Automatic road extraction framework based on codec network

Road extraction based on deep learning is one of hot spots of semantic segmentation in the past decade. In this work, we proposed a framework based on codec network for automatic road extraction from remote sensing images. Firstly, a pre-trained ResNet34 was migrated to U-Net and its encoding structure was replaced to deepen the number of network layers, which reduces the error rate of road segmentation and the loss of details. Secondly, dilated convolution was used to connect the encoder and the decoder of network to expand the receptive field and retain more low-dimensional information of the image. Afterwards, the channel attention mechanism was used to select the information of the feature image obtained by up-sampling of the encoder, the weights of target features were optimized to enhance the features of target region and suppress the features of background and noise regions, and thus the feature extraction effect of the remote sensing image with complex background was optimized. Finally, an adaptive sigmoid loss function was proposed, which optimizes the imbalance between the road and the background, and makes the model reach the optimal solution. Experimental results show that compared with several semantic segmentation networks, the proposed method can greatly reduce the error rate of road segmentation and effectively improve the accuracy of road extraction from remote sensing images.