基于深度学习的道路语义分割新方法  

A novel method of roads’semantic segmentation based on deep learning

作　　者：薛振彦 惠明[1] 程晋炳 董昊 XUE Zhenyan;HUI Ming;CHENG Jinbing;DONG Hao(Henan International Joint Laboratory of MXene Materials Microstructure, School of Physics and Electronic Engineering, Nanyang Normal University, Nanyang 473061, China;School of Art and Art Design,Nanyang Normal University,Nanyang 473061,China)

机构地区：[1]南阳师范学院物理与电子工程学院,河南省MXene材料微结构国际联合实验室,河南南阳473061 [2]南阳师范学院美术与艺术设计学院,河南南阳473061

出　　处：《南阳师范学院学报》2021年第4期22-27,共6页Journal of Nanyang Normal University

基　　金：中国博士后基金面上项目(2016M591741);河南省高校科技创新人才项目(21HASTIT032)。

摘　　要：针对道路图像语义分割效果不够精确的问题,提出一种基于深度学习的道路图像语义分割的改进方法,并与传统的全卷积神经网络模型(FCN)相结合实现道路图像语义分割.先验概率层利用道路图像的先验知识,在所有道路图像训练标签的基础上构建二维数组表示像素点的分类概率,并将它结合传统全卷积神经网络模型对道路图像进行语义分割.实验结果表明:提出的先验概率层的后处理算法能够优化传统全卷积神经网络模型的分割效果,使像素精确度由88.8%提高到91.3%,平均像素精确度由82.9%提高到85.7%,平均交并比值由72.5%提高到77.9%.Aiming at the problem that the semantic segmentation of road image is not accurate enough,this paper proposes an improved semantic segmentation of road image based on deep learning.This method introduces a post-processing method of prior probability layer,which combines the traditional full convolution neural network(FCN)model to segment road images.Based on the prior knowledge of the training road images,the prior probability layer is a two-dimensional array which can represent the classification probability of each pixel and is combined with FCN for the semantic segmentation of road images.The experimental results show that the proposed post-processing algorithm of prior probability layer can improve the segmentation performance of traditional convolutional neural network model,and increase the pixel accuracy from 88.8%to 91.3%,the average pixel accuracy from 82.9%to 85.7%,and the average intersection ratio from 72.5%to 77.9%.

关 键 词：深度学习 语义分割 全卷积神经网络 道路图像 先验概率层的后处理算法 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]