Existing deep learning techniques for tomato leaf disease recognition face several challenges, including external environmental interference, limited dataset size, imbalanced sample distribution, and overlapping characteristics among different diseases, which complicate accurate disease identification. Furthermore, complex models with a high number of parameters are often difficult to deploy on resource-constrained embedded devices. To address these challenges, this paper proposes a novel tomato leaf disease recognition method based on DGP-SNNet. Initially, to mitigate issues related to imbalanced samples and overfitting, we introduce a two-stage transfer learning technique alongside a partial convolution module (PConv) to decrease data dependency and enhance model stability. Subsequently, we propose a Global Grouped Location Attention (GGLA) mechanism that dynamically adapts to capture fine-grained disease information, thereby addressing the similarities between disease categories. Finally, we employ a joint compression method utilizing Network Slimming and Neuron Selectivity Transfer, which significantly reduces model size with minimal loss in accuracy. Experimental results demonstrate a classification accuracy of 99.55%, with FLOPs of 1011.88 MB and a parameter count of 4.93 MB. Compared to the baseline model, accuracy improved by 2.23%, FLOPs decreased by 63.39%, and the parameter count was reduced by 75.13%. Additionally, we achieved optimal performance through comparative analyses with other classical and state-of-the-art models, generalization experiments, and module effectiveness tests. In conclusion, the proposed method effectively recognizes various diseases in tomato leaves and offers a practical solution for the integration of deep learning into agricultural production processes.

中文翻译：

---

基于DGP-SNNet的番茄叶片病害鉴定


现有的用于番茄叶片病害识别的深度学习技术面临一些挑战，包括外部环境干扰、数据集大小有限、样本分布不平衡以及不同病害之间的特征重叠，这使得准确的病害识别变得复杂。此外，具有大量参数的复杂模型通常难以部署在资源受限的嵌入式设备上。为了应对这些挑战，本文提出了一种基于 DGP-SNNet 的新型番茄叶片病害识别方法。最初，为了缓解与不平衡样本和过拟合相关的问题，我们在部分卷积模块 （PConv） 旁边引入了两阶段迁移学习技术，以减少数据依赖性并增强模型稳定性。随后，我们提出了一种全局分组位置注意力 （GGLA） 机制，该机制可以动态适应以捕获细粒度的疾病信息，从而解决疾病类别之间的相似性。最后，我们采用了一种利用网络瘦身和神经元选择性转移的联合压缩方法，该方法显着减小了模型大小，而准确性损失最小。实验结果表明，分类准确率为 99.55%，FLOPs 为 1011.88 MB，参数计数为 4.93 MB。与基线模型相比，准确率提高了 2.23%，FLOPs 降低了 63.39%，参数数量减少了 75.13%。此外，我们通过与其他经典和最先进的模型、泛化实验和模块有效性测试的比较分析，实现了最佳性能。 总之，所提出的方法有效地识别了番茄叶片中的各种病害，并为将深度学习集成到农业生产过程中提供了实用的解决方案。