Scholarlink

<p>Real-time and accurate detection of tobacco leaf diseases and damage is essential for agricultural modernization. However, complex field backgrounds, highly similar pathological characteristics, and the inherent trade-off between high precision and lightweight design remain significant challenges. To address these issues, we propose MCAD-DETR (<b>M</b>ulti-scale <b>C</b>omplementary <b>A</b>dditive <b>D</b>iffusion <b>De</b>tection <b>Tr</b>ansformer), a lightweight real-time tobacco leaf disease and damage detection model. First, we construct an Efficient Feature Complementary Mapping Network (EFCMNet) as the feature extraction network, which enhances feature extraction capabilities while maintaining a low-parameter count. We also introduce the Efficient Additive Attention Block (EAA Block) to decouple high- and low-frequency features, thereby strengthening the representation of pathological details and suppressing background interference. Additionally, a Tri-Focal Diffusion Feature Pyramid Network (TFDFPN) is designed to improve multi-scale detection performance and optimize model efficiency. To address class imbalance and small-object detection in the dataset, we propose a multi-dimensional joint loss function. Experimental results demonstrate that MCAD-DETR achieves an <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\({\text {mAP}}_{50}\)</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mtext>mAP</mtext> <mn>50</mn> </msub> </math></EquationSource> </InlineEquation> of 89.3% and a Recall of 83.0%, outperforming the RT-DETR baseline by 2.0% and 1.5%, respectively. The model parameters are reduced from 19.9 to 6.48M, the storage size decreases from 77.0 to 25.3MB, and the computational cost is lowered from 57.0 to 46.1 GFLOPs, representing reductions of 67.44%, 67.14%, and 19.12%, respectively. Compared with existing methods, MCAD-DETR provides superior detection accuracy with substantially lower complexity, making it well-suited for real-time tobacco disease and damage detection applications. Core code and dataset are available at <a href="https://github.com/ghshow5007/MCAD-DETR">https://github.com/ghshow5007/MCAD-DETR</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

MCAD-DETR: a real-time lightweight tobacco leaf disease and damage detection model

  • 期刊论文

  • 发表日期：

    2026年6月27日

查看全文
  • Wenjun Zhao,
  • Hao Wu,
  • Hui Gao,
  • Haofeng Wang

摘要

Real-time and accurate detection of tobacco leaf diseases and damage is essential for agricultural modernization. However, complex field backgrounds, highly similar pathological characteristics, and the inherent trade-off between high precision and lightweight design remain significant challenges. To address these issues, we propose MCAD-DETR (Multi-scale Complementary Additive Diffusion Detection Transformer), a lightweight real-time tobacco leaf disease and damage detection model. First, we construct an Efficient Feature Complementary Mapping Network (EFCMNet) as the feature extraction network, which enhances feature extraction capabilities while maintaining a low-parameter count. We also introduce the Efficient Additive Attention Block (EAA Block) to decouple high- and low-frequency features, thereby strengthening the representation of pathological details and suppressing background interference. Additionally, a Tri-Focal Diffusion Feature Pyramid Network (TFDFPN) is designed to improve multi-scale detection performance and optimize model efficiency. To address class imbalance and small-object detection in the dataset, we propose a multi-dimensional joint loss function. Experimental results demonstrate that MCAD-DETR achieves an \({\text {mAP}}_{50}\) mAP 50 of 89.3% and a Recall of 83.0%, outperforming the RT-DETR baseline by 2.0% and 1.5%, respectively. The model parameters are reduced from 19.9 to 6.48M, the storage size decreases from 77.0 to 25.3MB, and the computational cost is lowered from 57.0 to 46.1 GFLOPs, representing reductions of 67.44%, 67.14%, and 19.12%, respectively. Compared with existing methods, MCAD-DETR provides superior detection accuracy with substantially lower complexity, making it well-suited for real-time tobacco disease and damage detection applications. Core code and dataset are available at https://github.com/ghshow5007/MCAD-DETR.