K-Nearest Neighbors (k-NN) search is the standard technique for today’s Content-Based Image Retrieval (CBIR) systems.   In this model, semantically similar images are generally assumed to be clustered together in a single neighborhood in the high-dimensional feature space.   Unfortunately semantically similar images with different appearances are often spread in distinct neighborhoods, which might be far away from each other in the feature space.   Hence, the confinement of the search results to a single neighborhood is the latent reason of the low recall rate of typical nearest neighbor techniques.   In our Query Decomposition system, the user query may be decomposed into multiple subqueries based on the user’s relevance feedback to cover multiple image clusters which contain semantically similar images.   The retrieval results are the k most similar images from multiple disjoint relevant clusters.

Project Description