Knowledge Distillation Facilitates the Lightweight and Efficient Plant Diseases Detection Model.

Area of Science:

• Agricultural technology
• Computer vision
• Artificial intelligence

Background:

• Timely plant disease diagnosis is crucial for preventing crop loss and ensuring food security.
• Current object detection methods for plant diseases are accurate but limited to single crops and often require large computational resources.
• Deploying disease diagnosis models on mobile agricultural devices is challenging due to model size and parameter count, often leading to reduced accuracy when models are compressed.

Purpose of the Study:

• To develop a lightweight and efficient plant disease detection method for diagnosing multiple diseases across various crops.
• To address the limitations of existing models in terms of single-crop specificity and deployability on mobile agricultural devices.
• To maintain high diagnostic accuracy while significantly reducing model parameters.

Main Methods:

• Proposed a plant disease detection method utilizing multistage knowledge distillation.
• Designed four lightweight student models (YOLOR-Light-v1, YOLOR-Light-v2, Mobile-YOLOR-v1, Mobile-YOLOR-v2) using two distinct strategies.
• Employed the YOLOR model as the teacher model for knowledge distillation.

Main Results:

• Achieved 60.4% mAP@0.5 on the PlantDoc dataset with significantly reduced model parameters.
• The proposed multistage knowledge distillation method improved the performance of lightweight models.
• Outperformed existing methods in terms of accuracy and efficiency for multi-crop, multi-disease diagnosis.

Conclusions:

• Multistage knowledge distillation effectively creates lightweight yet accurate models for plant disease detection.
• The developed technique is suitable for deployment on agricultural mobile devices, supporting smart agriculture.
• The methodology shows potential for extension to other computer vision tasks like image classification and segmentation.