Scholarlink

<p>To address the limitations of convolutional neural networks (CNNs) in current agricultural pest classification tasks, particularly their insufficient performance and lack of multi-scale feature extraction capability, this paper proposes a novel CNN architecture named <b>RepMSNet</b>. By employing parallel convolutions with varying kernel sizes, RepMSNet enhances multi-scale feature extraction. Additionally, structural re-parameterization is introduced to streamline the model architecture, further improving inference speed. Experimental results on the IP102 dataset demonstrate that RepMSNet achieves a classification accuracy of 72.0%, which rises to76.3%after ImageNet-1&#xa0;K pre-training and fine-tuning, surpassing existing single-model methods based on CNNs or Vision Transformers (ViTs). In terms of inference speed, RepMSNet reaches 134 FPS, outperforming ViT-based approaches and most CNN-based methods. Ablation studies provide a detailed analysis of the rationale behind each design component of RepMSNet. All model code and trained weights are open-sourced at: <a href="https://github.com/OnlyForWW/WorkForIP102">https://github.com/OnlyForWW/WorkForIP102</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

RepMSNet: multi-scale convolutional neural network combined with structural re-parameterization for the classification of agricultural pests

  • 期刊论文

  • 发表日期：

    2026年3月31日

查看全文
  • Junyan Wang,
  • Haoran Yu,
  • Feiyang Kang

摘要

To address the limitations of convolutional neural networks (CNNs) in current agricultural pest classification tasks, particularly their insufficient performance and lack of multi-scale feature extraction capability, this paper proposes a novel CNN architecture named RepMSNet. By employing parallel convolutions with varying kernel sizes, RepMSNet enhances multi-scale feature extraction. Additionally, structural re-parameterization is introduced to streamline the model architecture, further improving inference speed. Experimental results on the IP102 dataset demonstrate that RepMSNet achieves a classification accuracy of 72.0%, which rises to76.3%after ImageNet-1 K pre-training and fine-tuning, surpassing existing single-model methods based on CNNs or Vision Transformers (ViTs). In terms of inference speed, RepMSNet reaches 134 FPS, outperforming ViT-based approaches and most CNN-based methods. Ablation studies provide a detailed analysis of the rationale behind each design component of RepMSNet. All model code and trained weights are open-sourced at: https://github.com/OnlyForWW/WorkForIP102.