In order to solve the problem of road extraction information loss in remote sensing images, a road extraction algorithm based on residual neural networks was proposed. Firstly, an encoder-decoder network was constructed, combined with pre-coder and dilated convolution module to extract more semantic information.Secondly, the parallel designed dilated convolution module was added to the middle part of the encoder-decoder structure, which could extract features of different receptive field features. Finally, the encoder-decoder used jumper to perform multi-scale feature fusion, learning more low-dimensional and high-dimensional features.In the Massachusetts road dataset, this method had 11 %, 0.3 %, and 7.4 % improvement in Precision, Recall, and F1-score performance indicators. At the same time, it also achieved 97.9 % in the Accuracy index. Compared with other algorithms, the algorithm has certain application value.