Architectural Visualization of Knowledge Spaces: An Image-Based Study of Contemporary Library Design and Information Flow

Abstract

The application of artificial intelligence in architectural studies has mostly focused on the field of geometric reconstruction and generative design, whereas the semantic explanation of knowledge distribution in the interior spaces has not been studied in detail. This paper will suggest a computing model to explain how spatial arrangements of objects in the architectural environment encode knowledge. With the MIT Indoor-67 dataset (15,620 images and 67 categories), a ResNet50 deep learning model was trained on scene classification, with a Top-1 and Top-5 accuracy of 77.01 and 94.6, respectively, on 1,340 test samples. The matrices of scene-object frequency were constructed with the help of object-level annotations, and a Knowledge Density Index (KDI) was proposed to measure the knowledge-related object concentration. The spaces with high knowledge density (bookstores: KDI = 0.609 and libraries: 0.484) were significantly greater than the domestic spaces (bedrooms: 0.049). The similarity modeling based on TF-IDF weights produced a Knowledge Flow Network with high semantic relationships such as bookstore-library (0.8847 similarity). Object-centric model reasoning was validated by the Grad-CAM visualization. The findings indicate that the spatial knowledge distribution can be computedally measured, structured and can be interpreted using a combination of deep learning and semantic modeling methods. The framework leads to the development of AI-supported interpretive spatial intelligence in architectural image analysis.