Abstract:
"Traditional floor planning processes are often marred by complexity, lack of customization,
and the potential for costly mistakes. Current floor plan recommendations typically offer
minimal requirements and require more user-centric designs. They often need to align with
new floor plan trends and use similar data for all recommendations, disconnecting the architect
and the user. Users may need help to express their needs clearly, and architects may find it
challenging to design plans that meet those requirements, leading to significant time wastage
for both parties.
To solve this problem, the author extensively studied existing challenges in floor plan design
and implemented Deep Learning and GAN technologies into the software's architecture. The
system features a user-friendly interface that allows users to input their preferences directly.
Based on these inputs, the GAN-driven recommendation engine then refines and personalizes
floor plans. The technical approach involved designing a multi-layer Convolutional Neural
Network within the GAN framework to accurately model and predict spatial arrangements,
ensuring the generated plans are functional and aesthetically pleasing.
Based on the evaluation results from the Tell2Design dataset with CGAN, this model achieved
an Inception Score of 1.00 ± 2.0 and a Frechet Inception Distance of approximately 1600, even
though it was trained with a small dataset. These metrics indicate a high level of diversity and
quality in the generated floor plans despite the challenges posed by the dataset size. The IS
score suggests good perceptual quality, while the FID score reflects reasonable similarity to
real floor plans in the dataset. This performance underscores the effectiveness and reliability
of the CGAN-based approach in generating accurate and user-preferred floor plans, validating
its potential to enhance architectural and interior design practices through advanced AI
technologies."
