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

Abstract

In the design field, designers need to investigate and collect logo materials before designing logos and search a large number of design materials on well-known logo websites to find logos with similar styles as reference images. However, manual work is time-consuming and labor-intensive. To solve this problem, we propose a clustering method that uses K-Means clustering and visual transformer model to group the styles of the logo database. Specifically, we use the visual transformer model as a feature extractor to convert logo images into feature vectors and perform K-Means clustering, use the clustering results as pseudo-labels to further train the feature extractor, and continue to iterate the above process to finally obtain reliable clustering results. We validate our approach by creating the logo image dataset JN Logo, a proposed database for image quality and style attributes, containing 14922 logo design images. Our proposed deep transformer-based cluster (DTCluster) automatic style grouping method is used in JN Logo; the DBI reaches 0.904, and the DI reaches 0.189, which are better than those of other K-Means clustering methods and other clustering algorithms. We perform a subjective analysis of five features of the clustering results to obtain a semantic description of the clusters. Finally, we provide six styles and five semantic descriptions for the logo database.