Semantic Segmentation of Buildings in Remote Sensing Images Based on Dense Residual Learning and Channel Adaption

Deep fully convolution neural network has opened a new field in semantic segmentation for remote sensing images. In this paper, an improved U-net model is proposed to extract buildings at the pixel level, so as to obtain its contour and size information. In this improved U-net model, the highly modular ResNeXt50 network is used as the encoder of U-net and a parallel dense residual module based on atrous convolution is proposed to extract multi-scale semantic information for segmentation. Moreover, the transposed convolution whose stride is 2 is used for upsampling to obtain the grayscale segmentation mask. This modified model adopts the sum of jaccard loss and binary cross entropy loss as the total loss function. Our experimental results have demonstrated that our improved U-net model has excellent performance on semantic information encoding and multi-scale feature extraction. Finally, the mean pixel precision (MPA), the mean intersection of union (MIoU), F1 score and AUC of our improved model are 93.17%, 81.51%, 86.95% and 0.9928 respectively. The F1 Score and AUC of our improved model are 11.3% and 11.4% higher than those of the standard U-net model respectively.