In an article published in the journal Scientific Reports, researchers from China demonstrated how generative adversarial networks (GANs) can enhance design efficiency and creativity in spatial computing, a field encompassing architecture, interior design, and urban planning.
Study: GANs Revolutionize Spatial Computing in Design Fields. Image credit: Golden Dayz/Shutterstock
Additionally, they proposed a method for icon generation based on GANs and introduced the concept of interactive design and requirement condition features to construct a system with generation and optimization ability. Moreover, they discussed the potential application of the presented technique in various spatial design fields.
Background
Spatial computing is an important branch of the design field, which involves creating and manipulating spatial data in various forms, such as two-dimensional (2D), three-dimensional (3D), and four-dimensional (4D). It can be applied in fields, such as architecture, interior design, urban planning, and landscape design. Spatial design must address functional requirements and consider emotions, user experience, and aesthetics.
However, conventional design methodologies are constrained by experience accumulation and manual operations, making it challenging to effectively implement difficult design tasks. The emergence of GANs, a powerful deep-learning technique, presents possibilities for addressing this issue.
GANs are a network structure capable of developing realistic data by allowing two neural networks to compete with each other and copy, capture, and analyze the differences in a dataset. It gained excellent success in several tasks, such as speech synthesis and image generation, and has shown potential in multiple fields.
About the Research
In the present paper, the authors explored the application potential and optimization problems of GANs in spatial computing and designed an interactive design generation and optimization method based on multiple features and conditional constraints. They constructed a basic architecture by establishing icons, common data forms with high information density, and application scenarios in spatial design.
Icons could be used for mobile applications, website design, brand identification, and other purposes. Additionally, the study introduced interactive design and essential condition features to tailor to the specificity of spatial design and create a system with optimization and generation capabilities, aiming to meet diverse spatial design needs.
The researchers proposed an icon generation method based on GANs, leveraging the generator and discriminator components. The generator aims to produce icons similar to real icons, while the discriminator differentiates the generated icons and real icons. With continuous adversarial learning, the generator steadily enhances the quality of the generated icons, making them more realistic.
The paper also introduced conditional input, which are specified features or parameters, such as color, shape, size, and style, to customize the generated icons. The generator needs to combine conditional input with random noise to generate icons that match the conditions. The discriminator needs to determine whether the generated icon meets the conditions.
The authors further optimized the structure of conditional features by integrating multi-feature recognition modules into the discriminator, effectively capturing the important features of icons, such as color and shape. They used an improved auxiliary classifier GAN (ACGAN) model, which can generate images that meet the conditional features, while the discriminator can verify whether the images meet the conditions.
Moreover, they used a Wasserstein GAN (WGAN) model, which can measure the distance among distributions using Wasserstein distance and improve the stability and generation efficiency of the model.
The researchers validated the new method on two different icon datasets, namely Icons-50 and LLD-Icon, for the generation of black and white and color icons. Moreover, they compared the performance of the proposed method with other GAN models, such as GAN, WGAN, and ACGAN, using two evaluation indicators, namely inception score (IS) and Fréchet inception distance (FID). IS measures the authenticity and diversity of the generated samples, while FID measures the similarity between the distribution of the generated samples and the distribution of the real samples.
Research Findings
The outcomes showed that the newly developed method could generate high-quality icons that meet the design requirements and demonstrate the potential of GANs in spatial design. It could produce icons with more prominent shape features and more fine-grained color control and can effectively maintain the diversity and innovation of the generated icons while meeting the conditional features. Moreover, it achieved higher IS values and lower FID values than other GAN models, indicating that the proposed method can generate more realistic and diverse icons.
The proposed method can be applied to various spatial design fields, such as architecture, interior design, urban planning, and landscape design, to achieve more efficient and innovative design processes. It can generate icons suitable for mobile applications, website design, brand identification, and other purposes, while also allowing customization of the icons based on specific design requirements, such as color, shape, size, and style. Additionally, it has the potential to be extended to generate other data forms, including 2D, 3D, and 4D spatial data, facilitating the creation and manipulation of spatial data in diverse formats.
Conclusion
In summary, the researchers comprehensively explored the optimization problems of GANs in spatial computing and proposed a novel technique for interactive design generation and optimization. Their method outperformed other GAN models in authenticity and diversity metrics, indicating its potential for diverse spatial design applications.
Furthermore, the researchers recommended directions for future work, such as how to better determine and design conditional features, how to select and fuse multiple features, and how to explore more GAN structures, such as variational autoencoder GANs, to further improve the performance of the generation model.