结合全局注意力机制的实时语义分割网络  

Global attention mechanism with real time semantic segmentation network

作　　者：李涛[1,2] 高志刚 管晟媛 徐久成 马媛媛 LI Tao;GAO Zhigang;GUAN Shengyuan;XU Jiucheng;MA Yuanyuan(College of Computer and Information Engineering,Henan Normal University,Henan Xinxiang 453007,China;Engineering Lab of Henan Province for Intelligence Business&Internet of Things,Henan Xinxiang 453007,China;College of Software,Henan Normal University,Henan Xinxiang 453007,China;National Security Academy,People's Public Security University of China,Beijing 100038,China)

机构地区：[1]河南师范大学计算机与信息工程学院,河南新乡453007 [2]“智慧商务与物联网技术”河南省工程实验室,河南新乡453007 [3]河南师范大学软件学院,河南新乡453007 [4]中国人民公安大学国家安全学院,北京100038

出　　处：《智能系统学报》2023年第2期282-292,共11页CAAI Transactions on Intelligent Systems

基　　金：国家自然科学基金项目(61976082,62002103);河南省高等学校重点科研项目(22B520013);河南省科技攻关计划项目(222102210169)。

摘　　要：针对轻量化网络结构从特征图提取有效语义信息不足,以及语义信息与空间细节信息融合模块设计不合理而导致分割精度降低的问题,本文提出一种结合全局注意力机制的实时语义分割网络(global attention mechanism with real time semantic segmentation network,GaSeNet)。首先在双分支结构的语义分支中引入全局注意力机制,在通道与空间两个维度引导卷积神经网来关注与分割任务相关的语义类别,以提取更多有效语义信息;其次在空间细节分支设计混合空洞卷积块,在卷积核大小不变的情况下扩大感受野,以获取更多全局空间细节信息,弥补关键特征信息损失。然后重新设计特征融合模块,引入深度聚合金塔池化,将不同尺度的特征图深度融合,从而提高网络的语义分割性能。最后将所提出的方法在CamVid数据集和Vaihingen数据集上进行实验,通过与最新的语义分割方法对比分析可知,GaSeNet在分割精度上分别提高了4.29%、16.06%,实验结果验证了本文方法处理实时语义分割问题的有效性。To address the problems of insufficient extraction of effective semantic information from feature maps by lightweight network structure and the reduction of the segmentation accuracy due to the unreasonable design of the semantic information and spatial detail information fusion block.A real-time semantic segmentation network with the global attention mechanism(Global attention mechanism with real time semantic segmentation network,GaSeNet)is proposed in the paper.Firstly,a global attention mechanism is introduced into the semantic branch of the dual-branch structure,and the convolutional neural network is guided in the two dimensions of channel and space to focus on the semantic categories related to the segmentation task,so as to extract more effective semantic information.Secondly,a mixed hole convolution block is designed in the spatial detail branch,and the receptive field is enlarged while the size of the convolution kernel remains unchanged,so as to obtain more global spatial detail information and make up for the loss of key feature information.Then,the feature fusion module is redesigned and introduced the deep aggregation pyramid pooling module for deeply fuse feature maps of different scales,thereby improving the semantic segmentation performance of the network.Finally,the method proposed is tested on CamVid dataset and Vaihingen dataset.Compared with the latest semantic segmentation algorithm,it can be seen that the GaSeNet improves the segmentation accuracy by 4.29% and 16.06% respectively. The experimental results verify theeffectiveness of this method in dealing with real-time semantic segmentation problems.

关 键 词：实时语义分割 全局注意力机制 多尺度特征融合 混合空洞卷积 卷积神经网络 金字塔池化 感受野 特征提取 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]