• Journal of KIISE:Software and Applications
• /
• v.27 no.4
• /
• pp.422-431
• /
• 2000

This paper proposes a new codebook generation method, called a PCA-Based VQ, that incorporates the PCA (Principal Component Analysis) technique into VQ (Vector Quantization) codebook design. The PCA technique reduces the data dimensions by transforming input image vectors into the feature vectors. The cluster of feature vectors in the transformed domain is bisectioned into two subclusters by an optimally chosen partitioning hyperplane. We expedite the searching of the optimal partitioning hyperplane that is the most time consuming process by considering that (1) the optimal partitioning hyperplane is perpendicular to the first principal axis of the feature vectors, (2) it is located on the equilibrium point of the left and right cluster's distortions, and (3) the left and right cluster's distortions can be adjusted incrementally. This principal axis bisectioning is successively performed on the cluster whose difference of distortion between before and after bisection is the maximum among the existing clusters until the total distortion of clusters becomes as small as the desired level. Simulation results show that the proposed PCA-based VQ method is promising because its reconstruction performance is as good as that of the SOFM (Self-Organizing Feature Maps) method and its codebook generation is as fast as that of the K-means method.

• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.25 no.1A
• /
• pp.48-55
• /
• 2000

In this paper, an assumption to design a quantizer, is proposed that if one small region of a probability density function is represented larger probability and bigger total error than another neighbour region, then the quantizer is not optimal. It is tested when the probability functions are Gaussian, Laplacian and uniform density function by the computer simulations. A new LBG algorithm which originates from this assumption in addition to LBG algorithm, is designed for the vector quantizer. The new LBG algorithm presents better performance than the original LBG algorithm in the average error and the variance of the error.

• Journal of the Korean Institute of Telematics and Electronics B
• /
• v.33B no.1
• /
• pp.161-169
• /
• 1996

This paper presents a new search region prediction method using vector quantization for the motion estimation. We find motion vectors using the full search BMA from two successive frame images first. Then the motion vectors are used for training a codebook. The trained codebook is the predicted search region. We used the unsupervised neural network for VQ encoding and codebook design. A major advantage of formulating VQ as neural networks is that the large number of adaptive training algorithm that are used for neural networks can be applied to VQ. The proposed method reduces the computation and reduce the bits required to represent the motion vectors because of the smaller search points. The computer simulation results show the increased PSNR as compared with the other block matching algorithms.

• Journal of the Korean Institute of Telematics and Electronics S
• /
• v.36S no.11
• /
• pp.89-94
• /
• 1999

In vector quantization (VQ), mean squared error (MSE) is widely used as a distance measure between vectors. But the distance between averages appears as a dominant quantity in MSE. In the case of image vectors, the coincidence of edge pattern is also important considering human visual system (HVS). Therefore, this paper presents a new distance measure using the variance of difference (VD) as a criterion for the coincidence of edge pattern. By using this in the VQ encoding, we can reduce the degradation of edge region in the reconstructed image. And applying this to the codebook design, we can obtain the final codebook that has a lot of various edge codevectors instead of redundant shade ones.

• The Journal of the Acoustical Society of Korea
• /
• v.10 no.4
• /
• pp.38-47
• /
• 1991

The frequency variation among speakers is one of problems in the speech recognition. This paper applies fuzzy theory to solve the variation problem of frequency features. Reference patterns are expressed by fuzzified patterns which are produced by the peak frequency and the peak energy extracted from codebooks which are generated from training words uttered by several speakers, as they should include common features of speech signals. Words are recognized by fuzzy inference which uses the certainty factor between the reference patterns and the test fuzzified patterns which are produced by the peak frequency and the peak energy extracted from the power spectrum of input speech signals. Practically, in computing the certainty factor, to reduce memory capacity and computation requirements we propose a new equation which calculates the improved certainty factor using only the difference between two fuzzy values. As a result of experiments to test this word recognition method by fuzzy interence with Korean digits, it is shown that this word recognition method using the new equation presented in this paper, can solve the variation problem of frequency features and that the memory capacity and computation requirements are reduced.

• Journal of the Institute of Electronics Engineers of Korea CI
• /
• v.38 no.5
• /
• pp.27-36
• /
• 2001

This paper proposes new design algorithms based on stochastic relaxation (SR) for an on-line vector quantizer (VQ) design. These proposed SR methods solve the local entrapment problems of the conventional Kohonen learning algorithm (KLA). These SR methods cover two different types depending upon the use of simulated annealing (SA) : the one that uses SA is called the OLVQ SA and the other the OLVQ SR. These methods arc combined with the KLA and therefore preserve the its convergence properties. Experimental results for Gauss Markov sources, real speech and image demonstrate that the proposed algorithms can consistently provide better codebooks than the KLA.