A Comparative analysis of similarity-based retrieval for SVG logos

Abstract

Brand logos play a major role in modern digital branding, enabling companies to maintain a consistent and recognizable identity. Measuring the similarity between brand logos is essential for tasks such as logo verification, duplication detection, and visual consistency checks. Unlike recognition tasks that aim to identify or classify logos, similarity assessment focuses on quantifying how closely two visual instances resemble each other. While extensive research exists for raster image similarity, assessing similarity in simple 2D vector graphics such as SVG logos remains relatively underexplored. Vector graphics define an image as a set of points, lines, and curves to create scalable, resolution-independent visuals. The study compares algorithmic approaches, including Hausdorff distance, Procrustes distance, Earth Mover’s distance, and centroid distance and angle feature matching, alongside deep learning methods such as DeepSVG. Five methods were tested on a set of query images with varying levels of impairments against a brand visual database. Results show that both Procrustes Analysis and DeepSVG deliver strong matching performance, even when the query images are significantly distorted. Interestingly, the top algorithmic and deep learning methods achieved similar performance, suggesting that deep learning does not necessarily outperform traditional algorithmic approaches for measuring similarity in SVG logos.