结合MobileNet的改进DeepLabv3+遥感影像道路提取方法

Road extraction method based on improved DeepLabv3+remote sensing images combined with MobileNet

针对现有深度学习网络提取遥感影像中道路时存在的分割精度较低、错检率高、检测效率较低等问题,本文提出一种改进模型MB-DeepLabv3+。在编码器层,利用MobileNetv3作为特征提取网络,同时引入全局注意力机制计算空间与通道层面的样本注意力权重;在解码器层,以密集上采样卷积核代替双线性内插法进行特征图上采样。在Deep Globe数据集上的实验结果表明,本文所提出算法在准确率上达到了98.39%,较原始DeepLabv3+提高了2.6%,并且对单幅影像的提取计算效率也有所提升,面对不同复杂程度的道路影像,较对照组内其余模型在漏洞、错误提取等问题上有了显著的改善,能够实现高效精准的道路提取。

The existing deep learning network faces problems of low segmentation accuracy,high false detection rate,and low detection efficiency when extracting roads in remote sensing images.In order to address these issues,an improved model,namely MB-DeepLabv3+was proposed.At the encoder layer,MobileNetv3 was used as the feature extraction network,and a global attention mechanism was introduced to calculate the sample attention weight at the spatial and channel levels;at the decoder layer,a dense upsampling convolution kernel was used instead of the bilinear interpolation method to perform upsampling in feature maps.The experimental results on the Deep Globe dataset show that the accuracy of the proposed algorithm reaches 98.39%,which is 2.6%higher than that of the original DeepLabv3+,and the extraction and calculation efficiency on a single image is improved.For road images with different levels of complexity,the proposed algorithm significantly improves the problems of loopholes and false extraction and can realize efficient and accurate road extraction compared with other models in the control group.