PKD-YOLOv8: A Collaborative Pruning and Knowledge Distillation Framework for Lightweight Rapeseed Pest Detection.

Area of Science:

• Agricultural Science
• Computer Vision
• Artificial Intelligence

Background:

• Rapeseed is a vital global crop facing significant threats from pests.
• Effective pest management is crucial for maintaining rapeseed production and economic value.
• Existing detection methods may lack efficiency or real-time capabilities for field application.

Purpose of the Study:

• To develop a lightweight and efficient deep learning model for rapeseed pest detection.
• To compress a YOLOv8s model for improved performance on edge devices.
• To enhance the model's feature representation capabilities through a novel distillation strategy.

Main Methods:

• Utilized structured pruning based on model analysis and sensitivity evaluation.
• Proposed a novel lightweight model (LMGD) distillation strategy integrating Logit and MGD distillation.
• Developed a dedicated rapeseed pest dataset (ACEFP) for experimental validation.

Main Results:

• Achieved 96.7% mAP@0.5, 93.2% accuracy, and 92.7% recall on the ACEFP dataset.
• Compressed model size from 11.2 MB to 4.4 MB (60.7% reduction) and FLOPs from 28.3 G to 10.01 G (64.6% reduction).
• Maintained high accuracy with only a 0.1% reduction and achieved 11.76 FPS on a Jetson Nano for real-time inference.

Conclusions:

• The proposed lightweight collaborative compression learning method effectively reduces model size and computational load.
• The LMGD distillation strategy enhances the student model's performance without significant accuracy loss.
• The optimized model demonstrates excellent real-time pest detection capabilities suitable for edge deployment in agriculture.