Partial near-duplicate image identification with global geometric consistency of subset-of-features

The task of determining whether two images are near-duplicate or not becomes increasingly important in many applications, such as copyright infringement detection, sub- image retrieval and spam image filtering. Traditional methods for near-duplicate image identification (NDII) usually extract image local features firstly, and then quantize them as bag of words (BOW); the frequency histogram is finally taken as the representation of image for NDII. However, the mismatches between local features, the lower distinctiveness, polysemy and synonymy of visual words all degrade the accuracy of NDII, especially for partial NDII. Although some geometric verification procedures have been taken, these methods are still affected by the mismatches and ambiguity of visual words. In this paper, we propose a novel scheme for verifying the global geometric consistency of subset-of-features for improving the accuracy of BOW model. If there is a subset of matched pairs of local features obtained by BOW model, in which the ratios of scales and differences of orientations are consistent, we take these two images as near-duplicate images. The cardinality of the subset can also be used for measure the similarity of the two images. Experimental results show that the proposed method can improve the accuracy of NDII prominently, and it is also effective and robust for the retrieval of some typical partial near-duplicate images.