摘要
图像语义分割是视觉智能方向最重要的基础性技术之一,语义分割效果关系着智能系统对其应用场景的理解能力,因此在诸如无人驾驶、机器人认知与导航、安防监控与无人机着陆系统等重要领域均具有较大的应用价值。由于复杂环境下的目标存在非结构化、目标多样化、形状不规则化以及光照变化、视角变化、尺度变化与物体遮挡等各种干扰因素,给图像的语义分割带来了较大挑战。近年来,受益于深度学习理论的快速发展,图像语义分割方向涌现了一大批具有典型意义的研究成果。为启发图像语义分割领域的学术研究及其相关智能系统的工程化开发,文中首先全面阐述了图像语义分割方法的研究发展历程,并将其划分为:传统的图像语义分割方法、传统方法与深度学习相结合的图像语义分割方法、基于深度学习的图像语义分割方法;其次从复杂环境下图像语义分割面临的问题出发,重点对近年来涌现的各种面向复杂环境的语义分割方法的模型、算法、性能及存在的问题进行了详细地分析与对比,并按照强监督、弱监督、无监督图像语义分割方法分类进行阐述;然后归纳了当前主流的PASCALVOC,Cityscape,SUNRGB-D等9类包含各种复杂环境的数据集,以及3项评估指标PA,mPA和mIoU;最后对面向复杂环境的图像语义分割研究工作进行了总结,并对其在实时视频分割、三维场景重构及无监督语义分割等方向的发展进行了展望。
Image semantic segmentation is one of the most important fundamental technologies for visual intelligence.Semantic segmentation can greatly enable intelligent systems to understand their surrounding scenarios,so it has enormous value in application domains such as unmanned vehicles, robot cognition and navigation,video surveillance and drone landing systems.Great challenges also exist in the semantic segmentation of images,due to various interfering factors of targets in complex environments,such as unstructured targets,diversity of objectives, irregular shapes,illumination changes,different viewing angles,scale variation,object occlusion,etc.In recent years,benefiting from the great advancements in deep learning techniques,a large number of research approaches with practical significance emerge in ima- ge semantic segmentation.For having a comprehensive survey and inspiring the academic research,this paper extensively discussed the existing state-of-the-art image semantic segmentation methods,and further classified them into the traditional image semantic segmentation ones,the ones combining traditional and deep learning techniques,and those based purely on deep learning.In order to address these problems in complex environments, various semantic segmentation methods for complex environment emerged in recent years were analyzed and compared in detail,including the mo- dels ,algorithms and performance with the category of strong supervised,weak supervised and unsupervised semantic segmentation methods.Furthermore, the current main datasets such as PASCAL VOC,Cityscape,SUN RGB-D,which contains various complex environments and 3 evaluation indicators of PA, mPA,mIoU were summarized.Finally,the existing research of image semantic segmentation for complex environment was summarized,and its future trends were prospected such as optimization in real-time video,3d scene reconstruction and unsupervised semantic segmentation techniques.
作者
王嫣然
陈清亮
吴俊君
WANG Yan-ran;CHEN Qing-liang;WU Jun-jun(College of Information Science and Technology,Jinan University,Guangzhou 510632,China;School of Mechatronics Engineering,Foshan University,Foshan,Guangdong528225,China)
出处
《计算机科学》
CSCD
北大核心
2019年第9期36-46,共11页
Computer Science
基金
国家自然科学基金(61603103,61673125)
广东省自然科学基金(2016A030310293)
广州市科技计划科学研究专项(201707010013)资助
关键词
语义分割
视觉智能
深度学习
图像分割
卷积神经网络
Semantic segmentation
Visual intelligence
Deep learning
Image segmentation
Convolutional neural network
