Scholarlink

<p>Edge detection in dynamic video environments remains challenging due to noise, texture clutter, and real-time processing constraints. To address this, we propose a novel fusion framework that synergistically combines multi-scale Gabor filtering with a directionally enhanced Sobel operator, followed by adaptive weighting based on local consistency. Unlike deep learning approaches, our method requires no training data and operates efficiently on standard hardware. Evaluated on the BSDS500 and DAVIS 2017 datasets, our approach achieves an F1-score of 88.8%, outperforming classical baselines such as Canny (81.2%) and Sobel (75.8%) by 7.6 and 13.0 percentage points, respectively. Moreover, it attains an edge localization error of only 1.42 pixels, demonstrating superior spatial precision. The integration of FFT-based and separable convolutions enables real-time performance at over 45 FPS on 480p video streams, making it suitable for resource-constrained applications like autonomous drones or embedded surveillance systems. A sensitivity analysis confirms robustness, with F1-score variations under ±1.5% across parameter perturbations. By bridging frequency-domain selectivity with gradient-based sharpness, our method delivers both high accuracy and computational efficiency—offering a practical solution for real-world edge-aware video analytics where reliability, speed, and interpretability are critical.</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

Research on real-time edge detection algorithm for video frames based on Gabor filtering and directional enhanced Sobel operator

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月7日

查看全文
  • Xianghua Gao,
  • Shangtao Gao

摘要

Edge detection in dynamic video environments remains challenging due to noise, texture clutter, and real-time processing constraints. To address this, we propose a novel fusion framework that synergistically combines multi-scale Gabor filtering with a directionally enhanced Sobel operator, followed by adaptive weighting based on local consistency. Unlike deep learning approaches, our method requires no training data and operates efficiently on standard hardware. Evaluated on the BSDS500 and DAVIS 2017 datasets, our approach achieves an F1-score of 88.8%, outperforming classical baselines such as Canny (81.2%) and Sobel (75.8%) by 7.6 and 13.0 percentage points, respectively. Moreover, it attains an edge localization error of only 1.42 pixels, demonstrating superior spatial precision. The integration of FFT-based and separable convolutions enables real-time performance at over 45 FPS on 480p video streams, making it suitable for resource-constrained applications like autonomous drones or embedded surveillance systems. A sensitivity analysis confirms robustness, with F1-score variations under ±1.5% across parameter perturbations. By bridging frequency-domain selectivity with gradient-based sharpness, our method delivers both high accuracy and computational efficiency—offering a practical solution for real-world edge-aware video analytics where reliability, speed, and interpretability are critical.