Scholarlink

Background <p>Wheat diseases significantly impair production efficiency and grain quality in the wheat industry. In recent research, deep learning techniques have been widely applied to plant disease detection. However, wheat disease images collected in field conditions often face complex backgrounds and diverse lesion shapes, making accurate disease classification difficult. In real-world applications, agricultural disease recognition systems must also deal with limited computational resources and edge device constraints, emphasizing the need for lightweight methods.</p> Results <p>To solve these challenges, this paper introduces a lightweight dual-stream learning (LDSL) framework for wheat disease detection. The framework adopts a unique global–local dual-stream architecture that combines global semantic understanding with local discriminative analysis. The global learning stream extracts comprehensive semantic features and generates saliency maps to highlight key regions, while the local learning stream performs fine-grained inspection of these regions using a novel dynamic-static dual attention (DSDA) mechanism. Additionally, a Kullback–Leibler (KL) divergence perturbation strategy is implemented during training to boost the LDSL framework’s robustness in noisy and complex settings. Experimental results show that the proposed LDSL framework achieves an accuracy of 94.44%, a precision of 94.47%, a recall of 94.44%, and an F1-score of 94.45%, outperforming several mainstream classification models in wheat disease recognition, such as ConvNeXt-T (92.66% accuracy, 92.69% precision, 92.66% recall, and 92.63% F1). The proposed LDSL framework is lightweight, using only 4.41&#xa0;M parameters and 1.71G FLOPs. On the NVIDIA Jetson Orin Nano, it requires just 15.99&#xa0;MB of storage, 39.49&#xa0;MB of peak memory, and achieves an inference latency of 234.76&#xa0;ms/image, demonstrating good potential for real-world deployment.</p> Conclusions <p>This study provides a novel detection framework for wheat disease research, which significantly improves various classification metrics. With low parameter and computation costs, the framework demonstrates good potential for practical deployment.</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

LDSL framework: a lightweight dual-stream learning framework for wheat disease detection

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年10月24日

查看全文
  • Lei Feng,
  • Mingliang Li,
  • Guanshi Ye,
  • Qinghai Wu,
  • Chunyu Ning,
  • You Tang

摘要

Background

Wheat diseases significantly impair production efficiency and grain quality in the wheat industry. In recent research, deep learning techniques have been widely applied to plant disease detection. However, wheat disease images collected in field conditions often face complex backgrounds and diverse lesion shapes, making accurate disease classification difficult. In real-world applications, agricultural disease recognition systems must also deal with limited computational resources and edge device constraints, emphasizing the need for lightweight methods.

Results

To solve these challenges, this paper introduces a lightweight dual-stream learning (LDSL) framework for wheat disease detection. The framework adopts a unique global–local dual-stream architecture that combines global semantic understanding with local discriminative analysis. The global learning stream extracts comprehensive semantic features and generates saliency maps to highlight key regions, while the local learning stream performs fine-grained inspection of these regions using a novel dynamic-static dual attention (DSDA) mechanism. Additionally, a Kullback–Leibler (KL) divergence perturbation strategy is implemented during training to boost the LDSL framework’s robustness in noisy and complex settings. Experimental results show that the proposed LDSL framework achieves an accuracy of 94.44%, a precision of 94.47%, a recall of 94.44%, and an F1-score of 94.45%, outperforming several mainstream classification models in wheat disease recognition, such as ConvNeXt-T (92.66% accuracy, 92.69% precision, 92.66% recall, and 92.63% F1). The proposed LDSL framework is lightweight, using only 4.41 M parameters and 1.71G FLOPs. On the NVIDIA Jetson Orin Nano, it requires just 15.99 MB of storage, 39.49 MB of peak memory, and achieves an inference latency of 234.76 ms/image, demonstrating good potential for real-world deployment.

Conclusions

This study provides a novel detection framework for wheat disease research, which significantly improves various classification metrics. With low parameter and computation costs, the framework demonstrates good potential for practical deployment.