Image retrieval using extended bag-of-visual-words

Bag of Visual Word (BoVW) is a popular visual content based image retrieval method, applied successfully over the years. However, the standard BoVW representation takes into consideration the quantitative information of visual words discarding semantic and spatial information of the images, crucial for retrieval. Several techniques have been applied so far by the researchers to improve the BoVW using semantic of the images. In this paper, we extend the original BoVW by capturing semantic of the images using eigenvectors obtained from the patches of the training images. The spatial information of eigenvectors are utilized to build the extended BoVW (EBoVW), represented as a projection vector with dimension equal to the number of selected eigenvectors. For each training image the projection of the patches along each eigenvector is evaluated and the nearest eigenvector is identified. This information is used to modify the respective element of the projection vector which has been appended with the original BoVW to obtain the EBoVW. Performance of retrieval has been improved using EBoVW compared to the state-of-the-art image retrieval techniques as demonstrated in this paper using Coil-100 dataset.