基于深度学习的语义分割算法综述  

A Review of Semantic Segmentation Algorithm Based on Deep Learning

作　　者：赵霞[1] 白雨 倪颖婷 陈萌 郭松[3] 杨明川 陈凤 ZHAO Xia;BAI Yu;NI Yingting;CHEN Meng;GUO Song;YANG Mingchuan;CHEN Feng(Department of Electronics and Information Engineering,Tongji University,Shanghai 201804,China;Research Institute of Shanghai Aerospace System Engineering,Shanghai 200092,China;Shanghai Academy of Spaceflight Technology,Shanghai 201109,China)

机构地区：[1]同济大学电子与信息工程学院,上海200092 [2]上海宇航系统工程研究所,上海201109 [3]上海航天技术研究院,上海201109

出　　处：《上海航天》2019年第5期71-82,共12页Aerospace Shanghai

基　　金：上海航天科技创新基金(SAST2016018)

摘　　要：图像的语义分割是对图像中的每个像素标注其所属的类别。在航天领域,语义分割技术可用于定位航天器及其零部件,为航天器故障排除、部件维修、太空垃圾清理等在轨服务创造条件。近几年,全部或部分使用深度学习时,语义分割的效果获得了很大的提升。本文对基于深度学习的语义分割算法进行综述。首先介绍常用的数据集和通用的深度神经网络,随后对两类具有重大实用意义的分割算法:编码器-解码器算法和整合上下文信息算法进行总结。最后对语义分割的发展进行了展望。Semantic segmentation is the classification of each pixel in the image.In aerospace field,semantic segmentation can be used to locate spacecraft and its components,and to create conditions for spacecraft troubleshooting,component maintenance,space junk cleaning and other on-orbit services.In recent years,the effect of semantic segmentation has been greatly improved with full or partial use of deep learning.In this paper,the semantic segmentation algorithm based on deep learning is reviewed.Firstly,the commonly used datasets and deep neural networks are introduced,then two kinds of segmentation algorithms with great practical significance are summarized:encoder-decoder method and integrating context knowledge method.Finally,the development of semantic segmentation is prospected.

关 键 词：深度学习 语义分割 全卷积网路 编码器-解码器算法 整合上下文信息算法 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]