基于改进DeepLabV3+的高分辨率卫星影像分割方法

High-resolution satellite image segmentation method based on improved DeepLabV3+

针对利用DeepLabV3+进行高分辨卫星影像分割时存在的边缘分割精度不足和特征信息丢失等问题,本文提出了一种基于改进DeepLabV3+的高分辨率卫星影像分割方法。首先,调整空间金字塔池化(ASPP)模块的扩张率,使模型更好地适应高分辨率影像的特征,提高特征的提取能力;其次,在编码阶段引入注意力机制(CBAM),动态优化通道的权重和位置信息,加强对影像关键特征信息的学习;最后,在解码阶段引入自适应空间融合(ASFF)模块,将不同尺度的特征聚合成解码器的输入特征,提高语义分割结果的精度。结果表明,相较于传统语义分割,本文方法的总体精度、精确度、交并比均有明显提高,能够有效解决高分辨率卫星影像分割的问题。

As to the problems of poor edge segmentation accuracy and feature information loss when DeepLabV3+is used to perform high-resolution satellite image segmentation,this paper proposes a high-resolution satellite image segmentation method based on improved DeepLabV3+.Firstly,the expansion rate of the atrous spatial pyramid pooling(ASPP)module is adjusted to better adapt the model to the features of high-resolution images and improve the feature extraction capability.Then,the convolutional block attention module(CBAM)is introduced in the encoding stage to dynamically optimize the weight and location information of channels and enhance the learning of key feature information of images.Finally,the Adaptively Spatial Feature Fusion(ASFF)module is introduced in the decoding stage to aggregate features of different scales into the input features of the decoder and im‐prove the accuracy of the semantic segmentation.The results show that the overall accuracy,precision and intersection ratio from the proposed method are significantly improved compared with those from the traditional semantic segmentation,it can be used to effectively solve the problem of high-resolution satellite image segmentation.