文献类型：期刊文献

中文题名：基于深度学习的无人车夜视图像语义分割

英文题名：Semantic segmentation of night vision images for unmanned vehicles based on deep learning

作者：高凯珺[1,2];孙韶媛[1,2];姚广顺[1,2];赵海涛[3]

机构：[1]东华大学信息科学与技术学院,上海201620;[2]东华大学数字化纺织服装技术教育部工程研究中心,上海201620;[3]华东理工大学信息科学与工程学院,上海200237

年份：2017

卷号：38

期号：3

起止页码：421

中文期刊名：应用光学

外文期刊名：Journal of Applied Optics

收录：CSTPCD;;Scopus;北大核心:【北大核心2014】；CSCD:【CSCD2017_2018】；

基金：国家自然科学基金项目(61375007);上海市科委基础研究项目(15JC1400600)

语种：中文

中文关键词：夜视图像;语义分割;深度学习;反卷积;无人车

外文关键词：night vision image; semantic segmentation; deep learning; deconvolution; unmannedvehicle

摘要：为了增强无人车对夜视图像的场景理解,在夜间模式下更快更精确地探测和识别周围环境,将深度学习应用于夜视图像的场景语义分割,提出了一种基于卷积-反卷积神经网络的无人车夜视图像语义分割方法。在传统的卷积神经网络中加入反卷积网络,构建卷积-反卷积神经网络,无需手工选取特征。通过像素到像素的学习和训练,得到图像语义分割模型,可直接用该模型预测夜视图像中每个像素所属的场景语义类别,实现无人车夜间行驶时的环境感知。实验结果表明,该方法具有较好的准确性和实时性,平均IU达到68.47。
In order to assist unmanned vehicles in understanding scene of night vision images, de- tecting and identifying surrounding environment more quickly and accurately at night, a seman- tic segmentation method of unmmanned vehicle night vision images based on convolution-decon- volution neural network is proposed, which uses deep learning to segment scenery semant of night vision images. Convolution-deconvolution neural network is constructed by adding decon- volution network to traditional convolutional neural network, without selecting feature manual- ly. By learning and training tained. The model can be use P d ixels-to-pixels, image semantic segmentation model can be ob- to predict scene semantic category of each image, realizing environment perception of unmanned vehicles at night, wh pix ich el in night vision is import for au- tomatic driving at night. Experimental results show that this method has good accuracy and real- time performance, and average IU reaches 68.47.