On relevance feedback and similarity measure for image retrieval with synergetic neural nets

Abstract

In image retrieval, research issues relating to the design of a similarity function, which corresponds to human perception, remain open. Here we exploit a new interpretation of the control parameter, order vector, used in synergetic neural net (SNN) and use it as the basis of a similarity function for shape-based retrieval. More specifically, we have proven certain properties and theorems which give a formal basis for SNN based image retrieval. Based on the properties, an efficient affine invariant similarity measure has been developed for trademark images. Furthermore, we propose a self-attentive retrieval and relevance feedback mechanism for similarity measure refinement. Experiments also demonstrated that the proposed similarity measure is able to reflect the user's view of similarity through relevance feedback which in turn reinforces the retrieval ranking.