A Fuzzy Kwan–Cai Neural Network for Determining Image Similarity and for the Face Recognition

Similarity is a crucial issue in Image Retrieval [1–4]. It is relevant both for unsupervised clustering [5, 6] and for supervised classification [7]. In this study, we aim to provide an effective method for learning image similarity. To reach this aim we start of from the Fuzzy Kwan–Cai Neural Network (FKCNN) [8] and turn it into a supervised method for learning similarity. Unlike the classical unsupervised FKCNN [8], in which a class is represented by a single output neuron, the supervised FKCNN has more output neurons that each designate a class. These give a better performance than their unsupervised counterparts as in the case of the classical unsupervised FKCNN, a class is represented by a single output neuron while the supervised FKCNN has more output neurons that each designate a class. This concept is similar with the idea of replacing the binary decision about the membership (nonmembership) of a pattern to a class, with the introduction of a membership degree, between 0 and 1. In order to evaluate the performance of our proposed neural network, it is compared with two baseline methods: Self-organizing Kohonen maps (SOKM) and k-Nearest Neighbors (k-NN). The feasibility of the presented methods for similarity learning has been successfully evaluated on the Visual Object Classes (VOC) database [9], that consists of 10102 images in 20 object classes. The concept of similarity is important not only in almost every scientific field but it has [10, 11] deep roots in philosophy and psychology. Our work however deals more with the measure of similarity in computer science domain (Information Retrieval to be more specific, that has focused on images, video, and to some extent audio). “Measuring image similarity is an important task for various multimedia applications.”1