Nearest Neighbor Searching in Image Databases

iii Abstract A frequently encountered type of query in image database systems is to nd the k most similar images to a query image with respect to its feature. Processing such queries requires substantially diierent search algorithms than those for the normal k nearest neighbor problem: dimensionality of the feature may be very high and similarity measure may not be as simple as a Euclidean distance. For this problem, low-dimensional transformation techniques were proposed, which map high dimensional feature values into a lower dimensional space and evaluate the original similarity metric only for those within some bound in lower dimension. This reduces the number of the high dimensional matchings. However, to guarantee no false dismissals, the existing methods use a maximal search space, which may be too wider than is practically needed. We propose an incremental ltering algorithm that can reduce the search space. The proposed algorithm rst chooses a minimal search space and incremetally expands it until the target number of results is obtained. This algorithm also guarantees no false dismissals and lower-bounds the existing methods with respect to the number of evaluations of the expensive measure. Abstract iii 1 Introduction 1