GCAT-U-Net嵌入全局坐标注意力机制的遥感地块分割网络

GCAT-U-Net:Remote sensing plot segmentation network with global coordinate attention

摘要:

遥感影像的地块背景特征复杂，当前地块分割方法不能较好地处理模糊的边缘信息，导致分割精度不理想。文章利用注意力机制处理地块特征，提出了一种基于全局坐标注意力机制的遥感地块分割网络：GCAT-U-Net。该方法在U-Net网络基础上嵌入了全局坐标注意力机制，加强了深度神经网络对于遥感影像数据中重要特征的关注度。在公开的GID数据集上的实验结果表明，文章提出的模型将准确率从0.9041提升到了0.9227，比传统U-Net网络提高了2百分点；结合特征自身重要性和特征位置信息的全局坐标注意力机制有助于更精确的目标定位，其输出相较于嵌入单一注意力机制，地块边界更为清晰，提升效果更为显著。

Abstract:

The background features of the plots in remote sensing images are complex. The current plot segmentation methods cannot handle the fuzzy edge information well, and the segmentation accuracy is not ideal. This article uses the attention mechanism to process the land parcel features, and proposes a remote sensing land parcel segmentation network based on the global coordinate attention mechanism: GCAT-U-Net. This method embeds the global coordinate attention mechanism on the U-Net network, which strengthens the deep neural network's attention to important features in remote sensing image data. The experimental results on the public GID data set show that the model proposed in the article increases the accuracy from 0.9041 to 0.9227, which is 2% higher than the traditional U-Net network. The global coordinate attention mechanism that combines the importance of the feature itself and the feature location information is helpful for more accurate target positioning. Compared with the embedded single attention mechanism, the output of the global coordinate attention mechanism is clearer and the improvement effect is more significant.