基于全局注意力机制的语义分割方法研究

Research on Semantic Segmentation Based on Global Attention Mechanism

如何捕获更长距离的上下文信息成为语义分割的一个研究热点,但已有的方法无法捕获到全局的上下文信息。为此,文章提出了一种全局注意力模块,其通过计算每个像素和其他像素之间的关系生成一个全局关系注意力谱,然后通过该全局注意力谱来对深层卷积特征进行重新聚合,加强其中的有用信息,抑制无用的噪声信息。在具有挑战性的Cityscapes和PASCAL VOC 2012数据集上验证了所提出的方法具有有效性其优于现有的方法。

How to capture the context information with longer distance has become a research hotspot of semantic segmentation,but the existing methods can not capture the global context information.This paper proposes a global attention module,which generates a global relation attention spectrum by calculating the relationship between each pixel and other pixels,and then reaggregates the deep convolution features through the global attention spectrum to strengthen the useful information and suppress the useless noise information.The validity of the proposed method is verified on the challenging Cityscape and PASCAL VOC 2012 datasets,which is superior to the existing methods.