Scholarlink

<p>Accurate and real-time detection of maize pests and diseases plays a vital role in ensuring food security and advancing sustainable agriculture. However, existing detection methods suffer from limited performance on small targets, inadequate multi-scale feature fusion, and high computational costs, making them unsuitable for real-time agricultural monitoring applications. To overcome these challenges, this paper presents MCPDNet, an enhanced detection framework built upon YOLO11, specifically designed for efficient and real-time maize pest and disease detection. Specifically, a Grouped Multi-Scale Convolution (GMSConv) module is introduced to improve small target detection by extracting multi-scale features through parallel convolutions with varying kernel sizes, allowing the network to capture both fine-grained details and global contextual information simultaneously. Additionally, a Recurrent Adaptive Feature Pyramid Network (RAFPN) is developed to address insufficient cross-scale feature interaction, which leverages the Dynamic Aggregation for Scale Interaction (DASI) module to enable bidirectional information propagation and progressive feature refinement. Moreover, an Efficient Fused Detection Head (EFDHead) is designed to reduce computational burden by employing shared depthwise separable convolutions combined with learnable scale factors. Extensive experiments on the Corn-PD dataset demonstrate that MCPDNet achieves superior performance, attaining mAP50 of 96.8% and mAP50:95 of 81.5% with only 6.7 GFLOPs. Compared to the baseline YOLO11n, MCPDNet improves mAP50 by 3.5% and mAP50:95 by 3.7% while reducing model parameters by 3.9%, achieving an optimal balance between accuracy and efficiency. The inference speed reaches 78 FPS, significantly outperforming heavyweight models such as RT-DETR (102.5 GFLOPs, 45 FPS) while maintaining comparable computational cost to lightweight detectors. These results demonstrate that MCPDNet effectively addresses the accuracy-efficiency trade-off critical for real-world agricultural deployment, making it highly suitable for resource-constrained edge devices in precision agriculture applications.</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

MCPDNet: multi-scale feature fusion network for maize leaf disease and pest detection

  • 期刊论文

  • 发表日期：

    2026年3月12日

查看全文
  • Zhiming Li,
  • Yufen Li

摘要

Accurate and real-time detection of maize pests and diseases plays a vital role in ensuring food security and advancing sustainable agriculture. However, existing detection methods suffer from limited performance on small targets, inadequate multi-scale feature fusion, and high computational costs, making them unsuitable for real-time agricultural monitoring applications. To overcome these challenges, this paper presents MCPDNet, an enhanced detection framework built upon YOLO11, specifically designed for efficient and real-time maize pest and disease detection. Specifically, a Grouped Multi-Scale Convolution (GMSConv) module is introduced to improve small target detection by extracting multi-scale features through parallel convolutions with varying kernel sizes, allowing the network to capture both fine-grained details and global contextual information simultaneously. Additionally, a Recurrent Adaptive Feature Pyramid Network (RAFPN) is developed to address insufficient cross-scale feature interaction, which leverages the Dynamic Aggregation for Scale Interaction (DASI) module to enable bidirectional information propagation and progressive feature refinement. Moreover, an Efficient Fused Detection Head (EFDHead) is designed to reduce computational burden by employing shared depthwise separable convolutions combined with learnable scale factors. Extensive experiments on the Corn-PD dataset demonstrate that MCPDNet achieves superior performance, attaining mAP50 of 96.8% and mAP50:95 of 81.5% with only 6.7 GFLOPs. Compared to the baseline YOLO11n, MCPDNet improves mAP50 by 3.5% and mAP50:95 by 3.7% while reducing model parameters by 3.9%, achieving an optimal balance between accuracy and efficiency. The inference speed reaches 78 FPS, significantly outperforming heavyweight models such as RT-DETR (102.5 GFLOPs, 45 FPS) while maintaining comparable computational cost to lightweight detectors. These results demonstrate that MCPDNet effectively addresses the accuracy-efficiency trade-off critical for real-world agricultural deployment, making it highly suitable for resource-constrained edge devices in precision agriculture applications.