Unsupervised Visual Dictionary Construction Methods Via Clustering Analysis

Visual dictionary is the basic and key toefficiently represent image with Bag of Visual Words (BoVW) Model. The unsupervised visual dictionary construction methods based on clustering analysis are researched in this paper. Aiming at the problems that K-Means based clustering methods (K-Means and HKM) are sensitive to the initialization and easy to converge to the local optimum, a modified spectral clustering method is proposed to construct visual dictionary, which works in a way to cluster observations via eigenvalue decomposition according to their similarity matrix and can converge to the global optimal regardless of the distribution of feature space. The performance of visual dictionaries constructed by Random Sampling, K-Means, HKM and the improved Spectral Clustering is assessed through image retrieval experiments, and the effectiveness of clustering analysis especially the improved Spectral Clustering method to construct visual dictionary is validated.