In the remote sensing area, how to automatically and accurately extract buildings from images is a hot and challenging topic in these years. With the rapid development of sensor and computer hardware technologies, it gets easier to gain remote sensing images with very high-resolution and extract buildings from them by the popular deep learning models such as Fully Convolutional Networks (FCN). However, current FCN based models always lead to blurred building boundaries and have poor abilities on extracting small buildings. Therefore, in this paper, we propose the Gaussian Dilate Convolution, which is a cascade of a trainable Gaussian Filter and an dilate convolution with proper hyperparameter initializations. Also, we carefully design a hierarchical dense feature fusion structure following the dense connection manners. Finally, we embed the Gaussian Dilate Convolution into the hierarchical dense fusion structure and name it as Dense Hierarchical Spatial Gaussian Pool (Dense-HSGP). More specifically, the Gaussian Dilate Convolution has the advantages of the original dilate convolution but preserves much more context information, while the hierarchical dense connection structure of Dense-HSGP provides more abundant receptive fields and higher feature reused abilities within the model. We execute the experiments on the widely used Inrial Labelling Dataset to verify the efficiency of the proposed model. The experimental results show that the proposed model achieves 96.45 % average accuracy and 77.17% IoU respectively, which are distinct improvements rather than several recent state-of-the-art building extraction models.
|