Filtering near-duplicate samples

I'm playing around with SIFT descriptors for images in an attempt to use them in a 'bag-of-words' style approach to image classification.

I've cobbled together an assortment of random images, and after extracting features I have about 1.15 million assorted feature descriptors.

I've been mucking around with using k-means as a first blush at binning the features, but in the same way that it helps to pre-process words (removing pluralizations, prefixes, suffixes, etc), I want a reasonably concise way to consolidate near-duplicate features so as to avoid skewing clusters.

I'm looking for suggestions, whether that be a methodology, paper recommendation, or a smack to the side of the head.