Kernel methods based approaches to image classification and retrieval

Abstract

Support vector machine (SVM) is a pattern classification model suitable for classification and annotation of images using non-vectorial type representations of images. Varying length patterns extracted from image data correspond to sets of local feature vectors. Kernels designed for varying length patterns are called as dynamic kernels. The talk presents the issues in designing the dynamic kernel based SVMS for classification and annotation of images. Different methods for designing the dynamic kernels are presented. An intermediate matching kernel (IMK) for a pair of varying length patterns is constructed by matching the pairs of local feature vectors selected using a set of virtual feature vectors. For patterns corresponding to sets of feature vectors, a Gaussian mixture model (GMM) is used as the set of virtual feature vectors. The GMM-based IMK is considered for image processing tasks such as image classification, matching and annotation in content-based image retrieval. The talk presents results of experimental studies on image classification, annotation and retrieval of images using the kernel methods.