顾及多尺度特征及全局上下文的建筑提取方法

Building Detection Method Considering Multi-scale Features and Global Context

针对语义分割提取建筑物时,在特征提取过程中丢失局部细节信息,对全局上下文信息的感知能力及多尺度特征的提取不足,导致小建筑物漏提、建筑物提取不完整及内部孔洞的问题,提出了顾及多尺度特征及全局上文信息的建筑物提取方法。该方法采用编码-解码结构,利用并行的连续空洞卷积提取多尺度特征,并行使用压缩激励模块(SE)和条带池化模块(SPM)从通道和空间维度捕获全局上下文信息,提高网络对小建筑物的识别能力及提取结果的完整性,并减少内部孔洞。通过在WHU建筑数据集和Inria航空数据集上与常见的语义分割网络进行的对比实验表明,该方法在提高建筑物提取准确率的同时,较好地解决了小建筑物漏提、建筑物提取不完整及内部孔洞等问题。

Building extraction from remote sensing images is a semantic segmentation task.However,the local detail information may lost in the encoding stage.The perception ability of global context and the extraction of multi-scale features are insufficient,resulting in the omission of small buildings,incomplete extraction of buildings and internal holes.To solve the above problems,we propose a method considering global context and multi-scale features for building extraction.The method adopts an encoder-decoder structure and contains two core modules.One is the multiscale feature encoding module,which uses parallel continuous dilated convolution to extract multi-scale features.The other is the global context-aware module,which consists of the squeeze and excitation module and the strip pooling module,and is used to obtain sufficient global context information from the channel dimension and spatial dimension.Experimental results on the WHU building dataset and Inria aerial image labeling dataset indicate that the proposed method solves the problems of small buildings omission,incomplete extraction and internal holes while improving the accuracy.