Abstract: The traditional features extraction of residential building is mainly based on a large number of on-site observations, measurement and sampling, features summarization, the representative shape, color and structural style abstraction according to personal experience. Such methods are easily affected by personal subjective factors, and with fault on workload, reproducibility, stability, reliability and scientificity. To solve the above problems, a novel extraction method of plateau residential features in Xizang is proposed based on comprehensive application of computer vision, statistical analysis and remote sensing imaging. Firstly, the K-means unsupervised clustering method is used to extract the shape features of residential buildings and obtain typical characteristic indexes such as the length, width and aspect ratio of typical residential buildings. Secondly, the RGB three-color residential building samples are transformed into Lab color space, and the typical color features of residential buildings are extracted through color difference threshold segmentation and statistical analysis. Finally, the multi-scale structural similarity of samples is introduced to quantitatively characterize the architectural style, and typical residential architectural styles are extracted through sample selection, structural feature template extraction and maximum similarity matching. Two groups of comparative experiments with four classical mainstream methods and ours are carried out in Sangye Town and Qushui Town of Xizang as experimental areas. The experimental results show that the typical features extracted by the proposed method are consistent, stable and reliable, and less affected by human factors. At the same time, it is also found that the length-width ratio of the plateau residential buildings is about 1.1∶1, the length is about 24~29?m, and the width is slightly narrower than the length. The color of residential buildings is mainly light gray-brown, and the top 3 to 4 main colors with the largest proportion are very similar with that of local bare soil. The typical residential building style is mainly rectangle like the shape Chinese of “Hui”, with a courtyard in the middle and walls around.

Key words: residential building, typical feature, information extraction, computer vision, multi-scale structural similarity

摘要： 传统的民居建筑典型特征提取主要是通过大量现地观察、测量和采样后，凭借个人经验进行提炼总结，抽象概括出有代表性民居建筑形状、颜色和样式等造型特征。该类方法容易受个人主观因素影响，工作量大，可复现性不强，稳定性、可靠性和科学性不高。针对以上问题，结合计算机视觉、统计分析和遥感成像等技术手段提出一种西藏地区高原民居建筑典型特征遥感提取方法。利用K-means无监督聚类对民居建筑形状特征进行提取，获取典型民居建筑长度、宽度和长宽比等典型形状特征指标；将RGB三色的民居建筑样本转换到Lab色彩空间，通过色差阈值分割与统计分析提取民居建筑典型颜色特征；引入样本多尺度结构相似度对建筑样式进行定量化表征，通过样本精选、结构特征模板提取和最大相似度匹配提取典型民居建筑样式。以西藏地区的桑耶镇和曲水镇作为实验区分别选择四种经典主流方法开展了两组对比验证实验。实验结果表明，该方法提取的建筑典型特征一致性较好，结果稳定、可靠，受人为因素影响较小。同时也发现，高原民居建筑长宽比约为1.1∶1，长度约24~29?m，宽度略窄。民居建筑颜色主要为浅灰褐色，占比最大的前3~4种主色与当地的裸土颜色非常相似。典型民居建筑样式主要呈“回”字形结构，中间有庭院和四周为围墙。

关键词: 民居建筑, 典型特征, 信息提取, 计算机视觉, 多尺度结构相似度