Enhancing weed detection through knowledge distillation and attention mechanism.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Weeds compete with crops, reducing agricultural yields.
• Vision Transformers (ViT) show promise for weed detection but face deployment challenges in resource-limited agricultural robotics due to high training costs and model size.
• Traditional Convolutional Neural Networks (CNNs) are less effective than ViTs.

Purpose of the Study:

• To optimize Vision Transformer (ViT) models for weed detection in agricultural robotics.
• To address the computational constraints of ViTs, including model size and memory requirements.
• To maintain high weed detection performance while reducing computational expense.

Main Methods:

• Proposed a knowledge distillation method to optimize ViT models.
• Used ResNet-50 as a teacher model to distill knowledge into a compacted ViT student model.
• Facilitated parameter sharing and local receptive fields for the student model.

Main Results:

• The student ViT model achieved a mean Average Precision (mAP) of 83.47% for weed detection.
• The optimized model has only 5.7 million parameters, indicating minimal computational expense.
• Significant improvements in the student model's performance and efficiency were observed.

Conclusions:

• Knowledge distillation effectively optimizes ViT models for weed detection.
• The proposed method successfully addresses computational limitations for deploying ViTs in agricultural robotics.
• The optimized ViT model offers a viable solution for accurate and efficient weed detection in agriculture.