Deterministic Annealing Method with Equidistortion Principle for Vector Quantization

Abstract

The deterministic annealing (DA) method for vectorquantization can obtain better representative vectors thanthe LBG, CL, FSCL and SOFM. But it has two problems. One isthat quality of the results is sensitive to the reductionof parameter T which is a temperature in an annealingoperation. Another is that learning time is relativelylong. The equidistortion principle has been derived and itis shown effective for vector quantization. In this paper,to solve the above problems, a new algorithm named DAEP(deterministic annealing method with equidistortionprinciple) is proposed. In the DAEP, the idea ofequidistortion principle is introduced into the DA method.In the annealing operation, location of representativevectors are forced to move to the global optimum pointsusing the equidistortion principle. Judging from thequality of results and learning time, it is shown that theDAEP can obtain better results than the other methods.