Scholarlink

<p>Accurate plant disease detection is crucial for sustainable agriculture to reduce economic losses and optimize resources. Addressing challenges like insufficient valid samples, weak disease feature extraction, and accuracy-speed trade-off, we propose a novel MFVNet based on YOLOv9-T for accurate plant disease detection in the Intelligent Internet of Things. To enhance the quality of invalid disease samples, a Stable Diffusion-based data augmentation method is presented. In MFVNet’s backbone, a multi-scale edge information enhancer (MEIE) is designed to improve the capacity to capture edge information and multi-scale features, enabling an accurate distinction between disease regions from background noise. Then, a Re-FCA module is developed that integrates adaptive fine-grained channel attention into a reparameterized network with Cross-Stage Partial and Efficient Layer Aggregation Networks, effectively merging global and local information and refining disease feature extraction. Finally, a lightweight VanillaBlock-based pyramid network (VBPN) is introduced in the neck for feature aggregation. Its efficient computation and nonlinear enhancement drastically reduce computational overhead while enhancing detection accuracy. Experiments show that YOLOv9-T trained with the enhanced Stable-PlantDoc dataset achieved a 13.9% higher mAP50 than the original PlantDoc dataset trained one. On the Stable-PlantDoc dataset, MFVNet reduces FLOPs by 54.2%, increases mAP50 by 4.2%, and mAP50-95 by 3.7% compared to YOLOv9-T. Furthermore, MFVNet outperforms the state-of-the-art methods on the Stable-PlantDoc and AI Challenger 2018 datasets. For example, MFVNet achieves 81.6% mAP50 on the Stable-PlantDoc dataset, outperforming YOLOv10n by 5.7%, with 2.73 M parameters and 132 FPS. These results demonstrate that MFVNet achieves better accuracy-speed trade-off for plant disease detection.</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

MFVNet: a novel lightweight deep neural network for sustainable agriculture plant disease detection

  • 期刊论文

  • 发表日期：

    2026年5月11日

查看全文
  • Yunfan Chen,
  • Teng Gong,
  • Xiangkui Wan,
  • Yuting Li

摘要

Accurate plant disease detection is crucial for sustainable agriculture to reduce economic losses and optimize resources. Addressing challenges like insufficient valid samples, weak disease feature extraction, and accuracy-speed trade-off, we propose a novel MFVNet based on YOLOv9-T for accurate plant disease detection in the Intelligent Internet of Things. To enhance the quality of invalid disease samples, a Stable Diffusion-based data augmentation method is presented. In MFVNet’s backbone, a multi-scale edge information enhancer (MEIE) is designed to improve the capacity to capture edge information and multi-scale features, enabling an accurate distinction between disease regions from background noise. Then, a Re-FCA module is developed that integrates adaptive fine-grained channel attention into a reparameterized network with Cross-Stage Partial and Efficient Layer Aggregation Networks, effectively merging global and local information and refining disease feature extraction. Finally, a lightweight VanillaBlock-based pyramid network (VBPN) is introduced in the neck for feature aggregation. Its efficient computation and nonlinear enhancement drastically reduce computational overhead while enhancing detection accuracy. Experiments show that YOLOv9-T trained with the enhanced Stable-PlantDoc dataset achieved a 13.9% higher mAP50 than the original PlantDoc dataset trained one. On the Stable-PlantDoc dataset, MFVNet reduces FLOPs by 54.2%, increases mAP50 by 4.2%, and mAP50-95 by 3.7% compared to YOLOv9-T. Furthermore, MFVNet outperforms the state-of-the-art methods on the Stable-PlantDoc and AI Challenger 2018 datasets. For example, MFVNet achieves 81.6% mAP50 on the Stable-PlantDoc dataset, outperforming YOLOv10n by 5.7%, with 2.73 M parameters and 132 FPS. These results demonstrate that MFVNet achieves better accuracy-speed trade-off for plant disease detection.