• Speech Sciences
• /
• v.13 no.4
• /
• pp.177-186
• /
• 2006

In an effort to enhance the quality of feature vector classification and thereby reduce the recognition error rate of the speaker-independent speech recognition, we employ the Mahalanobis distance in the calculation of the similarity measure between feature vectors. It is assumed that the metric matrix of the Mahalanobis distance be diagonal for the sake of cost reduction in memory and time of calculation. We propose that the diagonal elements be given in terms of the variations of the feature vector components. Geometrically, this prescription tends to redistribute the set of data in the shape of a hypersphere in the feature vector space. The idea is applied to the speech recognition by hidden Markov model with fuzzy vector quantization. The result shows that the recognition is improved by an appropriate choice of the relevant adjustable parameter. The Viterbi score difference of the two winners in the recognition test shows that the general behavior is in accord with that of the recognition error rate.

• Proceedings of the Korean Institute of Intelligent Systems Conference
• /
• 2000.05a
• /
• pp.31-34
• /
• 2000

코호넨 자기조직 신경망을 사용하면 클러스터링뿐만 아니라 그 데이터가 할당된 클러스터의 대표값(Centroid)과의 거리 차이(Quantization Error)를 알아볼 수 있다 이를 이용하면 어떤 데이터가 정상적인 분포를 따르는지 정상적인 분포에서 벗어나는 비정상적인 데이터인지 알 수 있고, 유닉스 시스템 사용자의 명령어 사용 패턴에 적용하여 어떤 사용자의 명령어 사용 패턴이 정상적인 것인지 비정상적인 것인지 알 수 있다. 본 논문에서는 유닉스 시스템 사용자 8명의 명령어 패턴을 클러스터링한 후 Quantization Error를 이용하여 비정상 패턴을 탐지하는 오프라인에서의 비정상 행위를 탐지하는 시스템을 구현하였다. 그리고 통계적인 학습 방법을 적용한 비정상 패턴 탐지와의 비교를 통하여 두 가지 비정상 패턴 탐지 결과가 동일함을 확인하였다.

• 한국정보디스플레이학회:학술대회논문집
• /
• 2009.10a
• /
• pp.1350-1353
• /
• 2009

Conventional depth extraction in integral imaging is based on the disparity information between the elemental images. Since the disparity is measured in pixel unit, however, the extracted depth is discrete, resulting in the quantization error. Moreover, the quantization error grows as the object depth increases, which limits the accuracy of the depth extraction for distant objects. In this paper, we propose a new method for depth extraction in integral imaging using sub-pixel registration information between subimages to obtain linear and accurate depth.

• The KIPS Transactions:PartB
• /
• v.12B no.4 s.100
• /
• pp.379-386
• /
• 2005

This paper provides a rigorous theory for analysis of quantization effects and optimum filter bank design in quantized multidimensional subband filter banks. Even though subband filter design has been a hot topic for last decades, a few results have been reported on the subband filter with a quantizer. Each pdf-optimized quantizer is modeled by a nonlinear gain-plus-additive uncorrelated noise and embedded into the subband structure. Using polyphase decomposition of the analysis/synthesis filter banks, we derive the exact expression for the output mean square quantization error. Based on the minimization of the output mean square error, the technique for optimal filter design methodology is developed. Numerical design examples for optimum nonseparable paraunitary and biorthogonal filter banks are presented with a quincunx subsampling lattice. Through the simulation, $10\~20\;\%$ decreases in MSE have been observed compared with subband filter with no quantizers especially for low bit rate cases.

