A Deep Clustering Method with Generation Model and Metric Learning for Content-Based Retrieval Focused on Art Style

DOI

Bibliographic Information

Other Title
  • 画風に基づく作品検索に向けた生成モデルと距離学習に基づく深層クラスタリング手法

Abstract

This paper proposes a clustering model using deep metric learning for a content-based retrieval system focusing on the illustration art style. Unlike the conventional methods, which train models using heuristic features or probability distributions of specific labels, the proposed method jointly optimizes the latent space of the Variational Autoencoder with Triplet loss. Objective evaluation with unseen data revealed that the proposed method achieved an NMI score of 51.71%, 10.61% higher than the conventional method. Furthermore, when used in a similarity retrieval system, the proposed method was found to recommend images with perceptual similarities in the query about the art style.

Journal

Keywords

Details 詳細情報について

Report a problem

Back to top