House style recognition using deep convolutional neural network: Fid-Bau Portal

House style recognition using deep convolutional neural network

Yi, Yun Kyu / Zhang, Yahan / Myung, Junyoung

Abstract The recent development in deep learning has opened up a new era of possibilities, once difficult to achieve with conventional methods, to revolutionize image recognition, speech recognition, and natural language processing. Specifically, image recognition has been widely applied in various areas such as face recognition, object identification for security, and other purposes. Although it is rarely applied to discover new methods for use in architecture, image recognition has great potential in architectural design. For example, it can be used to identify the preference of the client and to design a building that satisfies a client's aesthetic preference. One of the major hurdles of utilizing image recognition in architecture is the architectural styles based on culture, location, and time. For that reason, it is difficult to identify an architectural style by non-trained clients and sometimes certain buildings are composed of different styles that are difficult to identify by experts as one style. This paper explores the possibility of using state-of-the-art image recognition algorithms in house style recognition to find out its limitations and possibilities. Moreover, the paper adopted a convolutional neural network model for classifying house styles in the US. Although the final accuracy is not high due to the lack of image datasets, the trained model performed reasonable predictions with a limited test set. The results show the importance of properly defining style for image recognition to improve its accuracy.

Highlights • Use Deep learning for single house style • Categorized single house styles into 17 styles. • Use Deep Convolutional Neural Network for house style image recognition. • Able to predict different single house styles from images.