Road Information Extraction from Remote Sensing Images Based on Fully Convolutional Network

Aiming at the problem of low accuracy of road extraction from remote sensing images, a road extraction network MDC-Net (Multiscale Convolution Network) is proposed, which takes into account both the detail retention ability and the multi-scale feature extraction ability. This model extracts road feature information based on the encoder-decoder network structure. In order to reduce the convergence time of network training, a deep residual network Resnet34 is introduced as the pre-training model of the network. ASPP structure based on dilated convolution is added between the encoder and the decoder to extract the multi-scale feature information of the road. The experimental data adopts the DeepGlobe road extraction data set in the remote sensing image competition held by CodaLab in 2018, and compares and analyzes the experimental results with several classical fully convolutional network methods. The experimental results show that: (1) The road extraction model MDC-Net proposed in this paper has achieved good results in road extraction, which proves the feasibility of this method; (2) Compared with several classic network models in remote sensing images, roads The extraction effect shows better road extraction results in terms of road extraction accuracy and road connectivity.