针对遥感图像中目标类型多、目标尺寸小、目标大小差异较大、图像视场大、环境和背景复杂等特点导致其分割困难的问题,提出一种注意力空洞多尺度U-Net (ADMSU-Net)的遥感图像中目标分割方法。ADMSU-Net由收缩子网、扩展子网、空洞残差卷积连接和空间注意连接组成,在收缩和扩展子网中引入空洞多尺度Inception模块,在不增加计算成本的情况下学习多尺度高级特征,在跳跃连接中加入空间注意机制,提取特征之间的空间相关性,提高模型的分割性能。在包含多尺度目标的遥感图像数据集EORSSD上的实验结果表明,该方法是有效可行的,分割准确率为93%以上。

Aiming at the problem of object segmentation caused by the characteristics of multiple types, small size, large-size difference, large field-view, and complex environment-background of multi-scale targets in remote sensing images (RSIs), an attention dilated multi-scale U-Net (ADMSU-Net) is constructed for multi-scale target segmentation in RSIs (TSRSI). It consists of contracting subnet, expanding subnet, dilated convolutional residual connection and spatial attention connection. In the model, dilated multi-scale Inception module is introduced into its contracting and expanding subnets to learn the multi-scale advanced features without increasing computational cost, and the spatial attention mechanism is added into skip connection to capture the spatial correlation between features and improve TSRSI performance. The experimental results on the RSI dataset EORSSD containing multi-scale targets show that this method is effective and feasible, and the segmentation accuracy is more than 93%.