Non-negative Sparse Coding Using Independent Multi-Codebooks for Near-Duplicate Image Detection

Abstract

In this paper, we propose an efficient approach for detecting near-duplicate images and make three contributions as follows. First, for each sub-region of spatial pyramid, we learn one distinct codebook such that independent multi-codebooks (IMC) are produced. IMC is more accurate than traditional codebook because it considers the spatial information of visual words to a certain extent. Second, we adopt non-negative sparse coding (NSC) technique to encode features. This encoding scheme can effectively encourage similar features to share similar sparse representations. Third, we design an improved intersection kernel (IIK) to compute image similarity. We validate our approach on two datasets respectively, namely our 6K dataset where images are collected from three web image search engines and publicly available University of Kentucky dataset. The experimental results demonstrate our technique achieves significant performance gain compared with state-of-the-art approaches.