A Robust Deep-Learning-Based Detector for Real-Time Tomato Plant Diseases and Pests Recognition

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Plant diseases and pests pose significant threats to agriculture, causing substantial economic losses.
• Accurate and timely detection of these issues is crucial for effective management and crop yield preservation.
• Deep Neural Networks (DNNs) have shown promise in enhancing object detection and recognition accuracy.

Purpose of the Study:

• To develop and evaluate a deep-learning-based system for detecting diseases and pests in tomato plants using in-place captured images.
• To identify the most suitable deep learning architecture for this specific agricultural application.
• To improve detection accuracy and reduce false positives through advanced annotation and data augmentation techniques.

Main Methods:

• Investigated three deep learning meta-architectures: Faster Region-based Convolutional Neural Network (Faster R-CNN), Region-based Fully Convolutional Network (R-FCN), and Single Shot Multibox Detector (SSD).
• Combined meta-architectures with deep feature extractors, including VGG net and Residual Network (ResNet).
• Implemented local and global class annotation and data augmentation strategies for model training and validation on a large, diverse dataset.

Main Results:

• The developed system demonstrated effective recognition of nine distinct tomato diseases and pests.
• The approach proved capable of handling complex scenarios, including variations in infection status and location.
• Performance evaluation confirmed the system's ability to deal with images from various camera resolutions.

Conclusions:

• Deep learning architectures, particularly when combined with robust feature extractors and data augmentation, offer a powerful solution for automated plant disease and pest detection.
• The proposed system provides a viable tool for early detection, contributing to reduced economic losses in agriculture.
• Further research can explore optimizing these architectures for real-time field deployment and broader crop applications.