一种基于图卷积的图像注意力模型

An Image Attention Model Based on Graph Convolution

现有的注意力模型往往没有考虑到空间与通道的联系,还忽视对全局特征信息。文章提出了一种基于图卷积网络的注意模型,将特征图压缩并映射成图结构,再借助图卷积网络处理拓扑结构数据的能力,得以提取全局特征信息学习特征权重图。另外,空间注意力与通道注意力一体化结构能够更有效地学习特征权重。通过多个实验测试表明,在图像分类任务中,展现了优异的性能和普遍适用性。

Existing attention models often do not consider the connection between space and channel, and also ignore the global feature information. this paper proposes an attention models based on graph convolutional network, which compresses and maps the feature map into a graph structure, and then uses the ability of graph convolutional network to process topology data to extract global feature information and learn feature weight map. In addition, the integrated structure of spatial attention and channel attention can learn feature weights more effectively. Multiple experimental tests show that it exhibits excellent performance and general applicability in image classification tasks.