多孔全卷积网络的语义分割算法研究  

Fully convolutional networks with atrous convolution for semantic segmentation

作　　者：戴伟达 霍智勇[2] 严邓涛 陈钊 DAI Weida;HUO Zhiyong;YAN Dengtao;CHEN Zhao(College of Telecommunications ＆ Information Engineering,Nanjing University of Posts and Telecommunications,Nanjing 210003,China;Jiangsu Province Key Lab on Image Proessing ＆ Image Communication,Nanjing University of Posts and Telecommunications,Nanjing 210003,China)

机构地区：[1]南京邮电大学通信与信息工程学院,江苏南京210003 [2]南京邮电大学江苏省图像处理与图像通信重点实验室,江苏南京210003

出　　处：《南京邮电大学学报（自然科学版）》2018年第4期96-102,共7页Journal of Nanjing University of Posts and Telecommunications：Natural Science Edition

基　　金：国家自然科学基金(61471201);江苏省高校自然科学研究重点项目(13KJA510004);江苏省自然科学基金青年基金(BK20130867);江苏省"六大人才高峰"(2014-DZXX-008);南京邮电大学"1311人才计划"资助项目

摘　　要：全卷积网络(Fully Convolutional Network,FCN)中由于连续的最大池化和下采样操作造成特征分辨率急剧降低,使最终上采样恢复的特征图失去对图像的细节敏感性。文中采用多孔卷积替代全卷积网络中的标准卷积,从而使得卷积网络在计算特征响应时能精确控制图像的分辨率,同时在不增加参数数量以及计算量的前提下,有效地扩大了滤波器的感受野。数值实验表明,文中算法利用更多的上下文信息获取了更稠密的特征,有效地提高了分割精度。In the fully convolutional network（FCN）,the feature resolution is drastically reduced due to the successive max-pooling and down-sampling operations,so that the feature map finally recovered by the up-sampling loses the detail sensitivity to the image. The atrous convolution is used to replace the standard convolution in the fully convolutional network,so that the convolution network can precisely control the resolution of the image when calculating the characteristic response. Meanwhile,the receptive field of the filter is effectively expanded without increasing the number of parameters or the amount of computation. Numerical experiments show that the algorithm uses more context information to obtain more dense features,thus improving the segmentation accuracy.

关 键 词：卷积神经网络 语义分割 最大池化 感受野 多孔卷积 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]