Scholarlink

<p>Corn holds a significant position in grain cultivation and occupies a notable place in agricultural production systems. The objective of the paper is to enhance the accuracy of lightweight convolutional neural networks for the recognition of corn diseases and to facilitate their implementation on agricultural Internet of Things (IoT). The complex teacher network was constructed based on ResNet50 and incorporated an instance relationship graph that included instance features, instance relationships, and feature space transformations to guide the student network, which was designed with reference to ResNet18. The model was augmented with a Convolutional Block Attention Module (CBAM) following the first convolutional layer to improve its adaptive performance. Additionally, the activation function was modified to the Swish Linear Unit (SILU) to enhance the model’s ability to capture complex data patterns. A spatial pyramid pooling module was integrated after the fifth convolutional layer, enabling the transformation of feature maps of varying dimensions into a fixed-length feature vector. To further optimize the model, pruning operations were implemented to reduce the complexity of the original architecture, resulting in a simplified representation. The experimental results based on the corn disease dataset indicate that the student network achieved an accuracy of 98.25% following knowledge distillation, with a model parameter size of less than 5&#xa0;MB and an average inference time of under 40&#xa0;ms. This system effectively meets the requirements of agricultural IoT mobile devices.</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

IRG-ResNet: distillation model for corn disease recognition

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月18日

查看全文
  • Shaoqiu Zhu,
  • Lujie Bai,
  • Haitao Gao

摘要

Corn holds a significant position in grain cultivation and occupies a notable place in agricultural production systems. The objective of the paper is to enhance the accuracy of lightweight convolutional neural networks for the recognition of corn diseases and to facilitate their implementation on agricultural Internet of Things (IoT). The complex teacher network was constructed based on ResNet50 and incorporated an instance relationship graph that included instance features, instance relationships, and feature space transformations to guide the student network, which was designed with reference to ResNet18. The model was augmented with a Convolutional Block Attention Module (CBAM) following the first convolutional layer to improve its adaptive performance. Additionally, the activation function was modified to the Swish Linear Unit (SILU) to enhance the model’s ability to capture complex data patterns. A spatial pyramid pooling module was integrated after the fifth convolutional layer, enabling the transformation of feature maps of varying dimensions into a fixed-length feature vector. To further optimize the model, pruning operations were implemented to reduce the complexity of the original architecture, resulting in a simplified representation. The experimental results based on the corn disease dataset indicate that the student network achieved an accuracy of 98.25% following knowledge distillation, with a model parameter size of less than 5 MB and an average inference time of under 40 ms. This system effectively meets the requirements of agricultural IoT mobile devices.