A self-sparse generative adversarial network for autonomous early-stage design of architectural sketches

This study develops an autonomous design method for architectural shape sketches by a novel self-sparse generative adversarial network (self-sparse GAN), thereby overcoming the problems regarding excessive reliance on sufficient aesthetic knowledge and excessive time consumption in traditional human design. First, a new architectural shape dataset denoted “Sketch” is built by using the eXtended difference-of-Gaussians operator. Second, a self-adaptive sparse transform module (SASTM) is designed following each deconvolution layer of the proposed self-sparse GAN to utilize the sparsity of sketch images by the sparsity decomposition and feature-map recombination. Third, the Frechet inception distance (FID) is adopted to evaluate the quality of the generated sketches by comparing the distribution of the real and generated datasets. Finally, two common image generation approaches, Wasserstein GAN with gradient penalty and self-attention GAN, are compared with the proposed self-sparse GAN, and results show the proposed method achieves the best performance with a relative decrease in the FID score of 11.87%. The proposed autonomous design method can give tens of thousands of sketches for a class of buildings in a few seconds using the trained network, which can help architects to choose the architectural form and/or inspire architects to consider unique schemes in the early stages of design.