Research On Architectural Style Recognition Based On Deep Learning

Architectural styles are not only a comprehensive reflection of historical periods,geographical regions and cultures,but also a unique imprint of social background and humanistic spirit.They are an important part of cultural heritage and have irreplaceable value.Digital documentation is one of the effective means to preserve architectural cultural heritage.It not only provides data support for archival records and value assessment of architectures,but also provides diverse forms of expression for stylistic analysis and educational display of architectures,as well as scientific analysis tools for conservation management and utilization planning of architectures.However,the creation of digital documentation is a complex and arduous task,which requires both the handling of large amounts of data and high costs,as well as relying on professional talents to effectively organize,classify and analyze the data.Computer vision techniques can be an effective tool to enhance the efficiency and scientific rigor of digital analysis of architectural cultural heritage by automatically identifying their styles.This paper presents a comprehensive study of architectural style recognition using computer vision algorithms,based on the summarization and analysis of the characteristics of different architectural styles.The research covers the following aspects:(1)Existing studies on architectural styles mainly focus on Western architecture,while there is a relative lack of research on Chinese ancient architectural styles.Therefore,this thesis generates the first dataset by images of Chinese ancient architecture styles,which contains images of ancient architectures and their annotation information in different chronological and cultural contexts.Firstly,the existing ancient architectures are studied selected in detail,the ages and characteristics of each architecture are specified.Next,a web crawler method based on the Scrapy framework is used to search and download images of architecture exteriors taken in the field based on architecture names.Then,the corrupted image data are eliminated by judging the file format of the images.After that,the histogram features of the images are calculated and the similarity between the histogram features is measured using the Barthian distance to remove duplicate images.Next,non-building images and images with mismatched architecture names are manually screened out.Finally,referring to ancient architecture-related literature,Chinese ancient architecture are classified according to historical periods and architectural forms.The Chinese Architecture-roof(CNA-roof)and Chinese Architecture-age(CNA-age)datasets were established.(2)To address the problems of incomplete feature extraction of architectural elements and difficulties in the recognition of similar architectural styles,we propose a Salient Region Suppression and Multi-scale Feature Fusion(SRSMFF)architectural style recognition method.First,the improved Resnet18 extracts the initial architectural features.Secondly,the Salient Region Suppression Module(SRSM)is designed.SRSM expands the perceptual range of architectural elements and enhances the key feature information through the interaction of suppression and salient regions.Then,Multi-scale Feature Fusion(MSFF)is proposed to fuse multi-resolution features using different pooling strategies.MSFF improves the different scale information of building images and enhances the spatial representation capability.Next,channel attention is used to quantify the importance of each channel and assign corresponding weights to enhance the important channel information.Finally,the large-margin softmax loss function is introduced.It maximizes the decision boundary distance of the feature embedding space and improves the performance of similar architectural style recognition.The experimental results show that our model achieves 64.44% and 80.21% accuracy on the 25-class and10-class public architectural style datasets.It achieves an accuracy of 88.21% on the dataset of ancient Chinese architectural styles.These results demonstrate the effectiveness of our method.(3)To solve the problems of unclear architectural texture features and insufficient utilization of contextual information,a hybrid model CVi T is proposed for architectural style recognition.First,the Primary Feature Perception(PFP)module is designed.This module uses an initial convolutional layer to segment the input image into multiple subregions,and uses an inverse residual network to perform feature enhancement on each sub-region.Thus,the low-level feature information of the building image is extracted.Secondly,the feature embedding encoder(FEE)is used to perform down-sampling and pixel attention operations on the feature map.It encodes the location information of features efficiently and improves the representation of architectural spatial features.Then,an improved E-Transformer structure is used.Efficient multi-headed self-attentiveness performs linear projection downscaling and pixel reorganization up-sampling on features.And the average pooling branch is introduced in the feedforward network.The structure enhances the texture features of the building facade and improves the global perception of building features.Finally,the modules are combined at different scales to form a multilevel structural model.The experimental results on the 25-class and 10-class public building style datasets show that our method achieves 67.36% and 82.64% accuracy.Moreover,our method achieves 95.63% accuracy on the self-built CNA-age dataset.These results confirm the effectiveness of our method.