Surface Defect Detection System for Carrot Combine Harvest Based on Multi-Stage Knowledge Distillation.

Area of Science:

• Agricultural Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Carrot quality and safety are enhanced by surface defect detection before market entry.
• Real-time defect detection during combine harvesting is crucial for efficiency.

Purpose of the Study:

• To develop an efficient lightweight network for carrot surface defect detection during combine harvesting.
• To adapt the network to image blur caused by harvester vibrations using knowledge distillation.

Main Methods:

• Proposed an improved knowledge distillation network (mobile-slimv5s) using YOLOv5s as the teacher and a pruned MobileNetV2 as the student.
• Trained the networks on standard and motion-blurred datasets to handle vibration-induced blur.
• Utilized multi-stage feature connections with weighted guidance for knowledge transfer.

Main Results:

• The optimal mobile-slimv5s network achieved a model size of 5.37 MB.
• Achieved 90.7% accuracy with specific training parameters (LR=0.0001, BS=64, Dropout=0.65).
• Demonstrated superior performance compared to other algorithms.

Conclusions:

• The mobile-slimv5s network enables simultaneous carrot harvesting and defect detection in field environments.
• This research provides a foundation for applying knowledge distillation in smart agriculture for crop sorting.
• The study significantly improves in-field crop sorting accuracy and contributes to agricultural automation.