【摘要】 遥感影像中的建筑物是城市大数据采集、分析的重要来源。大规模、高精度的遥感影像建筑物提取模型对智慧城市时空大数据建设、推动城市智能计算具有重要意义。当前建筑物提取模型通常利用大型卷积神经网络模型或多种网络模型串联，并辅以其他边界细化算法来提高建筑物提取的精度。但是，网络模型的大型化、复杂化对计算资源消耗高，需要更多的训练时间或算力，不利于大规模快速的网络模型训练预测及在便携式等终端设备上部署应用。因此，研究面向大规模快速的遥感影像建筑物提取，提出一种轻量化全卷积神经网络模型和特征融合方案，模型参数较轻量化前减少约40%,GPU内存占用下降33.61%，平均训练时间和预测时间分别下降32.40%和26.31%。融合后的模型在公开数据集测试得到的MIoU精度在74.14%左右，达到了保证高精度建筑物提取前提下模型轻量化的预期。更多还原

【Abstract】 Buildings in remote sensing images are an important source of urban big data collection and analysis.Developing large-scale and high-precision models for extracting buildings from remote sensing images is significance for building spatio-temporal big data platform of smart city. It also helps to the promote of urban intelligent computing. Current building extraction models usually use large-scale convolutional neural networks or multiple networks in series, supplemented by other boundary refinement algorithms to improve the accuracy of extracting buildings. However, the large-scale and complex network models consume high computing resources and require more training time or computing power, which is not conducive to large-scale and rapidly train network and do prediction. Also, the large-scale and complex models limit the deployment and application on portable and other terminal devices. Therefore, based on the idea of a cancellation balance method, this paper proposes a lightweight full convolutional neural network model and feature fusion scheme for large-scale and rapid building extraction from remote sensing images. The model parameters are reduced by about 40% compared with before lightweight. GPU memory usage is reduced by 33.61%, and the average training time and prediction time are reduced by 32.40% and 26.31%, respectively. The MIoU accuracy of the model after feature fusion is about 74.14% in the public data set, which meets the expectation of lightweight for building extraction models under the premise of ensuring high-precision accuracy.更多还原