摘要
针对复杂边界处语义分割信息在多尺度情况下表现出的不一致性,引起的现有模型产生错误语义分割标记预测问题,受到空间语境关系启发,提出一种多尺度条件随机场的深度卷积网络模型。首先,不同尺度的初始标记通过深度卷积网络获得,从而构建语义标记的多尺度表达;其次,引入多尺度表达的条件随机场模块,并在该模块上添加同层二元关系和异层二元关系,构建新的条件随机场能量评价;最后,针对模型中的深度网络参数和条件随机场参数学习过程,设计一种两阶段的训练过程,兼顾了模型的收敛速度和精度。实验给出了公认的PASCAL VOC 2012数据集的语义分割结果,能够说明该方法较现有主流方法有一定优势。
Complex boundaries can generate inconsistent semantic image segmentation at different scales,which can produce prediction problems of wrong semantic segmentation labels of the original model.Inspired by the relationship of the spatial context,a deep convolutional network model with multi-scale conditional random fields is proposed in the paper.Firstly,the initial labels of different scales are obtained by deep convolutional network to construct the multi-scale representation of semantic regions.Secondly,we propose a conditional random field module based on multi-scale representation,and we further design the same-layer pairwise constrains and cross-layer pairwise constrains in this module to construct a new conditional random field energy evaluation.Finally,our model have two parts of parameters,including deep network parameter and conditional random field parameter,and we give consideration to both the convergence speed and accuracy to design a two-stage training process for parameter learning.Experimental results show us recognized PASCAL VOC 2012 database of semantic image segmentation which proves that the method used in the paper exceeds the current mainstream approach.
作者
汪萍
WANG Ping(Anhui Vocational and Technical College of Press and Publication,Hefei 230601,China)
出处
《宿州学院学报》
2019年第7期69-74,共6页
Journal of Suzhou University
基金
安徽省高校人文社会科学研究重点项目(SK2017A0919)
安徽省质量工程项目(2017jyxm0829)
关键词
语义图像分割
多尺度CRF
跨层二元约束
深度卷积网络
空洞卷积
Semantic image segmentation
Multi-scale CRF
Cross-layer pairwise constraints
Deep convolutional network
Atrous convolution
