Scholarlink

<p>Real-time diagnosis of grape leaf diseases holds significant importance in agricultural production. The main challenges in this field currently are the diversity of disease scales and achieving model lightweighting while maintaining high accuracy. To overcome these challenges, we present a lightweight grape leaf disease detection model, DBPC-DETR. First, the proposed model integrates and improves the feature extraction modules of two lightweight backbones, proposing a novel dual backbone network of PPHGNetV2 and CSPnet (DBPCNet). This network performs multi-feature interactive fusion of grape disease leaves while reducing computational redundancy. Next, utilize the efficient attention mechanism HiLo to capture high-frequency and low-frequency features of multi-scale features. Finally, to mitigate the performance degradation caused by small-scale diseases, we abandon the conventional feature pyramid architecture and propose a Context-Guided Spatial Cross-Scale Fusion (CGSCF) network, which incorporates a novel approach for small-object detection. Additionally, pruning and Masked Generative Distillation (MGD) methods are employed to further optimize redundant channels in the DBPC-DETR network, thereby accelerating detection speed while preserving accuracy. Experimental results on the self-constructed dataset demonstrate that DBPC-DETR achieves a detection accuracy of 84.4% and a mean average precision of 87.6%, while maintaining a frame rate of 158 FPS. The model parameters are reduced to 9.3M, with a computational cost of 16.0 gigaflops per second, corresponding to reductions of 53.1% and 71.9%, respectively, relative to the RT-DETR model. This not only furnishes technical support for grape disease monitoring but also yields valuable insights for disease detection in other crop species.</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

DBPC-DETR: a lightweight dual backbone network for real-time identification of multi-scale grape leaf diseases under field conditions

  • 期刊论文

  • 发表日期：

    2026年1月10日

查看全文
  • Dayu Xu,
  • Xiaobo Mao,
  • Fang Xia

摘要

Real-time diagnosis of grape leaf diseases holds significant importance in agricultural production. The main challenges in this field currently are the diversity of disease scales and achieving model lightweighting while maintaining high accuracy. To overcome these challenges, we present a lightweight grape leaf disease detection model, DBPC-DETR. First, the proposed model integrates and improves the feature extraction modules of two lightweight backbones, proposing a novel dual backbone network of PPHGNetV2 and CSPnet (DBPCNet). This network performs multi-feature interactive fusion of grape disease leaves while reducing computational redundancy. Next, utilize the efficient attention mechanism HiLo to capture high-frequency and low-frequency features of multi-scale features. Finally, to mitigate the performance degradation caused by small-scale diseases, we abandon the conventional feature pyramid architecture and propose a Context-Guided Spatial Cross-Scale Fusion (CGSCF) network, which incorporates a novel approach for small-object detection. Additionally, pruning and Masked Generative Distillation (MGD) methods are employed to further optimize redundant channels in the DBPC-DETR network, thereby accelerating detection speed while preserving accuracy. Experimental results on the self-constructed dataset demonstrate that DBPC-DETR achieves a detection accuracy of 84.4% and a mean average precision of 87.6%, while maintaining a frame rate of 158 FPS. The model parameters are reduced to 9.3M, with a computational cost of 16.0 gigaflops per second, corresponding to reductions of 53.1% and 71.9%, respectively, relative to the RT-DETR model. This not only furnishes technical support for grape disease monitoring but also yields valuable insights for disease detection in other crop species.