• Proceedings of the Korean Institute of Communication Sciences Conference
• /
• 1985.10a
• /
• pp.220-223
• /
• 1985

• Journal of information and communication convergence engineering
• /
• v.12 no.4
• /
• pp.215-220
• /
• 2014

We consider quantization optimized for distributed estimation, where a set of sensors at different sites collect measurements on the parameter of interest, quantize them, and transmit the quantized data to a fusion node, which then estimates the parameter. Here, we propose an iterative quantizer design algorithm with a weighted distance rule that allows us to reduce a system-wide metric such as the estimation error by constructing quantization partitions with their optimal weights. We show that the search for the weights, the most expensive computational step in the algorithm, can be conducted in a sequential manner without deviating from convergence, leading to a significant reduction in design complexity. Our experments demonstrate that the proposed algorithm achieves improved performance over traditional quantizer designs. The benefit of the proposed technique is further illustrated by the experiments providing similar estimation performance with much lower complexity as compared to the recently published novel algorithms.

• Journal of the Korean Institute of Telematics and Electronics S
• /
• v.34S no.2
• /
• pp.43-51
• /
• 1997

In this paper, we quantize the image data to simplify the systolic array architecture for block matching algorithm. As the number of bits for pixel data to be processed is reduced by quantization, one can simplify the hardware of systolic array. Especially, if the bit serial input is used, one can even more simplify the structure of processing element. First, we analize the effect of quantization to a block matching. then we show the structure of quantizer and processing element when bit serial input is used. The simulation results applied to standard images have shown that the proposed block matching method has less prediction error than the conventional high speed algorithm.

• Proceedings of the Acoustical Society of Korea Conference
• /
• 1998.06c
• /
• pp.197-200
• /
• 1998

본 논문에서는 유전자 알고리즘(genetic Algorithm)을 사용하여 벡터 양자화(vector quantization : VQ)를 수행하는 방법을 제안하고자 한다. 벡터 양자화를 수행하여 코드북(codebook)을 생성할 때 생성된 코드북과 학습벡터(training vector)사이에는 반드시 양자화 오차(quantization error)가 발생하는데 기존의 K-means 알고리듬을 사용하여 코드북을 생성했을 경우 양자화 오차를 줄이는데 한계가 있었다. 본 논문에서 제안하는 유전자 알고리즘을 이용한 벡터 양자화는 이 양자화 오차를 감소시키기 위해서 연구되었다. 제안한 방법의 성능을 평가하기 위해 음성데이터를 기존의 K-means 알고리즘에서 클러스터의 중심을 선택하는 방법중의 하나인 Minimax방법으로 코드북을 생성하여 제안한 방법과 양자화 오차를 비교한 결과 양자화 오차가 감소됨을 알 수 있었다.

• Journal of information and communication convergence engineering
• /
• v.9 no.2
• /
• pp.155-160
• /
• 2011

In distributed source localization where sensors transmit measurements to a fusion node, we address the sensor selection problem where the goal is to find the best set of sensors that maximizes localization accuracy when quantization of sensor measurements is taken into account. Since sensor selection depends heavily upon rate assigned to each sensor, joint optimization of rate allocation and sensor selection is required to achieve the best solution. We show that this task could be accomplished by solving the problem of allocating rates to each sensor so as to minimize the error in estimating the position of a source. Then we solve this rate allocation problem by using the generalized BFOS algorithm. Our experiments demonstrate that the best set of sensors obtained from the proposed sensor selection algorithm leads to significant improvements in localization performance with respect to the set of sensors determined from a sensor selection process based on unquantized measurements.

• Speech Sciences
• /
• v.11 no.4
• /
• pp.17-28
• /
• 2004

In this paper, we propose a statistical approach to improve the performance of spectral quantization of speech coders. The proposed techniques compensate for the distortion in a decoded line spectrum pairs (LSP) vector based on a statistical mapping function between a decoded LSP vector and its corresponding original LSP vector. We first develop two codebook-based probabilistic matching (CBPM) methods based on linear mapping functions according to different assumption of distribution of LSP vectors. In addition, we propose an iterative procedure for the two CBPMs. We apply the proposed techniques to a predictive vector quantizer used for the IS-641 speech coder. The experimental results show that the proposed techniques reduce average spectral distortion by around 0.064dB.