基于改进U^(2)Net网络和无人机影像的城市绿化识别方法

Urban greening recognition method based on improved U^(2)Net andunmanned aerial vehicle images

针对城市绿化识别中存在的无可用公开数据集、人工标注数据任务大、图像边界分割不精确的问题,提出结合无人机影像和深度学习网络的城市绿化自动识别框架.首先建立基于无人机影像的城市绿化数据集,其次,运用交互式自动标注工具EISeg对数据进行标注,引入U^(2)Net用于无人机影像下的城市绿化识别,最后,在网络的特征提取模块引入最大池化索引来加强对目标边界的分割能力.结果表明,相较于其它深度学习网络,U^(2)Net在小规模数据集中有着优异的表现且改进后的网络在1 000张的训练集中达到了97.16%的分类精度,较原始的U^(2)Net提高0.68%,模型参数量、计算量、内存均未显著提升.改进后的方法提升了城市绿化的识别精度,可以为城市绿化识别提供一种新的自动解译方法.

An automatic urban greening identification framework that combines UVA images and deep learning networks was proposed for the problems of no available public dataset,large manual labeling data task,and imprecise image boundary segmentation in urban greening identification.Firstly,an urban greening dataset based on UAV images was established.Then,the interactive automatic labeling tool EISeg was used to label the data,and introduction of U^(2)Net for urban greening identification under UAV images.At last,the max pooling indices was introduced in the feature extraction module of the network to strengthen the segmentation ability of the target boundary.The experimental results show that U^(2)Net has excellent performance in small-scale datasets and the improved U^(2)Net network achieved overall accuracy of 97.16%in a small datasets of 1000,which is 0.68%better than the original U^(2)Net,and the number of model parameters,computation,and memory has not been significantly improved.The improved U^(2)Net network combined with UAV images can better recognize the green coverage area in the city and provide a new automatic decoding method for urban green recognition.