Distance Measures in the Training Phase of Self-Organizing Map for Color Histogram Generation in Spectral Image Retrieval

Abstract

The usefulness of different distance measures in the training phase of self-organizing map (SOM) for color histogram generation for spectral image retrieval purposes is examined. The calculation of the best-matching unit (BMU) in the training phase of SOM is done by using Euclidean distance, Kullback-Leibler distance, Jeffrey divergence, and CIEL*a*b* color difference as distance measures. One-dimensional SOMs are generated for two different data sets consisting of 1269 Munsell color chips and 1, 440, 000 color spectra collected from a real spectral image database. The suitability of the introduced measures is first evaluated by calculating the average color differences between the Munsell data set and its BMUs in the SOMs trained by Munsell data. The achieved results are validated by a practical application, in which the queries from a real spectral image database are performed. Furthermore, the ability of SOMs trained by different distance measures to distinguish between spectral images of real human skin and magazine prints of human skin is examined. The achieved results are promising and indicate that two-dimensional self-organizing maps, which are trained by using Euclidean distance and Jeffrey divergence as distance measure and color histograms that correspond the spectral images as training data, could be used for classifying spectral images.