Scholarlink

<p>Sign language serves as a vital communication medium for individuals with hearing impairments, yet conventional convolutional architectures often suffer from significant feature degradation, particularly in high-frequency details and multi-scale feature representation. This paper introduces a novel method, YOLO-FPD, which leverages Fast Fourier Transform (FFT) to construct a dual-domain decoupled feature representation framework. A Parallel Frequency-domain Attention Module (PFMLP) is integrated to dynamically enhance key responses in both frequency and spatial domains, while a Dynamic Heterogeneous Multi-scale Cross-stage Fusion Module (DHMCS-FM) is proposed to improve multi-scale and high-frequency gesture feature capture. Experimental results on public datasets demonstrate that YOLO-FPD achieves state-of-the-art accuracy (mAP@50 of 93.2% on the ASL dataset and 92.4% on the Expression dataset) while maintaining real-time performance, outperforming several mainstream models. Our approach not only addresses the challenges of high-frequency detail loss and multi-scale feature representation but also establishes a collaborative mechanism between frequency-domain and spatial-domain processing, paving the way for more robust and efficient sign language recognition systems. <a href="https://github.com/wtc0214/YOLO-FPD">https://github.com/wtc0214/YOLO-FPD</a></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

Enhanced YOLO with spectral recalibration for accurate and real-time sign language detection

  • 期刊论文

  • 发表日期：

    2026年4月24日

查看全文
  • Yong Yang,
  • Tianci Wan,
  • Menglu Zhang,
  • Ling Guo

摘要

Sign language serves as a vital communication medium for individuals with hearing impairments, yet conventional convolutional architectures often suffer from significant feature degradation, particularly in high-frequency details and multi-scale feature representation. This paper introduces a novel method, YOLO-FPD, which leverages Fast Fourier Transform (FFT) to construct a dual-domain decoupled feature representation framework. A Parallel Frequency-domain Attention Module (PFMLP) is integrated to dynamically enhance key responses in both frequency and spatial domains, while a Dynamic Heterogeneous Multi-scale Cross-stage Fusion Module (DHMCS-FM) is proposed to improve multi-scale and high-frequency gesture feature capture. Experimental results on public datasets demonstrate that YOLO-FPD achieves state-of-the-art accuracy (mAP@50 of 93.2% on the ASL dataset and 92.4% on the Expression dataset) while maintaining real-time performance, outperforming several mainstream models. Our approach not only addresses the challenges of high-frequency detail loss and multi-scale feature representation but also establishes a collaborative mechanism between frequency-domain and spatial-domain processing, paving the way for more robust and efficient sign language recognition systems. https://github.com/wtc0214/YOLO-FPD