Abstract
INTRODUCTION: Target detection is a pivotal technology for precise monitoring of leaf-used Ginkgo biloba diseases in precision agriculture. However, complex plantation environments impose significant constraints on existing detection systems, manifesting as degraded detection accuracy, suboptimal efficiency, and prohibitive computational overhead for edge deployment. This study aims to develop a lightweight deep learning model tailored for real-time disease detection on resource-constrained embedded devices. METHODS: First, a comprehensive multi-class dataset was constructed, containing 7,158 augmented images covering three disease categories: chlorosis, insect pest, and physical damage. Five lightweight architectures were systematically evaluated, and an optimized reconstructed backbone network was adopted. To maintain architectural efficiency, attention mechanisms, an improved detection head, and efficient convolution techniques were integrated, along with a custom feature fusion module designed to address small target feature loss-forming the base model LCNet-FusionYOLO. Subsequently, Layer-Adaptive Magnitude-based Pruning (LAMP) was applied to reduce model scale while enhancing performance, yielding the final PLFYNet model. RESULTS: The PLFYNet model achieves 94.5% mAP@0.5 with only 3.0M parameters, surpassing the baseline YOLOv7-tiny by 4.8% while using merely half the parameters. Deployment on the Jetson Orin Nano embedded platform demonstrates real-time inference at 50.5 FPS, validating its practical applicability in field scenarios. DISCUSSION: This work establishes a paradigm for developing high-precision, computationally efficient disease detection systems. By balancing accuracy and resource efficiency, PLFYNet provides a practical edge-based monitoring solution for sustainable Ginkgo biloba cultivation, addressing the key deployment challenges of existing detection systems in complex agricultural environments.