Nonlinear Transformation of the Distance Function in the Nearest Neighbor Image Recognition

Abstract

Conventional image recognition methods usually include dividing the keypoint neighborhood (for local features) or the whole object (for global features) into a grid of blocks, computing the gradient magnitude and orientation at each image sample point and uniting the orientation histograms of all blocks into a single descriptor. The query image is recognized by matching its descriptors with the descriptors of reference images. The matching is usually done by summation of distances between descriptors of corresponding blocks. Unfortunately, such approach does not lead to a correct distance between vector of points (histograms of each block) if popular square of Euclidean distance is used as a discrimination. To calculate the correct discrimination, we propose to sum the square roots (or, more generally, appropriate nonlinear transformation) of distances between block histograms. Such approach is experimentally examined in a face recognition problem with FERET and AT&T datasets. The results support the statement that the proposed approach provides higher accuracy (up to 5.5%) than state-of-the-art methods not only for the Euclidean distance but for other popular similarity measures (L ₁, Kullback-Leibler, Jensen-Shannon, chi-squared and homogeneity-testing probabilistic neural network).