Focusing the Normalised Information Distance on the Relevant Information Content for Image Similarity

This paper investigates the normalised information distance (NID) proposed by Bennet et al as an approach to measure the visual similarity (or dissimilarity) of images. Earlier studies (e.g.) suggest that compression-based approximations to the NID can yield dissimilarity measures that correlate well with visual comparisons. However, results also indicate that conventional feature-based dissimilarity measures often outperform those that are based on the NID. This paper proposes that a theoretical decomposition of the NID can help explain why the NID-based dissimilarity measures might not perform well compared to feature-based approaches. The theoretical decomposition considers the perceptually relevant and irrelevant information content for image similarity. We illustrate how the NID-based dissimilarity measures could be improved by discarding the irrelevant information, and applying the NID on only the relevant information.