FEGW-YOLO: A Feature-Complexity-Guided Lightweight Framework for Real-Time Multi-Crop Detection with Advanced Sensing...

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Real-time object detection on edge devices is crucial for precision agriculture and autonomous systems.
• Integrating multi-modal sensors (RGB-D, thermal, hyperspectral) presents significant computational challenges for resource-constrained hardware.

Purpose of the Study:

• To introduce FEGW-YOLO, a lightweight object detection framework designed to balance efficiency and accuracy for edge devices.
• To enable fine-grained visual perception and multi-modal sensor compatibility in precision agriculture applications.

Main Methods:

• Developed a Feature Complexity Descriptor (FCD) for adaptive, layer-wise network compression.
• Integrated Feature Engineering-driven Ghost Convolution (FEG-Conv) for parameter reduction.
• Employed Efficient Multi-Scale Attention (EMA) to counteract compression-induced information loss and Wise-IoU loss for improved localization.

Main Results:

• FEGW-YOLO achieved 95.1% mAP@0.5, reducing model parameters by 54.7% and GFLOPs by 53.5% compared to a YOLO-Agri baseline.
• Real-time inference achieved 38 FPS on NVIDIA Jetson Xavier with low power consumption (12.3 W).
• Field deployment demonstrated an 87.3% harvesting success rate with a 2.1% fruit damage rate.

Conclusions:

• FEGW-YOLO advances efficient agricultural sensing through metric-guided compression and multi-modal integration.
• The framework is validated for practical edge deployment in autonomous harvesting and precision monitoring.
• FEGW-YOLO bridges the gap between research and real-world application for resource-constrained agricultural systems.