Automatic building extraction has significant socio-economic applications such as population estimation, urban planning, rapid disaster response, monitoring illegal land acquisition, and change detection. The traditional methods fail to deal with the challenges entirely associated with building extraction, i.e. different shape, size, texture of buildings, missing and incomplete buildings due to the occlusion, and high intra-class variation. The existing CNN-based approaches are incapable of recovering the boundary information, especially when the building structures are small and complex. To alleviate the issues faced by current methods, we propose a lightweight attention mechanism-based model – refined cross attention neural network (RCA-Net) for precisely extracting the coarse-to-fine building features. Unlike recent attention mechanism-based approaches, the RCA-Net utilizes spatial and channel attention to capture the long-range multi-scale context. Then, we introduce an efficient attention module, the Global Attention Fuse (GAF) module, that fuses the local and global cross-channel relationships to capture the essential features without enhancing the computational complexity. A loss function, unified loss, is also presented that combines BCE loss and dice loss to alleviate the imbalanced class distribution problem. Experimental results show that our proposed method outperforms the latest method DSNet by 2.06% and 1.47% in IoU and 2.11% and 1.27% in F1-score on the publicly available datasets: Massachusetts building dataset and Inria Aerial Image Labeling dataset, respectively.

自动建筑物提取具有重要的社会经济应用，例如人口估计、城市规划、快速灾害响应、监测非法征地和变化检测。传统方法无法处理与建筑物提取完全相关的挑战，即不同形状、大小、纹理的建筑物，由于遮挡导致的建筑物缺失和不完整，以及高的类内变化。现有的基于 CNN 的方法无法恢复边界信息，尤其是在建筑结构小而复杂的情况下。为了缓解当前方法面临的问题，我们提出了一种基于轻量级注意力机制的模型——精细交叉注意力神经网络（RCA-Net），用于精确提取从粗到细的建筑特征。与最近基于注意力机制的方法不同，RCA-Net 利用空间和通道注意力来捕获远程多尺度上下文。然后，我们引入了一个高效的注意力模块，全局注意力保险丝（GAF）模块，它融合了局部和全局的跨通道关系，在不增加计算复杂性的情况下捕获基本特征。还提出了一个损失函数，统一损失，它结合了 BCE 损失和骰子损失来缓解不平衡的类分布问题。实验结果表明，在公开可用的数据集：Massachusetts building 数据集和 Inria Aerial Image Labeling 数据集上，我们提出的方法在 IoU 和 F1-score 上分别优于最新方法 DSNet 2.06% 和 1.47% 和 2.11% 和 1.27%。