Artwork Style Clustering with Caption-Derived Latent Features

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing artwork classification methods primarily focus on art movements and lack effective unsupervised clustering at the style level, failing to capture the diverse and evolving artistic styles of individual artists.

Innovation Solution

A method and system for style-based clustering of artworks using natural language style annotations, involving generating style-based artwork representations through captions or style concepts, followed by latent feature extraction with autoencoders and deep embedded clustering.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If art movement-based classification methods are used, then artworks can be categorized into broad categories, but fine-grained style-level clustering is not achieved

Engineering Contradiction:
Improvestyle-level classification precisionVSAvoidclustering system complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the artwork classification task into multiple processing stages: (1) feature extraction from artwork images, (2) generation of style-based captions using GPT-4, (3) extraction of style keywords from captions, and (4) clustering based on combined visual and textual features. This segmentation enables fine-grained style-level classification while managing system complexity through modular processing steps.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces style-based captions generated by GPT-4 as an intermediary between visual artwork features and clustering algorithms. These natural language captions serve as a bridge that translates visual style characteristics into structured textual representations, enabling more precise style-level clustering without directly increasing the complexity of the clustering mechanism itself.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Reliability

If unsupervised clustering methods are used, then labeled data requirements are reduced, but clustering accuracy at style level is insufficient

Engineering Contradiction:
Improveclustering accuracyVSAvoidlabeled data availability
Core Design Contradiction:
ReliabilityVSQuantity of substance

Solution Approach 1:

The patent performs preliminary action by generating style-based captions using GPT-4 before the clustering process. These pre-generated natural language descriptions capture style characteristics in advance, providing rich semantic information that guides the subsequent unsupervised clustering. This preliminary textual annotation enhances clustering accuracy without requiring manual labeled data for training.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent combines multiple feature types into a composite representation for clustering: visual features from artwork images, textual features from GPT-4 generated captions, and extracted style keywords. This composite feature representation integrates diverse information sources, enabling accurate style-level clustering while maintaining an unsupervised approach that does not rely on manually labeled training data.

Inventive Principle:
Principle #40Composite materials

Data Source

PatentEP4672041A1Method and system for style based clustering of artworks with natural language style annotations
Publication Date: 2025.12.31 TATA CONSULTANCY SERVICES LTD
  • EP4672041A1 patent drawingFigure 1
  • EP4672041A1 patent drawingFigure 2
  • EP4672041A1 patent drawingFigure 3

AI summary

This disclosure relates generally to a system and method for style-based clustering of artworks with natural language style annotations. The conventional methods generate generic image feature representations derived from deep neural networks and do not specifically deal with the artistic style. The present disclosure, generates style-based artwork representations based on caption with style-based keywords and style concept annotations by leveraging image captioning model, vision language model and text encoder. Further style-based latent feature representations are generated from the style-based artwork representations for performing unsupervised clustering. The clustering of style-based latent feature representations is done based on deep embedded clustering using dynamic or static initialization of clusters. The present disclosure helps in discovering finer-grained style concepts within a corpus of artwork in an unsupervised manner. It also helps explore and create the art style evolution-based narratives and curative practices.