EFFICIENT BROWSING IN IMAGE DATABASES USING A HIERARCHY OF KERNEL PCA SUBSPACES

We present a novel approach for designing the search functionality in large unlabeled image databases. It combines Relevance Feedback, Hierarchical Browsing and Kernel PCA, uses a Mixture-of-Gaussian to model feature space distributions and different visualization techniques of high dimensional feature spaces. Given an image database, finding a specific single or set of pictures is achieved by assisting the user to find an as-short-as-possible browsing path through the database. Our system relies on describing each picture with an appropriate feature vector that results from applying Kernel PCA to image and textual based similarity matrices. We solve the page-zero-problem by presenting the centroids of a hierarchical clustering in feature space as initial suggestions. The user can then steer the search by selecting positive and negative examples which define a Mixture-of-Gaussian density in the parameter space. New suggestions are drawn according to this density and the user is thus directed to the desired image category. A user study proved our system to be practical and beneficial for category search tasks.