Generative Design of Outdoor Green Spaces Based on Generative Adversarial Networks

Abstract

Generative Adversarial Networks (GANs) possess a significant ability to generate novel images that adhere to specific guidelines across multiple domains. GAN-assisted generative design is a design method that can automatically generate design schemes without the constraints of human conditions. However, more research on complex objects with weak regularity, such as parks, is required. In this study, parks were selected as the research object, and we conducted our experiment as follows: (1) data preparation and collection; (2) pre-train the two neural network, then create the design layout generation system and the design plan generation system; (3) realize the data augmentation and enhanced hundred level dataset to thousand level dataset; (4) optimized training; (5) test the optimized training model. Experimental results show that (1) the machine learning model can acquire specific park layout patterns, quickly generate well-laid-out plan layout plans, and create innovative designs that differ from the human designer’s style within reasonable limits; (2) GAN-driven data augmentation methods can significantly improve the generative ability of algorithms, reduce generative pressure, and achieve better generative results; (3) pix2pix is prone to mode collapse, and CycleGAN has fixed rule errors in expressing certain design elements; and (4) GAN has the ability to mine design rules in the same way as humans.