【作者】 王宁；

【导师】 王桥；

【作者基本信息】 东南大学 ， 信号与信息处理， 2021， 硕士

【摘要】 城市意象主要研究市民对城市的认知和感受,是体现城市文化特征和城市风貌的重要方式。我国经济社会进入新的常态化发展阶段,城市规划者逐渐认识到城市意象研究对城市建设的重要性,城市空间品质的提升和城市特色的塑造也日益得到重视。面对城市风貌“千城一面”的问题,在多源大数据和人工智能技术的背景下,对城市街景图片的识别和分析逐渐成为研究城市空间意象的主要研究手段。本文基于街景图片数据集,采用和改进现有的机器学习和深度学习技术,实现了对街景建筑风貌的识别和分类,并对街景要素进行了分析和生成。进而,这些结果可用于探索影响城市街道空间意象感知的因素,协助建设合理舒适的城市街道空间。首先,本文在有标签的街景数据集的基础上进行了城市意象的初步研究。基于Res Net152模型对南京的街景功能进行分类与识别;采用公开的cityscapes数据集训练Deep Lab V3+模型,计算出了南京中心城区各街景要素的占比等特征,进而对中心城区的空间舒适度进行分析。其次,本文提出一种基于生成对抗网络的街景建筑功能聚类的方法。通过分析Info GAN的模型特点,将Info GAN模型应用于图像的无监督聚类。通过对街景图片采用了掩膜和栅格处理,提取出街景图片中的建筑元素,进而使用Info GAN模型对图片进行聚类,得到建筑学领域有意义的聚类结果。通过实验发现,在街景聚类效果上,本文提出的无监督聚类方法与已有方法相比在聚类指标上获得较优的结果,同时有利于城市意象研究中街景图片的快速标定以及大规模的建筑风貌识别。最后,本文提出一种基于变分自编码器和Info GAN的街景图片生成模型。通过将变分自编码器与Info GAN模型结合起来,不仅可以生成街景要素风格变化的街景图片,而且可以生成指定图片的街景要素变化图,为后续城市意象中的街景要素谱系研究提供了技术手段。更多还原

【Abstract】 City image is the study of people’s perceptions and feelings of the city,and it is an important way to reflect the city’s regional culture and characteristics.With China’s economy and society entering a new normal development stage,people gradually understand the importance of urban image research to the city.Under the background of multi-source big data and artificial intelligence technology,the research on urban street scene image recognition and street scene element analysis has become the main research goal of urban street space image.This thesis aims to recognize and classify the functions of large-scale streetscape buildings,and analyze the streetscape elements by using machine learning and deep learning technology.Moreover,the generation of street view elements is experimented and explored by constructing a model based on variational self encoder and generation countermeasure network.Firstly,this thesis conducts a preliminary study of city image on the basis of a labeled street view data set.Based on Res Net152 model,the street view function of Nanjing is classified and identified.The Deep Lab V3 plus model is trained with the cityscapes dataset.Then this thesis calculates the proportion of street view elements in the central urban area of Nanjing,and analyzes the spatial comfort of the central urban area.Secondly,this thesis proposes a method of street view building function clustering based on generative adversarial networks.By analyzing the characteristics of Info GAN model,this thesis firstly applies the Info GAN model to unsupervised clustering of images.Experiments show that,compared with the existing clustering methods,the unsupervised clustering method proposed in this thesis obtains better results in the clustering index.it is conducive to the rapid calibration of street view images and large-scale architectural style recognition in the study of city image.Finally,this thesis proposes a street view image generation model based on variational autoencoder and Info GAN.By combining the variational autoencoder with the Info GAN model,not only the street view images with different styles of street view elements can be generated,but also the street view element change maps of the specified images can be generated,which provides contents for the subsequent study of the street view element pedigree in city image.更多还原