CGAN-Assisted Renovation of the Styles and Features of Street Facades—A Case Study of the Wuyi Area in Fujian, China

Lei Zhang, Liang Zheng, Yile Chen, Lei Huang, Shihui ZhouSustainability 2022 · DOI

TL;DR

This paper introduces a novel approach using Conditional Generative Adversarial Networks (CGANs) to automatically generate architectural facade designs that adhere to specific regional styles, addressing the challenge of maintaining aesthetic consistency in traditional villages and towns undergoing development. The method demonstrates practical applications in renovation, new building design, and even product design, offering a significant efficiency boost for architects and local governments.

What Problem Does This Paper Address?

Rapid urbanization and infrastructure development often clash with the preservation of traditional architectural styles in villages and towns. This creates a significant challenge for local governments and architects tasked with balancing modernization with the protection of cultural heritage. Specifically, achieving a unified and consistent architectural style across numerous buildings, especially in areas like the Wuyi Mountains (a UNESCO World Heritage site), is a time-consuming and labor-intensive process. The paper aims to bridge this gap by developing an efficient, machine-learning-assisted method for generating building facades that respect and replicate regional architectural characteristics, thereby enhancing design efficiency and promoting stylistic consistency.

Methods

The core of this research lies in the application of Conditional Generative Adversarial Networks (CGANs) for architectural facade style transfer. The methodology involves a five-step process:

  1. Data Capturing: This involves field trips, interviews, and collecting visual data of existing buildings that embody the target architectural style (Wuyi style in this case).
  2. Data Processing: Raw building images are processed into three distinct representations:
  3. Building Exterior Profile (BEP): The basic contour of the building facade.
  4. Functional Segmentation Layout (FSL): A simplified representation highlighting key architectural elements like doors, windows, walls, guardrails, and roofs using distinct colors. This allows the machine to "understand" the functional components.
  5. Building Elevation (BE): A cleaned and simplified representation of the final architectural rendering.
  6. Model Training: Two CGAN models are trained:
  7. Model 1: Trains to generate FSL from BEP.
  8. Model 2: Trains to generate BE from FSL. This two-stage approach allows for a more controlled and effective generation process.

  9. Model Evaluation: The trained models are assessed based on loss values (tracking the generator and discriminator's performance) and visual inspection of generated images to ensure accuracy and avoid artifacts.

  10. Data Application: The trained models are then applied to various design tasks, including new building design, facade renovation, and product design.

The novelty lies in the specific application of CGANs to architectural facade generation with a focus on style transfer and the structured processing of building data into BEP, FSL, and BE for a more robust training pipeline.

Datasets

The study utilizes a dataset of 164 representative facade images of Wuyi-style buildings collected from tourist areas in the Wuyi Mountains region. These images were uniformly scaled to 512x512 pixels. For training, these images were further processed into three categories: Building Exterior Profile (BEP), Functional Segmentation Layout (FSL), and Building Elevation (BE), resulting in a total of 492 images (164 for each category). The FSL samples used specific color mappings for different architectural elements (e.g., red for doors, green for windows).

Key Insights and Ideas

  • CGANs as a powerful tool for architectural style preservation: The research demonstrates that CGANs can effectively learn and replicate complex regional architectural styles, offering a scalable solution for maintaining aesthetic consistency in development projects.
  • Structured data representation is crucial: Breaking down building facades into BEP, FSL, and BE allows for a more granular and effective training process, enabling the CGAN to learn both structural outlines and functional elements.
  • Two-stage generation improves results: Training separate models for BEP-to-FSL and FSL-to-BE generation leads to more accurate and detailed final building elevations.
  • Beyond buildings: CGANs for cultural product design: The study highlights the versatility of the CGAN approach by successfully applying it to the graphic design of tourist products, suggesting a broader potential for promoting regional culture.
  • Efficiency gains are substantial: The CGAN method significantly reduces the time required for generating facade designs, transforming a manual, time-consuming process into a near-instantaneous one for batch applications.

Key Results

  • The trained CGAN models (Model 1 and Model 2) demonstrated the ability to generate building facades with a high degree of accuracy, closely resembling real images after sufficient training generations (200 generations).
  • Model 2, which generated Building Elevations from Functional Segmentation Layouts, showed particularly strong performance, capable of generating doors, windows, railings, roofs, and even decorative elements that reflect the local style.
  • In application tests, the CGAN was able to generate reasonable results for new building designs, facade decorations, and unfinished building restorations, with the generated facades showing a good degree of adherence to the regional style.
  • For product design, while road signs and house numbers with text showed mixed results (some blurring), generated designs for fans and cups were impressive and could be used almost immediately, indicating strong potential for cultural product branding.

Limitations and Future Work

  • Limited Dataset Size: The authors acknowledge that the generated effect is "still insufficient" due to the limited number of samples in the CGAN model.
  • Rudimentary Image Processing: CGANs are limited in their ability to fully convey the nuanced cultural qualities and "local spirit" of a region solely through image processing.
  • Inability to Replace Human Expertise: The CGAN cannot substitute human labor for highly expressive architectural cultural projects.
  • Future Work: The authors suggest increasing the training dataset size and optimizing the CGAN model to overcome current limitations. Further research could explore more sophisticated methods for capturing cultural nuances and expanding the application scope of the CGAN model in architectural design and related fields. The study also points to the need for thorough examination of building qualities and site conditions in practical applications, as the generated results are currently best suited for theoretical demonstration.

Summary updated

Abstract

With the development of society and the economy, the unified planning of architectural styles has become a significant problem in the balance between urban expansion and the protection of traditional buildings in villages and towns. This also allows people to re-examine the appearance of and quality of life, experienced by those in traditional village buildings. This research employs a conditional generative adversarial network (CGAN) to develop a generative technique for designing building facades in villages and cities. The provided results can be used to develop schemes and as design references for building facade design, enhancing the design efficiency of building facades. Simultaneously, we utilized this model for the rehabilitation of building facades in villages and towns, as well as in the visual design of rural tourism products, demonstrating its practical usefulness and design-related potential. We took villages and towns in the Wuyishan area of China as an example and carried out model training, image generation, and a comparison of the derivation results of different assumed buildings and product contours. The research shows that: (1) CGAN can be used to produce and supply reference schemes for conventional civil construction facade design in rural and urban areas. (2) In terms of adaptability, CGAN may develop architectural facade design schemes with a reference value for the hypothetical experimental building facades, and it can play a role in other design domains, as well. (3) The construction of this method is not only applicable to villages and towns in the World Heritage es Cities Programme, but can be further promoted and used in the future for cities and villages that have a demand for architectural style consistency.
BibTeX
@inproceedings{zhang2022cgan,
  author = {Lei Zhang and Liang Zheng and Yile Chen and Lei Huang and Shihui Zhou},
  title = {{CGAN-Assisted Renovation of the Styles and Features of Street Facades—A Case Study of the Wuyi Area in Fujian, China}},
  booktitle = {Sustainability},
  year = {2022},
  doi = {10.3390/su142416575},
}
Figure