Performance of similarity measures based on histograms of local image feature vectors

We investigate similarity measures for image retrieval from databases based on histograms of local feature vectors. The feature vectors are obtained from grouping quantized block transforms coefficients and thresholding. After preliminaries on block transforms we are introducing binary DC and AC feature vectors. Subsequently ternary DC and AC vectors are defined. Next we show how the histograms of vectors defined can be combined to form similarity measure for image retrieval from database. We formulate the database training and retrieval problem using the defined similarity measures. Performance results are shown using widely used FERET and ORL databases and the cumulative match score evaluation. We show that despite simplicity the proposed measures provide results which are on par with best results using other methods. This indicates that statistics based retrieval should not be underestimated comparing to structural methods.