• Journal of Korea Multimedia Society
• /
• v.7 no.12
• /
• pp.1671-1679
• /
• 2004

This study has proposed visual-feature extraction methods for each band in wavelet domain with both spatial frequency features and multi resolution features. In addition, it has brought forward similarity measurement method using fuzzy theory and new color feature expression method taking advantage of the frequency of the same color after color quantization for reducing quantization error, a disadvantage of the existing color histogram intersection method. Experiments are performed on a database containing 1,000 color images. The proposed method gives better performance than the conventional method in both objective and subjective performance evaluation.

• The Transactions of the Korea Information Processing Society
• /
• v.6 no.4
• /
• pp.988-998
• /
• 1999

This paper is to investigate the method of reducing the number of fuzzy rules with the help of LVQ. a large number of training patterns usually leads to a large set of fuzzy rules that require a large computer memory and take a long time to perform classification. so, in order to solve these problems, it is necessary to study to minimize the number of fuzzy rules. However, so as to minimize the performance degradation resulting from the reduction of fuzzy rules, fuzzy rules are generated after training the high-quality initial reference pattern. Through the simulation, we confirm that the proposed method is very effective.

• The Journal of the Korea institute of electronic communication sciences
• /
• v.9 no.11
• /
• pp.1221-1226
• /
• 2014

In an effort to improve the efficiency of the speech recognition, we investigate the effect of the number of phoneme clusters. For this purpose, codebooks of varied number of phoneme clusters are prepared by modified k-means clustering algorithm. The subsequent processing is fuzzy vector quantization (FVQ) and hidden Markov model (HMM) for speech recognition test. The result shows that there are two distinct regimes. For large number of phoneme clusters, the recognition performance is roughly independent of it. For small number of phoneme clusters, however, the recognition error rate increases nonlinearly as it is decreased. From numerical calculation, it is found that this nonlinear regime might be modeled by a power law function. The result also shows that about 166 phoneme clusters would be the optimal number for recognition of 300 isolated words. This amounts to roughly 3 variations per phoneme.

• Proceedings of the Korean Institute of Intelligent Systems Conference
• /
• 2004.10a
• /
• pp.153-157
• /
• 2004

본 논문은 패턴인식을 위해 사용할 수 있는 감독학습을 이용한 supervised IAFC neural network 1과 supervised IAFC neural network 2를 제안하였다 Supervised IAFC neural network 1과 supervised IAFC neural network 2는 LVQ(Learning Vector Quantization)를 퍼지화한 새로운 퍼지 학습법칙을 사용하고 있다. 이 새로운 퍼지 학습 법칙은 기존의 학습률 대신에 퍼지화된 학습률을 사용하고 있는데, 이 퍼지화된 학습률은 조건 확률을 퍼지화 한 것에 근간을 두고 있다. Supervised IAFC neural network 1과 supervised IAFC neural network 2의 성능과 오류역전파 신경회로망의 성능을 비교하기 위하여 iris 데이터를 사용하였는데, 실험결과 supervised IAFC neural network 2 의 성능이 오류역전파 신경회로망의 성능보다 우수함이 입증되었다.

• The Transactions of the Korea Information Processing Society
• /
• v.5 no.7
• /
• pp.1933-1942
• /
• 1998

분류 벡터 양자화(classified vector quantization: CVQ)〔2의 부코드북을 설계함에 있어서, 경쟁 학습 네트워크〔5〕-〔7〕 는 소속도의 이분법적 표현으로 상당한 소속도를 가지는 벡터들이 학습 과정에 무시되는 경향을 가진다. 이를 개선하기 위해 제안된 퍼지 경쟁 학습 네트워크〔8〕는 각 클러스터가 연속적인 소속도를 가진다는 개념을 도입하여 이와 같은 문제들을 해결했다. 그러나 퍼지 경쟁 학습 네트워크를 CVQ에 적용할 경우, 각 부코드북의 크기를 시행착오로 결정해야 하는 문제점을 여전히 가지고 있으며, 이러한 문제점들의 개선을 위하여 본 논문에서는 수정 퍼지 경쟁 학습 네트워크(modified fuzzy competitive learning network)를 제안한다. 수정 퍼지 경쟁 학습 네트워크는 퍼지 학습 네트워크가 가지는 이 분법적 소속도를 연속적인 소속도로 확장하여, 학습 과정중에 나타날 수 있는 지역 최소점 도달을 억제하였다.

