorithm is utilized to refine the results. Experimental results conducted on the NEU-DET dataset demonstrate the superior performance of IFRC compared to Faster R-CNN. Moreover, IFRC achieves better segmentation results than DLRSD.