Improving computer vision for plant pathology through advanced training techniques.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Convolutional neural networks (CNNs) are crucial for disease detection in cocoa (Theobroma cacao).
• Recent advancements in CNN accuracy for image classification have stagnated.
• Improving model generalizability and robustness is essential for real-world agricultural disease management.

Purpose of the Study:

• To investigate advanced training techniques for enhancing CNN performance in cocoa disease detection.
• To address the stagnation in accuracy improvements in computer vision for image classification.
• To develop more robust and generalizable deep learning models for agricultural applications.

Main Methods:

• Employed semi-supervised learning to reduce overfitting and enhance generalizability.
• Introduced a non-cocoa class to expose models to diverse features, improving robustness.
• Developed and utilized dynamic focal loss, a novel loss function that weights images based on empirical difficulty.
• Used Grad-CAM for qualitative assessment of model behavior.

Main Results:

• Semi-supervised learning significantly improved performance on subtle disease symptoms.
• The inclusion of a non-cocoa class enhanced model robustness in challenging cases.
• Dynamic focal loss provided superior handling of difficult images.
• ResNet18 combined with semi-supervised learning and dynamic focal loss demonstrated the strongest performance for practical deployment.

Conclusions:

• Semi-supervised learning and advanced loss functions hold significant potential for improving deep learning in agricultural disease management.
• The study introduces a new, high-quality benchmark dataset of 7220 images for cocoa disease detection, presenting a more realistic challenge.
• The developed methods offer a pathway to more effective and reliable automated disease detection systems in agriculture.