Scholarlink

<p>Real-time detection of small and fast-moving objects remains challenging due to limited spatial evidence, motion blur, and background interference. Shuttlecock detection is particularly difficult because of its tiny footprint and rapid, irregular motion. To address these issues, we develop a lightweight yet expressive architecture tailored for efficient feature extraction and multi-scale representation learning. The core <Emphasis FontCategory="NonProportional">GDModule</Emphasis> introduces ratio-controlled feature decomposition and depthwise–separable refinement, yielding more discriminative high-frequency cues at low computational cost. Efficient Channel Attention (<Emphasis FontCategory="NonProportional">ECA</Emphasis>) is integrated throughout the network to enhance inter-channel dependency modeling, while Asymmetric Convolution (<Emphasis FontCategory="NonProportional">AsyConv</Emphasis>) strengthens sensitivity to directional motion patterns. We further design a compact multi-scale <Emphasis FontCategory="NonProportional">EnFusion</Emphasis> module that adaptively promotes cross-scale feature interaction. Extensive experiments show that the proposed framework markedly improves accuracy, precision, recall, and F1 score over the baseline, while significantly reducing parameters and sustaining real-time performance. These results demonstrate the effectiveness and practical applicability of our method for shuttlecock detection in fast-paced sports 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 lightweight attention-guided multi-scale fusion network for real-time shuttlecock detection

  • 期刊论文

  • 发表日期：

    2026年1月14日

查看全文
  • Gang Li,
  • FengYi Wang,
  • YongQiang Fan,
  • YunKang Nie

摘要

Real-time detection of small and fast-moving objects remains challenging due to limited spatial evidence, motion blur, and background interference. Shuttlecock detection is particularly difficult because of its tiny footprint and rapid, irregular motion. To address these issues, we develop a lightweight yet expressive architecture tailored for efficient feature extraction and multi-scale representation learning. The core GDModule introduces ratio-controlled feature decomposition and depthwise–separable refinement, yielding more discriminative high-frequency cues at low computational cost. Efficient Channel Attention (ECA) is integrated throughout the network to enhance inter-channel dependency modeling, while Asymmetric Convolution (AsyConv) strengthens sensitivity to directional motion patterns. We further design a compact multi-scale EnFusion module that adaptively promotes cross-scale feature interaction. Extensive experiments show that the proposed framework markedly improves accuracy, precision, recall, and F1 score over the baseline, while significantly reducing parameters and sustaining real-time performance. These results demonstrate the effectiveness and practical applicability of our method for shuttlecock detection in fast-paced sports scenarios.