耦合轻量级网络GhostNet的无人机树冠检测研究

Research on UAV Canopy Detection Coupled with Lightweight Network GhostNet

在高分辨率遥感影像树冠检测研究中,基于深度学习方法虽取得较大进展,但基于无人机航拍影像检测效果欠佳,并且检测速度较慢,实时性较差。针对此类问题,旨在无人机获取树冠影像,从3个方面对YOLOv4网络模型进行改进,提高检测速度及精度。首先,使用多种数据增强方法,扩增实验样本,提高模型鲁棒性;其次,对YOLOv4网络模型进行改进,使用轻量级网络GhostNet作为改进算法的主干特征提取网络,减少模型参数,使得模型轻量化;最后,使用深度可分离卷积替代部分普通卷积,在保证提取特征同等情况下,缩短提取时间。实验结果表明,基于改进的YOLOv4模型检测精度高达98%,单张检测速度达到0.166 ms,能够有效节约时间成本和人工成本。

Although great progress has been made in the research of canopy detection based on high-resolution remote sensing images,the detection effect based on UAV aerial images is not good,and the detection speed is slow and the real-time performance is poor.In order to solve such problems,this paper aims to obtain tree canopy images from UAVs,improve the YOLOv4 network model from three aspects,and improve the detection speed and accuracy.Firstly,a variety of data augmentation methods are used to amplify experimental samples and improve the robustness of the model.Secondly,the YOLOv4 network model is improved,and the lightweight network GhostNet is used as the backbone feature extraction network of the improved algorithm,which reduces the model parameters and makes the model lightweight.Finally,the depth separable convolution is used to replace part of the ordinary convolution,which shortens the extraction time under the condition that the extracted features are equally extracted.Experimental results show that the detection accuracy of the improved YOLOv4 model is as high as 98%,and the detection speed of a single sheet reaches 0.166 ms,which can effectively save time and labor costs.