APPLICATIONS OF ITAKURA-SAITO TYPE SPECTRAL DISTORTION MEASURES TO IMAGE ANALYSIS AND CLASSIFICATION
Yusuf Öztürk, H. Abut · 2005
In this study we have applied Itakura-Saito (IS) type spectral distortion measures for classifying segments of digitized images using the multidimensional multichannel linear prediction (MLPC) theory. ItakuraSaito type spectral distortion measures have been successfully applied in the fields of speech coding, speech recognition and speaker identification and verification. But there has been only one such study to apply the gain-normalized Itakura-Saito distortion measure for classification of arbitrarily shaped image textures. Maragos et al. [l] have demonstrated that textures can be classifed by two-dimensional linear predictive models under a spectral distortion measure. We have extended this case to the more difficult, yet much more meaningful, multichannel linear predictive modeling [3] of arbitrarily shaped image textures. All three forms of the Itakura-Saito distortion measure [4,5], gain sensitive (GS), gain optimized (GO), and gain normalized (GN), have been extended to the multichannel linear prediction case. We have successfully used them in the template design and the classification modes.