Deep Learning Road Extraction Model Based on Similarity Mapping Relationship

This paper proposed a deep learning road extraction model based on similarity mapping relationship for extracting roads from high resolution remote sensing images. The model adopted a full convolutional network structure, and used the similarity between each pixel in the input image and the typical label sample to perform mapping relationship search, thereby, the classification of each pixel in the image to be classified is marked to the classification category of the typical sample. The model breaks through the traditional deep learning mindset and directly stores knowledge in the network, instead of just learning a set of feature extraction and integrated network parameters. Under the training data of 10,000 small samples, the corresponding image results are obtained after 100,000 iterations. The experimental results show that the accuracy of the method is 0.87, which proves that this method is more accurate and efficient than traditional algorithms.