Fashion Style Quantification and Clustering Network Based on Unsupervised Latent Space Representation Learning
Journal of Design Intelligence · 10.07.2026 · 1 (3) · 39-50
DOI: 10.64509/jdi.13.119 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Design Intelligence, online erschienen am 10.07.2026. Verfasst von Zhiyuan Yu und 2 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Amidst the rapid evolution of fashion e-commerce, efficiently matching massive inventories with personalized demands requires the objective quantification of aesthetic styles. However, traditional manual annotations are subjective and costly, while low-level visual features struggle with the "semantic gap" against high-level user preferences. To address these challenges, this study proposes LUCID (Latent Unsupervised Clustering with Interactive Display), an unsupervised latent space representation learning and clustering visualization network driven by vision-language priors. LUCID first utilizes a pre-trained Contrastive Language-Image Pre-training (CLIP) image encoder to extract high-dimensional semantic visual features. To overcome the "curse of dimensionality," it employs an auto-encoder integrated with a contrastive learning mechanism to compress these features into denoised, low-dimensional latent vectors. Subsequently, the K-means algorithm automatically clusters these representations, and Uniform Manifold Approximation and Projection (UMAP) generates an intuitive two-dimensional fashion style distribution map. This end-to-end workflow completely circumvents the reliance on manual labels, achieving effective feature dimensionality reduction and semantic refinement. Extensive experiments demonstrate that LUCID's unsupervised clustering aligns highly with human consensus, achieving ACC, NMI, and ARI scores of 84%, 75%, and 69% on the DeepFashion dataset, with strong generalization on Fashion-MNIST (68%, 69%, 56%). Ultimately, this research effectively bridges the aesthetic semantic gap, providing robust technical support for intelligent fashion analysis systems.
Zitieren
APA
Yu, Z., Guo, Z., & Hong, Y. (2026). Fashion Style Quantification and Clustering Network Based on Unsupervised Latent Space Representation Learning. Journal of Design Intelligence, 1(3), 39-50. https://doi.org/10.64509/jdi.13.119
Vancouver
Yu Z, Guo Z, Hong Y. Fashion Style Quantification and Clustering Network Based on Unsupervised Latent Space Representation Learning. Journal of Design Intelligence. 2026;1(3):39-50. doi:10.64509/jdi.13.119
BibTeX
@article{Yu2026,
author = {Yu, Zhiyuan and Guo, Ziyi and Hong, Yan},
title = {Fashion Style Quantification and Clustering Network Based on Unsupervised Latent Space Representation Learning},
journal = {Journal of Design Intelligence},
year = {2026},
volume = {1},
number = {3},
pages = {39--50},
doi = {10.64509/jdi.13.119}
}
EUVIMED