Scholarlink

<p>Automated bird species identification plays a crucial role in biodiversity monitoring. However, fine-grained bird image classification (FBIC) faces a significant challenge: balancing the capture of subtle diagnostic phenotypic traits with the computational efficiency required for real-time field deployment. To address this, this paper proposes EffBirdNet, a lightweight network that optimizes the balance between high identification accuracy and low computational cost. First, the MBConv-S<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(^{2}\)</EquationSource> <EquationSource Format="MATHML"><math> <mmultiscripts> <mrow /> <mrow /> <mn>2</mn> </mmultiscripts> </math></EquationSource> </InlineEquation> AM module is designed to capture discriminative local details by enhancing spatial perception. Second, to maintain low computational overhead, a multi-feature fusion network is constructed using efficient depthwise separable convolutions to integrate micro-scale morphological cues and macro-scale ecological context. These features are adaptively aggregated by a gated fusion block (GFBlock). Experimental results demonstrate that EffBirdNet achieves 95.8% accuracy on the Xixi Wetland Bird Dataset and surpasses most existing models on CUB-200-2011. Crucially, the model is deployed on an Orange Pi 5 Plus edge device using NPU acceleration. It achieves a real-time inference speed of 28.9 FPS, demonstrating its edge deployment capability and suitability for high-performance in situ biodiversity monitoring systems.</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

EffBirdNet: an efficient and lightweight network based on synergistic feature fusion for fine-grained bird image classification

  • 期刊论文

  • 发表日期：

    2026年5月4日

查看全文
  • Jian Gao,
  • Beiping Hou,
  • Wen Zhu,
  • Shuzong Xie,
  • Rui Liang,
  • Aihua Yu

摘要

Automated bird species identification plays a crucial role in biodiversity monitoring. However, fine-grained bird image classification (FBIC) faces a significant challenge: balancing the capture of subtle diagnostic phenotypic traits with the computational efficiency required for real-time field deployment. To address this, this paper proposes EffBirdNet, a lightweight network that optimizes the balance between high identification accuracy and low computational cost. First, the MBConv-S \(^{2}\) 2 AM module is designed to capture discriminative local details by enhancing spatial perception. Second, to maintain low computational overhead, a multi-feature fusion network is constructed using efficient depthwise separable convolutions to integrate micro-scale morphological cues and macro-scale ecological context. These features are adaptively aggregated by a gated fusion block (GFBlock). Experimental results demonstrate that EffBirdNet achieves 95.8% accuracy on the Xixi Wetland Bird Dataset and surpasses most existing models on CUB-200-2011. Crucially, the model is deployed on an Orange Pi 5 Plus edge device using NPU acceleration. It achieves a real-time inference speed of 28.9 FPS, demonstrating its edge deployment capability and suitability for high-performance in situ biodiversity monitoring systems.