An Integrated Approach to Near-duplicate Image Detection

Near-duplicate image detection is a task to find clusters of images that are considered to be the same pictures in human view. This is important in image recommendation systems, because when the systems recommend candidate images, redundancies of retrieved candidate images need to be avoided. In addition, in the era of big-data where image data is overflowing, its importance in terms of saving storage resources further increases. In this paper, we propose a robust model for detecting various types of near-duplicate images by integrating four different detection modules, where we use multiple image feature extractors such as Gabor filter and deep networks. The four modules are then integrated to conduct the multivariate log-likelihood ratio test for detecting duplication. Through computational experiments, we confirmed that our method reaches state-of-the-art performance.