融合注意力机制的端到端的街道场景语义分割  

End-to-end Semantic Segmentation of Street Scene with Attention Mechanism

作　　者：吴琼 瞿绍军 WU Qiong;QU Shao-jun(College of Information Science and Engineering,Hunan Normal University,Changsha 410081,China;Hunan Xiangjiang Artificial Intelligence Academy,Hunan Normal University,Changsha 410081,China)

机构地区：[1]湖南师范大学信息科学与工程学院,长沙410081 [2]湖南师范大学湖南湘江人工智能学院,长沙410081

出　　处：《小型微型计算机系统》2023年第7期1514-1520,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(12071126)资助;湖南省教育厅科学研究项目(19C1149)资助。

摘　　要：随着注意力机制的出现,研究人员开始通过对神经网络的特征图中的不同关注角度添加注意力机制来提升图像语义分割准确率.现有融合注意力机制的方法,大多通过计算像素与其它像素之间的关系来求得该像素点的注意力权值,这些方法提升效果具有局限性.针对以上问题,设计了两种不同的注意力模块:空间-条形注意力和通道注意力,提出了一种融合注意力机制的端到端的街道场景语义分割方法.空间-条形注意力采用两种不同条形池化核,捕获条形分布区域之间的长距离依赖关系,可以有效捕获局部和全局上下文信息.在Cityscapes数据集上的实验表明,该算法在验证集和测试集上的分割精度分别为76.89%、77.8%,与现有的算法相比有更好的表现.本文结合人类视觉注意力的特点,提出的两个注意力模型,使得街道场景下的语义分割效果更加准确.With the emergence of attention mechanism,researchers began to add attention mechanism to different attention angles in the feature graph of neural network to improve the accuracy of image semantic segmentation.The existing methods of integrating attention mechanism mostly calculate the attention weight of the pixel point by calculating the relationship between the pixel and other pixels.These methods have limitations in improving effectiveness.To solve the above problems,two different attention modules are designed:spatial-bar attention and channel attention,and a semantic segmentation method of street scene based on the fusion attention mechanism of fully convolution network is proposed.Spatial-bar attention can capture local and global context information effectively by using two different bar pooling cores to capture long distance dependencies between strip distribution regions.Experiments on Cityscapes data set show that the segmentation accuracy of the proposed algorithm is 76.89%and 77.8%,respectively,which is better than the existing algorithms.Combining with the characteristics of human visual attention,this paper proposes two attention models to make the semantic segmentation effect more accurate in street scenes.

关 键 词：注意力机制 全卷积网络 条形池化 语义分割 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]