An Effective Feature Fusion Solution for Image Retrieval
Yuanyuan Li, Zhiyang Li, Wenyu Qu, Jun Wang, Junfeng Wu and Changqing Ji
For the Bag-of-Words based image retrieval approaches, most of them rely on the local feature. However, the simple SIFT feature has led to false positive matches since the SIFT feature only describes the local gradient distribution. It has been shown that combining multiple complementary features such as shape, color or texture is a challenging and promising task which can improve the retrieval accuracy. However, most of the state-of-the-art feature fusion methods are usually requested to provide some weight assignments for the features, and providing such weight assignments for the features is not always easy without any initial knowledge about the image database. Furthermore obtaining the weight value requires an additional learning stage. In this paper, we present a new feature fusion solution, called SKFF, which takes advantage of the skyline operator to combine the local features and need not to provide weight value for each feature. The new solution incorporates the multiple dimensional feature vector into skyline operator, each dimension corresponds to one kind of feature similarity. The candidate results of our method include not only the images which are more similar to the query image in multiple cues, but also in simple cue. According to detailed analysis, we show that the proposed solution can overcome the deficiencies of simple feature image retrieval. Moreover, we evaluate the proposed solution through extensive experiments on several public image sets. Our experiments show that the proposed mechanism improves the retrieval accuracy significantly compared to existing techniques.
Keywords: Feature fusion, skyline, image retrieval, BoW