Deep Neural Network Based on Sparse Auto-Encoder for Road Extraction

Road extraction from aerial image has realistic significance for GIS data updating. In view of the complexity challenging for acquiring road information, this paper proposes supervised model that combines Convolutional Neural Network (CNN) with Sparse Auto-Encoder (SAE) to cope with the road extraction task. First, the road features are extracted from the amount of non-annotated data using SAE model that aim to train the road features using CNN principle with implementing convolution and pooling to reduce model complexity. Second, the encoder network completes the operation, and after the deep pooling and deconvolution operations, the intermediate features are extracted by the decoder network and sampled back to the input image of the same size on the map. Third, the soft-max classifier categorizes images into roads and non-roads. Finally, the experiments verify that the proposed method outperforms the traditional methods and could achieve the satisfy result.