• Speech Sciences
• /
• v.12 no.4
• /
• pp.7-16
• /
• 2005

We investigate the effect of membership concentration on the performance of the speaker-independent recognition system by FVQ/HMM. For the membership function, we adopt the result obtained from the objective function approach by Bezdek. Membership concentration is done by varying the exponent in the membership function. The number of selected clusters is constrained to two for the sake of cheap computational cost. Experimental results showed that the recognition rate has its maximum value when the membership function was taken to be inversely proportional to the distance of the input vector from the cluster centroid. When the membership concentration was two weak or too strong, the performance was found to be relatively poor as expected. Except these extreme cases, the membership concentration was not shown to affect the recognition rate significantly. This is in accordance with the general observation that the fuzzy system is not much sensitive. to the detailed shape of the membership function as long as it is overlapped over multiple classes.

• Journal of Institute of Control, Robotics and Systems
• /
• v.6 no.8
• /
• pp.710-718
• /
• 2000

In this paper we described an approach to automation of visual inspection of solder joint defects of SMC(Surface Mounted Components) on PCBs(Printed Circuit Board) by using neural network and fuzzy rule-based classification method. Inherently the surface of the solder joints is curved tiny and specular reflective it induces difficulty of taking good image of the solder joints. And the shape of the solder joints tends to greatly vary with the soldering condition and the shapes are not identical to each other even though the solder joints belong to a set of the same soldering quality. This problem makes it difficult to classify the solder joints according to their qualities. Neural network and fuzzy rule-based classification method is proposed to effi-ciently make human-like classification criteria of the solder joint shapes. The performance of the proposed approach is tested on numerous samples of commercial computer PCB boards and compared with the results of the human inspector performance and the conventional Kohonen network.

• Journal of the Korean Institute of Intelligent Systems
• /
• v.20 no.5
• /
• pp.672-676
• /
• 2010

This paper presents a learning rule that weights more on data near decision boundary. This learning rule generates better decision boundary by reducing the effect of outliers on the decision boundary. The proposed learning rule is integrated into IAFC neural network. IAFC neural network is stable to maintain previous learning results and is plastic to learn new data. The performance of the proposed fuzzy neural network is compared with performances of LVQ neural network and backpropagation neural network. The results show that the performance of the proposed fuzzy neural network is better than those of LVQ neural network and backpropagation neural network.

• The Journal of the Acoustical Society of Korea
• /
• v.15 no.4
• /
• pp.36-42
• /
• 1996

In this paper, we have proposed the robust estimation of HMM parameters which is based on CDDN(codeword dependent distance normalization)and codeword weighting by distance. The proposed method has used a distance normalization based on the characteristics of a codeword dependent distribution and have computed fuzzy contributions of codeword to a input vector with a fuzzy objective function. From experimental results, we have shown the effectiveness of the proposed method in that the correction rate of the proposed method is improved 4.5% over the conventional FVQ based method. Especially, the application of distance weighting to smoothing of output probability is improved the performance of 2.5% compared to distance based codeword weighting.

• Phonetics and Speech Sciences
• /
• v.1 no.2
• /
• pp.3-8
• /
• 2009

In an effort to improve speech recognition, we investigated the effect of the number of clusters. In usual LBG clustering, the number of codebook clusters is doubled on each bifurcation and hence cannot be chosen arbitrarily in a natural way. To have the number of clusters at our control, we combined adaptive resonance theory (ART2) with LBG and perform the clustering in two stages. The codebook thus formed was used in subsequent processing of fuzzy vector quantization (FVQ) and HMM for speech recognition tests. Compared to conventional LBG, our method was shown to reduce the best recognition error rate by 0${\sim$}0.9% depending on the vocabulary size. The result also showed that between 400 and 800 would be the optimal number of clusters in the limit of small and large vocabulary speech recognitions of isolated words, respectively.