基于深度学习的无人机地物图像分割方法  

Unmanned Aerial Vehicle Ground Object Image SegmentationMethod Based on Deep Learning

作　　者：陈国军[1] 尹冲 滕一诺 王雯璇 CHEN Guojun;YIN Chong;TENG Yinuo;WANG Wenxuan(College of Computer Science and Technology,China University of Petroleum(East China),Qingdao 266580)

机构地区：[1]中国石油大学(华东)计算机科学与技术学院,青岛266580

出　　处：《计算机与数字工程》2023年第3期706-711,共6页Computer & Digital Engineering

基　　金：山西省交通建设科技项目(编号:20190568)资助。

摘　　要：针对无人机地物图像的传统分割方法工程量大,效率低下,深度学习的无人机地物图像分割算法在复杂场景下精度不高和数据集的类别不均衡(长尾数据)等问题,提出一种基于深度学习的高分辨率无人机地物图像分割方法,用于提高不同地貌区域的分割精度。在语义分割模型DeepLabv3的基础上进行改进,将原始主干网络ResNet101替换为ResNet152并添加预训练模型,调整扩张卷积空间金字塔池化模块的扩张率,采用类别平衡损失函数来解决长尾数据问题。在采集的无人机地物图像数据集上进行训练并通过测试集的分割效果证明模型改进方法的有效性。根据实验模型分割效果表明,改进后的方法在测试集上平均交并比达到70.8%,相比原始模型提升了27.2%,能够得到效果更好的分割结果。Traditional segmentation methods for UAV ground object images have a large amount of engineering and low effi⁃ciency.Deep learning UAV ground object image segmentation algorithms have low accuracy in complex scenes and unbalanced data sets(long-tail data).A high-resolution UAV feature image segmentation method based on deep learning is proposed to improve the segmentation accuracy of different geomorphic regions.Improving on the basis of the semantic segmentation model DeepLabv3,re⁃placing the original backbone network ResNet101 with ResNet152 and adding a pre-training model,adjusting the expansion vol⁃ume The expansion rate of the product space pyramid pooling module uses the category balance loss function to solve long-tail data problem.Finally,train on the collected UAV ground object image data set and prove the effectiveness of the model improvement method through the segmentation effect of the test set.According to the segmentation effect of the experimental model,the improved method has an average intersection ratio of 70.8%on the test set,which is 27.2%higher than the original model,and can obtain bet⁃ter segmentation results.

关 键 词：深度学习 语义分割 长尾数据 类别平衡损失函数 无人机地物图像 

分 类 号：TP39[自动化与计算机技术—计算机应用技术] TP75[自动化与计算机技术—计算机科学与技术]