Abstract:

The current deep learning method has excessive segmentation of recognition results and poor recognition of small objects in high-resolution remote sensing images. In order to solve this problem, an improved network model AA-SegNet based on SegNet architecture is proposed, and an enhanced spatial pyramid pooling module and spatial attention fusion module are added. The network can enhance feature propagation and can effectively deliver higher levels of feature information to suppress noise of low-level features, and can enhance small-target feature learning. Based on the high-resolution 2 remote sensing image dataset and experiment, the overall recognition accuracy of AA-SegNet network is 96.61%, which is superior to SegNet, U-Net and DeepLab-V3 networks in recognition rate, [F1] score and training time.

Key words: deep learning, building identification, high resolution remote sensing, enhanced spatial pyramid model, attention mechanism, semantic segmentation

摘要：

为解决当前深度学习方法在高分辨率遥感图像中存在识别结果过度分割，以及小物体识别差的问题，提出一种基于SegNet架构改进的网络模型AA-SegNet，增加了增强的空间金字塔池化模块和空间注意力融合模块。该网络可以加强特征传播并能够有效传递更高级别的特征信息以抑制低级特征的噪声，并且可以增强小目标特征学习。基于高分二号遥感影像制作数据集并进行实验，AA-SegNet网络总体识别准确率为96.61%，在识别率、[F1]分数以及训练时间等方面也都优于SegNet、U-Net、DeepLab-V3网络。

关键词: 深度学习, 建筑识别, 高分辨率遥感, 增强型空间金字塔模型, 注意力机制, 语义分割