Detailed Investigation and Comparison of Various Binarization Algorithms for Hashing

The fact that the number of images stored continues to increase day by day makes content-based image retrieval (CBIR) systems more important. Considering that almost all of the datasets encountered today are large, it is evident that using the exact nearest neighbor (ENN) search will be inefficient and time-consuming. Therefore, hashing approach, which is seen as the most effective approximate nearest neighbor (ANN) method, is preferred in nearly all retrieval tasks. To learn these binary hash codes, feature extraction, dimension reduction, and binarization steps are generally applied. In this study, the effects of the approaches used in the binarization section on retrieval success are analyzed comparatively. For this purpose, the Kvasir dataset (includes gastrointestinal tract images) and 630 pyramid histograms of oriented gradients (PHOG) features of this dataset are used. First, the dimension of PHOG features is reduced to 16, 32, and 64 bits (usually hash code lengths) using principal component analysis (PCA). Then these different length feature vectors are converted into hash codes by binarization process. Five different threshold methods are used for the binarization process. Binarization is performed by means of techniques such as hard threshold, mean value threshold, adaptive threshold, line mean value, class mean value, total mean value. Finally, the retrieval performances of the hash codes are tested. The performances of the binarization methods are presented comparatively.