摘要
图像分割是机器视觉的关键步骤。本文针对目前经典图像分割方法、基于图论的图像分割方法与语义分割方法,尤其是卷积神经网络(Convolutional Neural Networks,CNN)应用于语义分割存在问题,系统比较编-解码器、多孔卷积等端到端语义分割架构的结构、体系演进、特点及趋势、技术水平,指出CNN端到端语义分割方法的并联模块设计、多孔卷积层多孔比例选取、其他多尺度特征提取算法均值得深入研究,分割方法与评价规则算法、辅助标注、边缘智能计算和并行调度计算结合是研究重要趋势。
Image segmentation is a key step in machine vision.The problems of applying CNN to semantic segmentation are point out,by expounding the status and principle among the classical image segmentation methods,graph theory-based image segmentation method and semantic segmentation method.The framework evolution,trends,and effectiveness of end-to-end semantic segmentation architecture,which also can divided into encoder-decoder architecture and atrous convolution architecture,are contrasted.A series of problem of CNN-based end-to-end semantic segmentation method are worth further study,such as atrous parameter selection,the parallel module of atrous convolution,and other multi-scale feature extraction algorithms.It is an important trend to combine semantic segmentation method with evaluation rule algorithm,auxiliary annotation,edge intelligent computing and parallel scheduling computing.
作者
黄坚
刘桂雄
HUANG Jian;LIU Guixiong(School of Mechanical and Automotive Engineering,South China University of Technology,Guangzhou 510640,China)
出处
《激光杂志》
北大核心
2019年第5期10-16,共7页
Laser Journal
基金
广州市产学研协同创新重大专项(2017010160641)
广东省现代几何与力学计量技术重点实验室开放课题(SCMKF201801)
关键词
语义分割
卷积神经网络
机器视觉检测
编解码器
多孔卷积
semantic segmentation
convolutional neural networks
machine vision inspection
encoder-decoder
atrous convolution
