Lightweight and efficient deep learning models for fruit detection in orchards.

Area of Science:

• Computer Vision
• Agricultural Robotics
• Machine Learning

Background:

• Accurate apple recognition is crucial for automated orchard equipment.
• Challenges include dense targets, occlusion, and natural environmental variations.
• Existing methods may lack efficiency or robustness in complex scenarios.

Purpose of the Study:

• To develop a real-time, lightweight object detection network for apple recognition in orchards.
• To address challenges posed by dense targets, occlusion, and natural environmental factors.
• To improve the accuracy and efficiency of automated fruit picking systems.

Main Methods:

• Construction of a diverse fruit dataset encompassing various orchard scenarios.
• Proposal of the Efficient Lightweight object Detector (ELD) network.
• Integration of novel modules: Efficient Ghost-shuffle Slim (EGSS) and Mix channel Attention (MCAttention).
• Implementation of SlimPAN for network compression and Shape-IOU loss for robustness.
• Application of knowledge distillation for enhanced accuracy and lightweight design.

Main Results:

• The ELD network achieved an accuracy of 87.4% for fruit detection.
• Demonstrated a low parameter count and computational load (1.7 GLOPs).
• Achieved a high frames per second (FPS) rate of 156.
• Outperformed existing networks in terms of accuracy and resource consumption.

Conclusions:

• The ELD network offers a highly accurate and efficient solution for apple detection in complex orchard environments.
• Its lightweight design and robust performance make it suitable for real-time automated picking applications.
• The proposed attention mechanisms and network optimization strategies significantly enhance detection capabilities.