Abstract:

In order to solve the problem of edge information loss in satellite image segmentation, a Multi-Attention Mechanism Network （MA-Net） algorithm is proposed to solve the problem of edge information loss. The framework of the algorithm adopts an end-to-end symmetric structure, which consists of two parts：encoding and decoding. In the coding part, the improved VGG16 network is used to extract the texture features of the lake, and in the decoding part, the Global average Pooling Attention fusion mechanism（GPA） is introduced to effectively fuse the texture features extracted in the coding part and obtain the high-resolution satellite image feature map. At the output of the network, attention mechanism module is added to fully extract the edge information of the lake and effectively segment the peninsula, island and small tributary of the lake. The experimental results show that the model has better segmentation accuracy than the existing semantic segmentation algorithm, and each segmentation index has been improved, and the model is verified to be universal on the public data set city scales.

Key words: semantic segmentation, satellite image segmentation, coding and decoding, attention mechanism module

摘要：

针对深度学习的语义分割法，在卫星图像分割中对半岛、小岛和湖泊细小支流的边缘信息提取丢失问题，提出了多注意力机制网络（MA-Net）卫星图像分割算法，弥补了边缘信息提取丢失问题。该算法的框架采用了端到端的对称结构，由编码和解码两部分组成。编码部分采用改进的VGG16网络提取湖泊的纹理特征，解码部分引入全局平均池化注意力融合机制（GPA），能够有效融合编码部分提取的纹理特征，得到高分辨率的卫星图像特征图。在网络的输出端加入注意力机制模块（Attention），充分提取湖泊边缘信息，有效分割出半岛、小岛和湖泊细小支流。实验结果表明，该模型相比现有语义分割算法，具有更好的分割精度，各项分割指标都有提升，并且在公共数据集City Scapes上验证了模型具有通用性。

关键词: 语义分割, 卫星图像分割, 编码和解码, 注意力机制模块