Scholarlink

<p>Circle detection plays a crucial role in diverse applications, including industrial inspection, medical imaging, and robotics. However, conventional methods such as the Circular Hough Transform (CHT), Randomized Circle Detection (RCD), and ED Circles remain vulnerable to noise, false positives, and computational inefficiency. This paper introduces a two-stage circle detection framework that integrates a deep learning model (YOLOv8) with classical methods. In the first stage, YOLOv8 provides bounding box data to restrict the search space, while in the second stage, conventional algorithms refine circle center and radius estimation. Bounding box information is further applied in post-processing to eliminate false detections. Experimental evaluation on three datasets (Mini, PCB, and Noisy) demonstrates that the proposed method achieves an average F-measure improvement of 9.07% and reduces the center location MSE and radius estimation MSE by 3.18% and 6.30%, respectively, compared with conventional approaches. The results indicate enhanced performance under noisy conditions and in the presence of overlapping circles. The proposed framework extends current circle detection methods without modifying their core principles and is applicable to practical engineering scenarios.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Two-Stage Circle Detection Framework Integrating YOLOv8 and Conventional Algorithms with Bounding Box-Based Refinement

  • 期刊论文

  • 发表日期：

    2026年3月2日

查看全文
  • Athtayu Yuthong,
  • Pornchai Phukpattaranont,
  • Kanadit Chetpattananondh

摘要

Circle detection plays a crucial role in diverse applications, including industrial inspection, medical imaging, and robotics. However, conventional methods such as the Circular Hough Transform (CHT), Randomized Circle Detection (RCD), and ED Circles remain vulnerable to noise, false positives, and computational inefficiency. This paper introduces a two-stage circle detection framework that integrates a deep learning model (YOLOv8) with classical methods. In the first stage, YOLOv8 provides bounding box data to restrict the search space, while in the second stage, conventional algorithms refine circle center and radius estimation. Bounding box information is further applied in post-processing to eliminate false detections. Experimental evaluation on three datasets (Mini, PCB, and Noisy) demonstrates that the proposed method achieves an average F-measure improvement of 9.07% and reduces the center location MSE and radius estimation MSE by 3.18% and 6.30%, respectively, compared with conventional approaches. The results indicate enhanced performance under noisy conditions and in the presence of overlapping circles. The proposed framework extends current circle detection methods without modifying their core principles and is applicable to practical engineering scenarios.