基于U-Net神经网络改进的建筑物分割算法研究

Research on improved building segmentation algorithm based on U-Net

为解决高分辨图像分割中存在的普遍问题,文章提出了一种基于U-Net改进的建筑物分割算法。在编码器中利用迁移学习解决标记数据集的稀缺性问题,在跳过连接中引入特征细化块来细化特征,在解码器中加入通道注意力机制选择更多具有辨别力的特征,算法在Inria航拍图像标记数据集上验证可行性。

To solve the common problems in high-resolution image segmentation,this paper propose an improved building segmentation algorithm based on U-Net.Transfer learning is used in the encoder to solve the scarcity problem of labeled datasets,a feature refi nement module is introduced in skip connections to refi ne features,and a channel attention mechanism is added in the decoder to select more discriminative features.The algorithm is validated on the Inria aerial image labeling dataset.