基于类别注意力卷积网络的地物分类方法

在近期的研究发展中，语义分割取得了巨大的进步。但大多数方法都是从空间角度出发，来获取更加丰富的上下文信息。与以往的方法不同，本文提出了一种基于类别注意机制的特征融合方法，从类别角度出发，来获取全局上下文信息，并与其他特征信息进行融合，这种方法能够更好地表示图像中各类目标的特征，具有更好的类内聚合性。为此，本文使用了一种ACF（类别注意力特征）模块，以计算和构建图像中各类目标的类别中心，以此为基础得到了一个基于类别注意力的多特征融合语义分割网络，以实现更好的地物分类性能。算法使用ISPRS数据集进行实验，与其他算法进行对比，本文方法具有更好的性能表现。 

Land cover classification based on class attention convolution network

In recent years, semantic segmentation has made great progress. But most of the methods are from a spatial perspective to obtain richer context information. Different from the previous methods, this paper proposes a feature fusion method based on class attention mechanism, which obtains the global context information from the perspective of category and fuses it with other feature. This method can better represent the features of various objects in the image and has better intra class aggregation. Therefore, this paper uses an ACF (attentional class feature) module to calculate and construct the category centers of all kinds of objects in the image. Based on this, a multi feature fusion semantic segmentation network based on category attention is obtained to achieve better classification performance. The algorithm uses ISPRS data sets for experiments, and compared with other algorithms, the proposed method has better performance.