[Paper Review] Image Recognition of Tea Leaf Diseases Based on Convolutional Neural Network

Xiaoxiao Sun, Shaomin Mu+3 more|arXiv (Cornell University)|Jan 9, 2019

Smart Agriculture and AI2 references17 citations

TL;DR

This paper proposes a convolutional neural network (CNN)-based approach for automated image recognition of tea leaf diseases, employing image preprocessing, data augmentation, and hyperparameter tuning to achieve 93.75% classification accuracy—surpassing SVM (89.36%) and BP neural networks (87.69%). The method effectively improves detection efficiency and accuracy in tea disease identification.

ABSTRACT

In order to identify and prevent tea leaf diseases effectively, convolution neural network (CNN) was used to realize the image recognition of tea disease leaves. Firstly, image segmentation and data enhancement are used to preprocess the images, and then these images were input into the network for training. Secondly, to reach a higher recognition accuracy of CNN, the learning rate and iteration numbers were adjusted frequently and the dropout was added properly in the case of over-fitting. Finally, the experimental results show that the recognition accuracy of CNN is 93.75%, while the accuracy of SVM and BP neural network is 89.36% and 87.69% respectively. Therefore, the recognition algorithm based on CNN is better in classification and can improve the recognition efficiency of tea leaf diseases effectively.

Motivation & Objective

• To develop an automated system for identifying tea leaf diseases using deep learning.
• To improve classification accuracy beyond traditional machine learning models like SVM and BP neural networks.
• To enhance recognition efficiency through image preprocessing and data augmentation techniques.
• To address overfitting in deep learning models using dropout and adaptive learning rate scheduling.
• To provide a practical, high-accuracy solution for early detection of tea leaf diseases in agricultural settings.

Proposed method

• Image segmentation and data augmentation were applied to preprocess tea leaf images, increasing dataset diversity and quality.
• A custom convolutional neural network architecture was trained using stochastic gradient descent with adaptive learning rate adjustment.
• Dropout layers were strategically inserted in the fully connected layers to mitigate overfitting during training.
• Hyperparameters such as learning rate and iteration count were tuned iteratively to optimize model convergence and performance.
• The trained CNN was evaluated on a real-world tea leaf disease dataset to assess classification accuracy.
• Comparative experiments were conducted using SVM and BP neural networks to benchmark performance.

Experimental results

Research questions

• RQ1 Can a CNN-based model achieve higher classification accuracy than traditional machine learning models in tea leaf disease recognition?
• RQ2 How do data augmentation and image preprocessing techniques affect the performance of deep learning models in this context?
• RQ3 To what extent does dropout regularization reduce overfitting in the CNN model for tea leaf disease classification?
• RQ4 What impact do learning rate and iteration count adjustments have on model convergence and accuracy?
• RQ5 Can the proposed method significantly improve the efficiency and reliability of tea leaf disease detection in real-world agricultural applications?

Key findings

• The proposed CNN model achieved a recognition accuracy of 93.75%, significantly outperforming SVM (89.36%) and BP neural network (87.69%).
• Data augmentation and image preprocessing enhanced the model's robustness and generalization capability.
• The use of dropout layers effectively reduced overfitting during training, especially in the fully connected layers.
• Iterative tuning of learning rate and iteration count improved model convergence and final performance.
• The CNN-based approach demonstrated superior classification performance and scalability for tea leaf disease detection.
• The results confirm that deep learning models like CNN are more effective than classical machine learning methods for complex image-based disease recognition in agriculture.