Adding spatial distribution clue to aggregated vector in image retrieval

This study proposes a novel algorithm that enhances the distinctiveness of the traditional vector of locally aggregated descriptors (VLAD) using spatial distribution clue of local features. The algorithm introduces a new method to compute the spatial distribution entropy (SDE) of clusters. Unlike conventional methods, this algorithm considers the distribution of full spatial information provided by local feature detectors rather than only utilizing the spatial coordinate statistics. For each cluster, the corresponding spatial distribution is computed using a histogram of spatial locations, scales, and orientations of all local features inside the cluster. Entropy is calculated from the spatial distributions of all clusters of an image to create a distribution function, which is further normalized and concatenated with the VLAD vector to generate the final representation. Image retrieval and classification experiments on public datasets are performed. Experimental results show that the proposed algorithms produce better or comparable retrieval performance than several state-of-the-art algorithms. In addition, we extend our SDE to the convolutional neural network (CNN) feature, which further improves the CNN feature result in image retrieval.