Dilated Convolutional Network for Road Extraction in Remote Sensing Images

Abstract

According to the characteristics of the road in the remote sensing images, we de-sign a dilated convolutional neural network (DCN) for road extraction. In order to minimize the problem of feature disappearance caused by down-sampling, we use dilated convolution of the traditional convolution in partial convolutional process instead. (dilated convolution has been used to partly replace traditional convolution) At the same time, DCN reduces the number of pooling layers. The ad-vantage is that the receptive field largely increases without increasing the number of parameters. As the road area is a small proportion in remote sensing images, the training set is pre-filtered according to the proportional distribution of positive and negative samples, which avoids the bad influences of the imbalance between positive and negative samples in training process. The experimental results show that DCN outperforms all the baselines in terms of precision, recall, and F1 scores.