An improved ShuffleNetV2 method based on ensemble self-distillation for tomato leaf diseases recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Agricultural Science

Background:

• Accurate tomato disease recognition is vital for crop yield.
• Large deep learning models are accurate but difficult to deploy on edge devices.
• Lightweight models are needed for real-time disease detection on edge devices.

Purpose of the Study:

• To develop an efficient and accurate tomato disease recognition method for edge devices.
• To improve the performance of the lightweight ShuffleNetV2 model.
• To enable timely disease detection for enhanced tomato production.

Main Methods:

• An ensemble self-distillation method was proposed and applied to ShuffleNetV2.
• Multiple shallow models were integrated with ShuffleNetV2 in a distillation framework.
• A depthwise separable convolution layer was used to extract fused feature information.

Main Results:

• The optimized ShuffleNetV2 achieved 95.08% accuracy, surpassing larger models like VGG16 and ResNet18.
• The model demonstrated the highest recognition accuracy among lightweight models with the smallest parameter count.
• The ensemble method significantly enhanced ShuffleNetV2's performance without altering its original structure.

Conclusions:

• The optimized ShuffleNetV2 is suitable for real-time tomato disease detection on edge devices.
• The ensemble self-distillation method effectively improves lightweight model performance.
• The approach offers flexibility for model deployment in agricultural applications.