Building NAS: Automatic designation of efficient neural architectures for building extraction in high-resolution aerial images.

Building extraction, which is a fundamental task in the community of remote sensing image analysis, has been extensively applied in various applications related to smart cities. Due to the complicated background information in urban areas, how to extract building footprints from high-resolution aerial images is challenging. The recent achievements of deep learning have shed light on building extraction and other remote sensing domain tasks. However, the heavy consumption of computational resources and the design of the neural architectures became the biggest bottleneck of utilizing deep learning techniques to improve the performance. In this work, we developed a Neural Architecture Search (NAS) algorithm, dubbed BuildingNAS, for building extraction from high-resolution aerial images. In particular, we built an efficient candidate operation set upon Separable Factorized Residual Blocks as our cell-level search space. Different from previous NAS in semantic segmentation tasks, we employed the hierarchical search space and proposed the Single-Path Sampling strategy to eliminate excessive GPU memory comsumption in searching process. In addition, we proposed an entropy regularized objective for the optimization of architecture parameters. As the result, the larger batch size can be adopted in the whole pipeline to accelerate the searching process, and make the resulted architecture more stable and accurate. We evaluated our proposed algorithm in WHUBuilding Dataset, the derived network achieved mIoU of 86.95% with only 2.05G FLOPs and 3.10 M parameters. The comparison results demonstrate that the network discovered by our algorithm can achieve great efficiency-accuracy trade-off. • The attempt of applying neural architecture search algorithm in segmentation of aerial images. • More lightweight hierarchical search space for neural architecture search of segmentation task. • Entropy regularized objective makes the searching process more effectively. • Better trades off the accuracy and efficiency of neural networks by using neural architecture search. • Achieved mIoU of 86.85% with only 2.05G Flops and 3.10 M parameters on WHUBuilding Dataset. [ABSTRACT FROM AUTHOR]