一种支持建筑群组相似模式检索的变分图卷积自编码模型

A Variational Graph Convolutional Auto-encoder Model Supporting Similarity Retrieval of Building Groups Patterns

针对当前建筑群组相似模式检索中空间域和谱域相似度量方法的模型泛化能力较弱,且空间信息利用不足的问题,本文提出了一种基于变分图卷积自编码器的相似检索模型。首先,利用最小生成树构建建筑群组图数据,并将建筑物描述特征嵌入图节点中,实现建筑群组模式的定量化表达;其次,在变分图卷积自编码器重参数化模块中,加入四维超混沌系统与高斯分布融合生成的混合噪声,通过增加采样的随机性来提高模型的泛化能力;然后利用图卷积和池化代替全连接层以保留更多的空间信息,并对低维特征编码进行处理,获得建筑群组的模式特征编码;最后,利用余弦距离来获得待检索建筑群组与模式编码库中建筑群组之间的相似度,从而实现建筑群组模式的相似检索。实验结果表明,该模型能够有效提取建筑群组的模式特征,通过无监督学习实现端到端的建筑群组相似模式检索,为建筑群组模式的自动分类与相似检索提供了新的思路和方法。

To address the problems of weak model generalization and insufficient utilization of spatial information due to the strict constraints in the fusion method of geometric features and spatial relations and the spectral domain similarity metric in the current building group similar pattern retrieval,this paper proposes a similar retrieval model based on the variational graph convolutional auto-encoder.Firstly,the minimum spanning tree is used to construct the building group graph structure and embed the building description features into the graph nodes to realize the quantified representation of the building group pattern.Secondly,in the reparameterization module of the variogram self-encoder,the mixed noise is generated by fusing the fourdimensional hyperchaotic system and Gaussian distribution to improve the generalization ability of the model by increasing the randomness of sampling.Then the graph convolution and pooling are used instead of fully connected layer to retain more spatial information.And the low-dimensional features are processed to obtain the feature encoding of building group patterns.Finally,the cosine distance is used to obtain the similarity between the building group to be retrieved and the building group in the pattern encoding library,so as to achieve the similarity retrieval of building group patterns.The experimental results show that the proposed model can effectively extract the pattern features of building groups and realize end-to-end similar pattern retrieval of building groups through unsupervised learning,which provides a new idea and method for automatic classification and similarity retrieval of building group patterns.