Building segmentation of aerial images in urban areas with deep convolutional neural networks

Abstract

Building segmentation of aerial images in urban areas is of great importance for many applications, such as navigation, change detection, areal monitoring and urban planning. However, due to the uncertainties involved in images, a detailed and effective solution is still critical for further applications. In this paper, we proposed a novel deep convolutional neural network for building segmentation of aerial images in urban areas, which was based on the down-sampling-then-up-sampling architecture. The suggested network is similar to that of the FCN, but with ours differs as it takes into account the multi-scale features using Atrous Spatial Pyramid Pooling. Additionally, motivated by the recent published works, the depth-wise separable convolution was also adopted to replace the standard convolution in our proposed method, which largely reduced the training parameters. To evaluate the performance of our proposed method, a very high resolution aerial image dataset (0.075 m) was used to train and test the images. In addition, two state-of-the-art methods named FCN-8s and SegNet were also compared with our method for performance evaluations. The experiments demonstrated that our method outperformed the state-of-the-art methods greatly both in terms of qualitative and quantitative performance.