Multiple Features Fusion Based Inverted Multi-index for Image Retrieval

In the community of image retrieval, single feature shows a very low discriminative power for matching, and typically the frequently-used SIFT feature only describes the local gradient distribution. Therefore, false matches occur prevalently. Besides, conventional inverted index usually returns long candidate lists for queries, with sparse subdivision of search space and limited accuracy. To tackle these problems, this paper proposes an effective image retrieval framework: multiple features fusion based inverted multi-index (mFFMI). Inside this framework, features are incorporated into inverted multi-index, providing respective complementary image information for accurate image retrieval. We fuse Root SIFT (a variant of SIFT), color and texture features into mFFMI framework, and experiments on benchmark datasets show that mFFMI structure yields favorable accuracy improvement, compared with state-of-the-art. Moreover, another significant characteristic of mFFMI is that it is well compatible with many other techniques, and a better performance can be achieved if combined with such prior techniques.