Build a real-time building segmentation network

Recently, most existing building segmentation networks have high memory usage and computational resource. And a large number of methods of the lightweight model have problems, such as the sharp decrease in segmentation performance due to parameter reduction, limited arithmetic power and real-time. Therefore, in this paper, we propose a new lightweight network based on the SGUnet segmentation network. Our network is a hourglass-like-depth-separable-convolution-based sgbunet segmentation network, pruning the sgnet and using residual connections. The core idea is that, shallow features are more important in lightweight networks. According to this point of view, increasing the proportion of features in the shallow part of the model successfully improved the accuracy of the model by 1%. Finally, the Cheap operation module is reused to further reduce the amount of network parameters. In the end, the parameter amount of the network is reduced from the original 260k to 70k, which greatly reduces the amount of model parameters.The sgbunet segmentation network have a reletive improvement, achieving 97.329 and 80.823 for Accuracy and Miou respectively in the building segmentation.