HFRA-Net:基于混合特征和残差注意力的遥感影像建筑物提取方法

HFRA-Net:a method for hybrid feature fusion and residual attention building extraction from remote sensing images

针对基于深度学习的遥感影像建筑物提取方法存在覆盖范围大的建筑物信息提取不完整和部分边界细节信息的丢失的问题,本文提出了一种新的采用混合特征融合和残差注意力机制的网络方法HFRA-Net。该方法引入混合特征融合模块,以并行结构捕获全局和局部特征,提高建筑物分割的完整性。同时,通过引入自上而下和自下而上双向注意力反馈的残差注意力机制,自适应地捕获不同尺度下的边界细节信息。最后,在两个公共数据集上的实验结果证明了提出方法对建筑物提取的有效性。

Building extraction from remote sensing images based on deep learning faces two challenges:incomplete extraction with large coverage and the loss of partial boundary details.To address these issues,a novel network model HFRA-Net using hybrid feature fusion and residual attention mechanism is proposed.The model introduces a hybrid feature fusion module to capture global and local features in a parallel structure,enhancing the integrity of building segmentation.At the same time,the residual attention mechanism of top-down and bottom-up bidirectional attention feedback is introduced to adaptively capture the boundary detail information at different scales.Finally,experimental results on two public datasets demonstrate the effectiveness of the proposed method for building extraction.