360 파노라마 실내공간 사진 기반 지능형 레퍼런스 데이터베이스 구축

Abstract

(Background and Purpose) Design communication plays an important role in helping designers share design information and make decisions, because the process of designing a space involves collaboration between several people or groups. Recently, with the advent of virtual reality, many observers have been able to experience a desired space from anywhere beyond the passive experience of space. This study focused on the use of intelligent technology to construct essential and important data in the process of building visual information for indoor spatial design. By analyzing the commonly used spatial photo-based approach to media and demonstrating the prototype of an indoor spatial reference image database through the 360° panoramic image-based approach, we intend to demonstrate its potential. (Method) In this study, we solved the problem using deep learning technology that automatically classified interior design images extracted from high-quality panoramic images. We will present an approach that simplifies the process of building reference image archives. The research methods and procedures were as follows: 1) 360° panoramic image preparation, 2) automatic reference image extraction for interior space design, 3) automatic classification and recognition of features appearing in each image to diversify image datasets, and 4) recognition information and storage of images. (Results) Based on the above-mentioned method, an executable prototype was produced and its utility was verified. The source of the 360° panoramic images utilized not only high-definition photos taken using a 360° camera, but also realistically rendered virtual images created with CAD or BIM. There is a parametric method and a preset cube map method for dividing a given indoor 360° panoramic image. The first method divides a given image into grids with specific viewing angles and sizes. However, the cube map method generates only six images with different fields of view and a fixed image size. The trained deep learning-based CNN model recognizes and labels design-related information, such as space usage and design style, on each segmented image. Segmented images are automatically stored in the database along with labeled data. (Conclusion) This thesis is expected to contribute to archiving and utilizing a vast amount of qualitative 360° panoramic image data in the field of architectural design. In particular, we have created a standardized way to keep interior design references that people treat arbitrarily. In future research, we intend to expand the scope of the recognition information (e.g., design-related objects and design styles) from a given interior design image and develop a method to utilize stored data in the design process.