Near-duplicate image detection with cascade method

Abstract

In this paper, a novel scheme to tackle the task of near-duplicate image detection is presented. The scheme is based on a two-level image similarity measure strategy, which reduces the overall computational cost. The second-level similarity measure considering spatial position relationship can find the small similar objects in two images. Given two input images, which are represented with multiple local features, the proposed algorithm can assert whether the reference image is a near-duplicate of the query image or not. The algorithm is demonstrated on some image or video keyframe pairs with scale change, viewpoint change, blur, noise and spatial deformation, which are extracted from INRIA copy dataset, etc. The experimental results show that proposed algorithm is simple and effective.