PLFYNet-based edge-deployable detection system for Ginkgo biloba leaf diseases.

Area of Science:

• Agricultural Science
• Computer Science
• Deep Learning

Background:

• Precision agriculture demands accurate monitoring of Ginkgo biloba diseases.
• Existing detection systems face challenges in complex environments, including low accuracy, efficiency, and high computational costs for edge devices.

Purpose of the Study:

• To develop a lightweight deep learning model for real-time disease detection on resource-constrained embedded devices.
• To address the limitations of current disease detection systems in agricultural settings.

Main Methods:

• A multi-class dataset of 7,158 augmented images was created for three disease categories.
• An optimized lightweight architecture (LCNet-FusionYOLO) was developed, incorporating attention mechanisms, an improved detection head, efficient convolutions, and a custom feature fusion module.
• Layer-Adaptive Magnitude-based Pruning (LAMP) was applied to create the final PLFYNet model.

Main Results:

• The PLFYNet model achieved 94.5% mAP@0.5 with only 3.0M parameters, outperforming YOLOv7-tiny by 4.8% with half the parameters.
• Real-time inference was demonstrated at 50.5 FPS on a Jetson Orin Nano embedded platform.

Conclusions:

• PLFYNet offers a high-precision, computationally efficient solution for disease detection in Ginkgo biloba cultivation.
• The model provides a practical edge-based monitoring system, overcoming deployment challenges in complex agricultural environments and supporting sustainable practices.