A Deep Learning-Based Approach to Generating Comprehensive Building Façades for Low-Rise Housing

• Published in: Sustainability
• Main Authors: Da Wan, Runqi Zhao, Sheng Zhang, Hui Liu, Lian Guo, Pengbo Li, Lei Ding
• Format: Article in Journal/Newspaper
• Language: English
• Published: MDPI AG 2023
• Subjects: deep learning; generative adversarial network (GAN); façade generation; Pix2Pix; generator comparison; Environmental effects of industries and plants; TD194-195; Renewable energy sources; TJ807-830; Environmental sciences; GE1-350; Attu
• Online Access: https://doi.org/10.3390/su15031816; https://doaj.org/article/f4f9c1787d9e4b43810c3db98f3d2e75

In recent years, as machine learning has been widely studied in the field of architecture, scholars have demonstrated that computers can be used to learn the graphical features of building façade generation. However, existing deep learning in façade generation has yet to generate only a single façade, without comprehensive generation of five façades including the roof. Moreover, most of the existing literature has utilized the Pix2Pix algorithm for façade generation experiments, failing to attempt to replace the original generator in Pix2Pix with a different generator for experiments. This study addresses the above issues by collecting and filtering entries from the international Solar Decathlon (SD competition) to obtain a data set. Subsequently, a low-rise residential building façade generation model based on the Pix2Pix neural network was constructed for training and testing. At the same time, the original U-net generator in Pix2Pix was replaced with three different generators, U-net++, HRNet and AttU-net, for training and test results were obtained. The results were evaluated from both subjective and objective aspects and it was found that the AttU-net generative network showed the best comprehensive generation performance for such façades. HRNet is acceptable if there is a need for fast training and generation