Identification of Diseases and Pests in Tea Leaves based on Deep Learning Techniques

Abstract

Tea is an significant crop and deeply loved by individuals.However, various climatic factors and other factors may lead to tea diseases, resulting in the decline of tea quality and yield. In earlier times, identification of tea leaf diseases and pests was manual and inefficient.

With the increasing application of AI(artificial intelligence),deep learning and image recognition technology in the field of agriculture, deep learning models have achieved a certain improvement in accuracy of pests and diseases classification and recognition in tea leaf. However, there remains potential for additional enhancement in both model parameters and overall performance,so this paper introduces a method with improved efficiency and precision for intelligent identification in tea leaf diseases and pests.The main research contents and conclusions of this thesis are as follows:

(1)Three classification models VGG16, Resnet50 and DenseNet169 were used to categorize and identify common amongst diseases of tea leaf. By accommodating the relevant parameters of these three identification models and the learning rates, the performances of the three convolutional neural networks were analyzed and compared. The simulated test results showed that VGG16, Resnet50, and DenseNet169 achieved the highest accuracy rates of 90.6%, 95.2%, and 99.0%, respectively.DenseNet169 provided the fastest convergence speed and the highest accuracy among the models.

(2)The deep learning target detection models were proposed, for automated recognition of common tea leaf diseases and pests. Firstly,this study proposed an improved YOLOv5 (You Only Look Once version 5) algorithm method based on deep learning,which improved the model structure of YOLOv5 and integrate Omni Dimensional Dynamic Convolution (ODConv) and Convolutional Block Attention Module (CBAM) attention mechanisms into YOLOv5. The results show that the accuracy of the improved YOLOv5 is higher than that before the improvement and superior to other methods, including Faster R-CNN(Faster region-based convolutional neural networks), SSD(Single Shot MultiBox Detector) .Then, the latest models of deep learning target detection YOLO such as YOLOv10s, YOLOv9s and YOLOv8s were applied to automatic identification of tea pests and diseases. The comparison of performance metrics indicated that the model based on YOLOv10s performed the best. As shown by the test evaluation results, precision, recall, mAP50(mean of Average Precision), F1-Score,these values are all higher than those achieved by others.

(3) Using the optimal YOLOv10s model, a software system with a graphical interface was developed using Python and PyQt5. It effectively supports detection of data such as images, videos, and cameras, and also supports saving the detection results.It is capable of real-time and off-site detection of tea diseases.

Based on this proposed model with deep learning techniques, identification of tea leaf diseases and pests will be significantly improved for all the terms of higher efficiency, less costs, as well as enhanced quality and sustainability of tea production.In addition, the models showed the possibility of the precise diagnosis and identification of other crop leaf diseases as well.

-