Zhang Xiaoli, Liang Kun, Zhang Yiying
College of Artificial Intelligence, Tianjin University of Science & Technology, Tianjin, China.
Front Plant Sci. 2024 Nov 21;15:1443815. doi: 10.3389/fpls.2024.1443815. eCollection 2024.
Plant pest and disease management is an important factor affecting the yield and quality of crops, and due to the rich variety and the diagnosis process mostly relying on experts' experience, there are problems of low diagnosis efficiency and accuracy. For this, we proposed a Plant pest and Disease Lightweight identification Model by fusing Tensor features and Knowledge distillation (PDLM-TK). First, a Lightweight Residual Blocks based on Spatial Tensor (LRB-ST) is constructed to enhance the perception and extraction of shallow detail features of plant images by introducing spatial tensor. And the depth separable convolution is used to reduce the number of model parameters to improve the diagnosis efficiency. Secondly, a Branch Network Fusion with Graph Convolutional features (BNF-GC) is proposed to realize image super-pixel segmentation by using spanning tree clustering based on pixel features. And the graph convolution neural network is utilized to extract the correlation features to improve the diagnosis accuracy. Finally, we designed a Model Training Strategy based on knowledge Distillation (MTS-KD) to train the pest and disease diagnosis model by building a knowledge migration architecture, which fully balances the accuracy and diagnosis efficiency of the model. The experimental results show that PDLM-TK performs well in three plant pest and disease datasets such as Plant Village, with the highest classification accuracy and F1 score of 96.19% and 94.94%. Moreover, the model execution efficiency performs better compared to lightweight methods such as MobileViT, which can quickly and accurately diagnose plant diseases.
植物病虫害管理是影响作物产量和质量的重要因素，由于种类丰富且诊断过程大多依赖专家经验，存在诊断效率和准确性低的问题。为此，我们提出了一种融合张量特征和知识蒸馏的植物病虫害轻量级识别模型（PDLM-TK）。首先，构建基于空间张量的轻量级残差块（LRB-ST），通过引入空间张量增强对植物图像浅层细节特征的感知和提取，并使用深度可分离卷积减少模型参数数量以提高诊断效率。其次，提出基于图卷积特征的分支网络融合（BNF-GC），利用基于像素特征的生成树聚类实现图像超像素分割，并使用图卷积神经网络提取相关特征以提高诊断准确性。最后，设计基于知识蒸馏的模型训练策略（MTS-KD），通过构建知识迁移架构训练病虫害诊断模型，充分平衡模型的准确性和诊断效率。实验结果表明，PDLM-TK在Plant Village等三个植物病虫害数据集上表现良好，最高分类准确率和F1分数分别为96.19%和94.94%。此外，与MobileViT等轻量级方法相比，该模型执行效率更高，能够快速准确地诊断植物病害。
