Research On Application Of Style Transfer And Conditional Generative Adversarial Networks In Urban Fabric Generation

Urban fabric plays an important role in urban design and architectural design,recording information such as building layout and spatial pattern,providing information for further urban space and architecture form design.However,urban fabric generation is complex,and it takes a lot of time to consider and need an appropriate methods.In addition,the traditional fabric generation methods can not achieve the fast,accurate and controllable generation goals.Deep learning models can be trained based on a large amount of data to learn the rules and logic in data,so as to generate real data.Therefore,this kind of model is more and more used in the field of generation.In this study,based on the principle of deep learning technology,different deep learning methods are used to generate urban fabric in order to achieve the fast,accurate and controllable generation goals.Style transfer and Conditional generative adversarial network are two mature generative models in deep learning.Style transfer can transfer the texture pattern style on image to target image and achieve the effect of overall coordination according to the reference style image.Conditional generative adversarial networks can generate the corresponding results according to different conditions.It is proved by experiments that these two kinds of models can quickly generate urban fabric,so as to assist in guiding design process.In this study,experiments were carried out on these two types of models to explore the differences in urban fabric generation among different models.In style transfer,Neural Style Transfer and Cycle GAN are used.Neural Style Transfer model can quickly generate urban fabric images,while Cycle GAN can train on a wider range of data sets to learn deep urban fabric formal logic.Both models are easy to use,and the generated image can be used to provide design inspiration for designers.In Conditional generative adversarial network,Pix2 pix model is mainly used,with summarizing the methods of data collection,processing and cleaning,and how to improve model performance,including model hyperparameter adjustment,model architecture adjustment,transfer learning enhancement.In addition,Gau GAN model is used to generate diverse outputs.Besides,an evaluation method based on deep learning models is proposed to evaluate the generation quality of different models.Finally,the well-trained Pix2 pix model is applied to the generation of urban fabric in different design scenarios.,and it is proved that this method can generate urban fabric quickly,accurately and controllably.Besides,a new human-computer interaction design mode is discussed in the study.The whole thesis contains about 65 000 words,162 pictures and charts.