Multi-resolution and Multi-bit Representation for Image Similarity Search

This paper explores the use of multi-bit quantisation of image features for similarity-based image retrieval. Our work builds on multi-resolution image similarity search algorithms which utilise one-bit representation of the largest magnitude wavelet coefficients. Given a query, images are ranked based on the number of quantised coefficients they have in common with the query. We explore the benefits of a finer-level quantisation (specifically with two bits) and one control parameter that can be chosen optimally based on the probability density of the wavelet coefficients. We show that this extension leads to significant performance improvements.