Research On Automatic Classification Method Of Building Styles Based On Live Images

Cities need to develop,but also need to be protected.How to effectively distinguish building types and classify urban functions in a growing city,and at the same time preserve the unique architecture of the city to overcome the phenomenon of uniformity,has become an important issue in urban preservation and renewal work.In recent years,with the continuous development of deep learning technology,it has been widely used in the field of image recognition and classification.Street view images,which have more perspectives and can show the characteristics of buildings more intuitively,have also been widely used in the management and evaluation of urban built environment.In this context,this paper carries out an automated classification and evaluation of urban building styles based on streetscape images and convolutional neural networks to provide auxiliary support for urban conservation and renewal work.The specific work mainly includes:(1)Firstly,the streetscape building images are analyzed,and it is found that the streetscape buildings have complex backgrounds and too much interference information,while considering the lack of streetscape building datasets for image classification in China at present.Therefore,this paper firstly establishes two datasets for domestic buildings,named CB＿FUN and CB＿STY,and preprocesses the sample datasets to enhance the usability of the sample datasets.At the same time,the semantic segmentation method is used to extract the target buildings and provide good quality data support for the design of building classification models.(2)Secondly,for the design of the building classification model,this paper selects the Res Net-50 and Inception-V3 networks with high usability as the base networks of the classification model by analyzing the currently popular convolutional neural networks,introduces the attention mechanism for the situation that the gap between classes is too small and the gap within classes is too large for the target buildings,and also fine-tunes the The network structure and optimized parameters are trained on the two types of data sets established in this paper.The training results show that the improved convolutional neural network achieves 86.2% and 88.4% classification accuracy on the two types of datasets,respectively.Both achieved better classification results compared with the base network.(3)Finally,according to the trained building classification model,the Nanluoguxiang neighborhood in Dongcheng District of Beijing was selected as the experimental area to classify and evaluate the buildings in the street,and visualized with the location information of Baidu Street View and Arc GIS tools.The results found that with the rapid development of commercialization of Nanluoguxiang,the percentage of buildings showing commercial characteristics in the area has reached 16.2%,which makes the style of some characteristic buildings suffer,and puts forward higher requirements for the protection and development of buildings in the area in the future.This study is based on street view image data and uses convolutional neural network to classify urban buildings for research,the method is not only simple in logic operation but also has strong usability.The proposed method system provides new research ideas for the study of the distribution characteristics of buildings in different functional areas of different cities,and also has a strong practical application value for the protection of characteristic buildings in the process of rapid urban development.