Clustering- and Transformer-Based Networks for the Style Analysis of Logo Images

Area of Science:

• Computer Vision
• Machine Learning
• Graphic Design

Background:

• Designers manually search for logo references, a time-consuming process.
• Existing methods lack efficient automatic style grouping for large logo databases.

Purpose of the Study:

• To develop an automated method for clustering logo styles.
• To improve the efficiency of logo material collection and style analysis for designers.

Main Methods:

• Utilized a visual transformer model for feature extraction from logo images.
• Applied K-Means clustering with iterative refinement using pseudo-labels.
• Developed a deep transformer-based cluster (DTCluster) method.

Main Results:

• Achieved high clustering performance on the JN Logo dataset (DBI: 0.904, DI: 0.189).
• DTCluster outperformed other K-Means and clustering algorithms.
• Identified six distinct logo styles with five semantic descriptions through subjective analysis.

Conclusions:

• The DTCluster method offers an effective solution for automatic logo style grouping.
• This approach enhances the efficiency of design research and material discovery.
• Provides a semantically rich categorization of logo styles.