Semantic similarity between images: A novel approach based on a complex network of free word associations

Abstract

Several measures exist to describe similarities between digital contents, especially for what concerns images. Nevertheless, distances based on low-level visual features embedded in a multidimensional linear space are hardly suitable for capturing semantic similarities and recently novel techniques have been introduced making use of hierarchical knowledge bases. While being successfully exploited in specific contexts, the human perception of similarity cannot be easily encoded in such rigid structures. In this paper we propose to represent a knowledge base of semantic concepts as a complex network whose topology arises from free conceptual associations and is markedly different from a hierarchical structure. Images are anchored to relevant semantic concepts through an annotation process and similarity is computed following the related paths in the complex network. We finally show how this definition of semantic similarity is not necessarily restricted to images, but can be extended to compute distances between different types of sensorial information such as pictures and sounds, modeling the human ability to realize synaesthesias.