Robust fusion of color and local descriptors for image retrieval and classification

Abstract

This paper introduces an optimized image descriptor that combines both global and local features for image retrieval and classification. Color histograms in HSV space are extracted and quantized as global features, while root scale-invariant feature transform (rootSIFT) descriptors are densely extracted as local features. The extracted features are fused and reduced to obtain a lower-dimensional descriptor and discriminate the underlying variances of data. Image descriptors are encoded by the visual locally aggregated features (VLAD) approach. The Corel image dataset is used for evaluation and benchmarking. The experimental results show that the proposed descriptor improves the classification accuracy by 5% as well as the retrieval accuracy by 10% and 20% over rootSIFT and HSV, respectively. Additionally, the retrieval model outperforms many state-of-the-art approaches.