Improved image-to-class distance based on feature selection

Abstract

We propose a novel, Naive Bayes Nearest Neighbor (NBNN) classifier considering discriminative features, NBNN-DF, for improving both classification accuracy and runtime performance. Unlike the original NBNN method, we define discriminative features among features extracted from training and query images, and perform the NBNN only with those discriminative features. As a result, we can ignore nondiscriminatory features extracted from background clutters or irrelevant objects from a class type of each image. To define discriminative features we measure a discriminative power for each feature based on a ratio of posterior probability that the feature is located in its positive class to that in its negative class. While it is easy to measure discriminative power for features extracted from training images, we face the chick-and-egg problem for query images, whose class type is unknown. To address this problem we hypothesize potential class types of a query image and perform the NBNN with discriminative features under its hypothetical classes, while considering a confidence level of each hypothesis. We have tested our method on the Caltech101 dataset, and compared it against other state-of-the-art techniques. We found that our method, NBNN-DF, achieves 32%, 27%, 17% relative accuracy improvement over the standard NBNN, the local NBNN , and NBNN with max-margin optimized weights under the image-to-class distance, respectively. Our technique achieves this mprovement while improving the overall runtime performance by using a smaller number of features.