结合全局注意力机制的实时语义分割网络

PDF下载HTML

Title:

Global attention mechanism with real-time semantic segmentation network

作者:

李涛^1,2, 高志刚³, 管晟媛⁴, 徐久成^1,2, 马媛媛¹

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

Author(s):

LI Tao^1,2, GAO Zhigang³, GUAN Shengyuan⁴, XU Jiucheng^1,2, MA Yuanyuan¹

1. College of Computer and Information Engineering, He’nan Normal University, Xinxiang 453007, China;
2. Engineering Lab of He’nan Province for Intelligence Business & Internet of Things, Xinxiang 453007, China;
3. College of Software, He’nan Normal University, Xinxiang 453007, China;
4. National Security Academy, People’s Public Security University of China, Beijing 100038, China

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

Keywords:

real-time semantic segmentation; global attention mechanism; multiscale feature fusion; hybrid dilated convolution; convolutional neural network; pyramid pooling; receptive field; feature extraction

分类号:

TP391

DOI:

10.11992/tis.202208027

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。

Abstract:

The lightweight network structure cannot sufficiently extract effective semantic information from feature maps, and the unreasonable design of the semantic information and spatial detail information fusion block leads to a decrease in segmentation accuracy. To address these problems, a global attention mechanism with a real-time semantic segmentation network (GaSeNet) is proposed in the paper. First, a global attention mechanism is introduced into the semantic branch of the dual-branch structure. The convolutional neural network is then guided in the two dimensions of channel and space to focus on the semantic categories related to the segmentation task to extract remarkably effective semantic information. Second, a mixed hole convolution block is designed in the spatial detail branch, and the receptive field is enlarged while maintaining the size of the convolution kernel to obtain additional global spatial detail information and compensate for the loss of key feature information. The feature fusion module is then redesigned, and the deep aggregation pyramid pooling module is introduced to fuse feature maps of different scales comprehensively, thereby improving the semantic segmentation performance of the network. Finally, the proposed method is tested on CamVid and Vaihingen datasets. Compared with the latest semantic segmentation algorithm, GaSeNet improves the segmentation accuracy by 4.29% and 16.06%. Experimental results verify the effectiveness of this method in dealing with real-time semantic segmentation problems.