Learning lightweight tea detector with reconstructed feature and dual distillation.

Area of Science:

• Computer Vision
• Machine Learning
• Agricultural Technology

Background:

• Deep neural networks excel in image recognition, including tea leaf detection, but large models are computationally intensive for real-world deployment.
• Existing lightweight models often sacrifice accuracy for efficiency, hindering real-time applications like automated tea picking in resource-constrained areas.

Purpose of the Study:

• To enhance the detection accuracy of lightweight deep learning models for tea leaves, addressing computational limitations in remote environments.
• To propose a novel method, Reconstructed Feature and Dual Distillation (RFDD), that improves the performance of existing lightweight detectors for tea leaf recognition.

Main Methods:

• Reconstructed Feature selectively masks student model features using teacher model attention maps, guided by a generation block to transfer knowledge.
• Dual Distillation, comprising Decoupled Distillation (foreground/background focus) and Global Distillation (recovering inter-pixel relationships), refines feature learning.
• RFDD is designed for easy integration with various detection frameworks, requiring only feature map loss calculation.

Main Results:

• Experiments on a tea dataset demonstrated performance improvements across different detector frameworks (e.g., RetinaNet, Faster RCNN) after RFDD guidance.
• A one-stage detector (RetinaNet) saw a 3.14% increase in Average Precision (AP), while a two-stage detector (Faster RCNN) achieved a 3.53% AP improvement.
• The proposed RFDD method effectively boosts the detection capabilities of lightweight models for tea leaves.

Conclusions:

• RFDD offers a practical solution for real-time tea leaf detection by enhancing lightweight models without prohibitive computational costs.
• The method shows significant promise for improving automated tea-picking technologies, particularly in challenging, remote geographical locations.
• RFDD's feature-based approach ensures broad applicability across diverse deep learning detection architectures.