基于PN-YOLO v8s-Pruned的轻量化三七收获目标检测方法

Lightweight Object Detection Method for Panax notoginseng Based on PN-YOLO v8s-Pruned

摘要:

为实现三七联合收获作业过程中的自适应分级输送和收获状态实时监测，本文针对三七根土复合体特征和复杂田间收获工况，提出一种基于YOLO v8s并适用于Jetson Nano端部署的三七目标检测方法。在YOLO v8s对三七准确识别的基础上，针对其新的模型结构特性，利用通道剪枝算法，制定相应剪枝策略，保证模型精度的同时提升实时检测性能。采用TensorRT推理加速框架将改进模型部署至Jetson Nano，实现了三七目标检测模型的灵活部署。试验结果表明，改进后的PN-YOLO v8s-Pruned模型在主机端的平均精度均值为93.71%，参数量、计算量、模型内存占用量分别为原始模型的39.75%、57.69%、40.25%，检测速度提升44.26%，与其他目标检测模型相比，本文改进模型在计算复杂度、检测精度和实时性方面具有更好的综合检测性能。在Jetson Nano端部署后，改进模型检测速度达18.9f/s，较加速前提升2.7倍，较原始模型提升5.8f/s。台架试验结果表明，4种输送分离收获作业工况下三七目标检测的平均精度均值达87%以上，不同输送分离收获作业工况和不同流量等级下的目标三七计数平均正确率分别达92.61%、91.76%。田间试验结果表明，三七目标检测平均精度均值达84%，计数平均正确率达88.11%，图像推理速度达31.0f/s。模型检测性能和计数效果能够满足复杂田间收获工况下目标三七的检测需求，可为基于边缘计算设备的三七联合收获作业自适应分级输送系统和收获作业质量监测系统提供技术支撑。

Abstract:

In order to realize the adaptive grading conveyance and real-time monitoring of harvesting status in the process of Panax notoginseng combined harvesting operation, focusing on the characteristics of Panax notoginseng root-soil complex and the complex field harvesting conditions, a Panax notoginseng object detection method based on YOLO v8s and suitable for deployment on the Jetson Nano was proposed. Based on the accurate recognition of Panax notoginseng by YOLO v8s, the channel pruning algorithm was utilized to formulate a corresponding pruning strategies for its new model structural characteristics, which ensured the accuracy and improved the real-time detection performance at the same time. The improved model was deployed to Jetson Nano by using the TensorRT inference acceleration framework, which realized the flexible deployment of the Panax notoginseng object detection model. The experimental results showed that the mean average precision of the improved PN-YOLO v8s-Pruned model on the host side was 93.71%, although it was decreased by 0.94 percentage points compared with that of the original model, the number of parameters, computational complexity, and model size were 39.75%, 57.69%, and 40.25% of the original model, respectively, and the detection speed was increased by 44.26%. Compared with other models, the improved model demonstrated superior comprehensive detection performance in terms of computational complexity, detection accuracy, and real-time performance. After deployment at the Jetson Nano, the improved model had a detection speed of 18.9 frames per second, which was 2.7 times higher than before acceleration and 5.8 frames per second higher than the original model, and the deployment detection effect was better than the original model. The results of the bench tests showed that the mean average precision of Panax notoginseng detection was more than 87% under four conveyor separation harvesting conditions. The average accuracy of the Panax notoginseng counting under different conveyor separation harvesting conditions and different flow levels reached 92.61% and 91.76%, respectively. The field test results showed that the mean average precision of Panax notoginseng detection was more than 84%, and the average accuracy of the Panax notoginseng counting reached 88.11%, which could meet the detection requirements of Panax notoginseng under complex field harvesting conditions, and could provide technical support for the monitoring system of harvesting quality and the adaptive grading transportation system of combined harvesting operation based on edge computing equipments.