融合深度神经网络和空洞卷积的语义图像分割研究  

Semantic Image Segmentation Based on Fusion of Deep Neural Networks and Atrous Convolution

作　　者：陈洪云 孙作雷[1] 孔薇[1] CHEN Hong-yun;SUN Zuo-lei;KONG Wei(College of Information Engineering,Shanghai Maritime University,Shanghai 201306,China)

机构地区：[1]上海海事大学信息工程学院

出　　处：《小型微型计算机系统》2020年第1期166-170,共5页Journal of Chinese Computer Systems

基　　金：上海市科委自然科学基金项目(18ZR1417200)资助

摘　　要：语义分割是计算机视觉中的基本任务,是对图像中的不同目标进行像素级的分割与分类.针对多尺度的目标分割难题,本文提出了一种基于Res Net网络的方法,通过定义并联支路,将浅层特征图像信息融合到深层特征图像中,提出新的空洞空间金字塔模块,该模块采用并行的不同采样率的空洞卷积进行特征提取与融合,从而更有效的提取不同层的特征以及上下文信息,并且在新模块中加入批规范化计算,增强参数调优的稳定性.本文还采用了Adam自适应优化函数,在训练的过程中,使得每个参数的更新都具有独立性,提升了模型训练的稳定性.本文结果在PASCAL VOC 2012语义分割测试集中取得了77.31%mIOU的成果,优于Deeplab V3的效果.Semantic segmentation is a basic task in computer vision,it’s a pixel-level segmentation and classification of different objects in an images.Aiming at the problem of multi-scale target segmentation,this paper proposes a method based on Res Net network,this method fuses the shallowfeature image information into the deep feature image by defining a parallel branch.A newartous space pyramid module is proposed,The module uses parallel convolution of different sampling rates for feature extraction and fusion,which can extract features and context information of different layers more effectively,and batch normalization calculation is added in the newmodule to enhance the stability of parameter tuning during model training.This paper also adopts the Adam adaptive optimization function,in order to improve the stability of the training,because in the process of training,the update of each parameter is independent.The results of this paper achieved 77.31%mIOU results in the PASCAL VOC 2012 semantic segmentation test set,better than Deeplab V3.

关 键 词：语义分割 神经网络 空洞卷积 空洞空间金字塔模块 

分 类 号：TP389[自动化与计算机技术—计算机系统结构]