Research Of Near-Duplicate Image Elimination

With the rapid development of the Internet and the widespread popularity of image processing software, more and more users like to share their favorite images on social network or personal home page, resulting in a large amount of redundant images, as well as a waste of network storage space and delay of image searching.Accordingly, reform and innovation of near-duplicate image elimination are keys for image management. Among the existing image de-duplication technologies, most researchers always focus on the study of near-duplicate image clustering, ignoring deep analysis in redundant image eliminating. Thus, this paper improves the performance of existing methods by proposing two superior redundancy elimination methods. The main contributions of our paper can be summarized as follows:(1) PageRank based near-duplicate image elimination method. This method combines both global feature and local feature for fast and accurate image clustering. For seed image selection, PageRank algorithm is used to analyze the contextual relationship between clustered images, and this analysis is based on images’ visual similarity. Finally, the most representative image will be selected to be reserved, and other redundant images will be eliminated. Experimental results show that this method makes a good performance on the dataset composed of natural near-duplicate images. However, its performance on artificial near-duplicate images needs to be improved.(2) IRGfusion analysis based near-duplicate image elimination method. This method is the improved version of the former method, which is also suitable for artificial near-duplicate images that have undergone complicated copy attacks. The main improvements in this method include the refinement of image features, the IRG construction and fusion. This method enhances the performance of near-duplicate elimination, which can accurately trace the original version of clustered near-duplicate image group. Experimental results show that this method improves the former method, which not only have high accuracy on natural near-duplicate image set, but also perform well on artificial near-duplicate image set. Moreover, the anti-interference ability of this method is also very strong.