In this paper, the application of an unsupervised parallel approach called the Fuzzy Hopfield Neural Network (FHNN) for vector qunatization in image compression is proposed. The main purpose is to embed fuzzy reasoning strategy into neural networks so that on-line learning and parallel implementation for codebook design are feasible. The object is to cast a clustering problem as a minimization process where the criterion for the optimum vector qunatization is chosen as the minimization of the average distortion between training vectors. In order to generate feasible results, a fuzzy reasoning strategy is included in the Hopfield neural network to eliminate the need of finding weighting factors in the energy function that is formulated and based on a basic concept commonly used in pattern classification, called the "within-class scatter matrix" principle. The suggested fuzzy reasoning strategy has been proven to allow the network to learn more effectively than the conventional Hopfield neural network. The FHNN based on the within-class scatter matrix shows the promising results in comparison with the c-means and fuzzy c-means algorithms.

This paper presents a novel classified vector quantizer (CVQ) design algorithm which can control the rate and storage size for applications of image coding. In the algorithm, the classification of image blocks is based on the edge orientation of each block in the wavelet domain. The algorithm allocates the rate and storage size available to each class of the CVQ optimally so that the average distortion is minimized. To reduce the arithmetic complexity of the CVQ, we employ a partial distance codeword search algorithm in the wavelet domain. Simulation results show that the CVQ enjoys low average distortion, low encoding complexity, high visual perception quality, and is well-suited for very low bit rate image coding.

We present a mechanism, named the law of the jungle (LOJ), to improve the Kohonen learning. The LOJ is used to be an adaptive vector quantizer for approximating nonstationary probability distribution functions. In the LOJ mechanism, the probability that each node wins in a competition is dynamically estimated during the learning. By using the estimated win probability, "strong" nodes are increased through creating new nodes near the nodes, and "weak" nodes are decreased through deleting themselves. A pair of creation and deletion is treated as an atomic operation. Therefore, the nodes which cannot win the competition are transferred directly from the region where inputs almost never occur to the region where inputs often occur. This direct "jump" of weak nodes provides rapid convergence. Moreover, the LOJ requires neither time-decaying parameters nor a special periodic adaptation. From the above reasons, the LOJ is suitable for quick approximation of nonstationary probability distribution functions. In comparison with some other Kohonen learning networks through experiments, only the LOJ can follow nonstationary probability distributions except for under high-noise environments.

This letter presents a novel variable-rate vector quantizer (VQ) design algorithm, which is a hybrid approach combining a genetic algorithm with the entropy-constrained VQ (ECVQ) algorithm. The proposed technique outperforms the ECVQ algorithm in the sense that it reaches to a nearby global optimum rather than a local one. Simulation results show that, when applied to the image coding, the technique achieves higher PSNR and image quality than those of ECVQ algorithm.

We have developed and fabricated an LSI called the FMPP-VQ64. The LSI is a memory-based shared-bus SIMD parallel processor containing 64 PEs, intended for low bit-rate image compression using vector quantization. It accelerates the nearest neighbor search (NNS) during vector quantization. The computation time does not depend on the number of code vectors. The FMPP-VQ64 performs 53,000 NNSs per second, while its power dissipation is 20 mW. It can be applied to the mobile telecommunication system.

This paper focuses on competitive learning algorithms for WTA (winner-take-all) networks which perform rotation invariant pattern classification. Although WTA networks may theoretically be possible to achieve rotation invariant pattern classification with infinite memory capacities, actual networks cannot memorize all input data. To effectively memorize input patterns or the vectors to be classified, we present two algorithms for learning vectors in classes (LVC1 and LVC2), where the cells in the network memorize not only weight vectors but also their firing numbers as statistical values of the vectors. The LVC1 algorithm uses simple and ordinary competitive learning functions, but it incorporates the firing number into a coefficient of the weight change equation. In addition to all the functions of the LVC1, the LVC2 algorithm has a function to utilize under-utilized weight vectors. From theoretical analysis, the LVC2 algorithm works to minimize the energy of all weight vectors to form an effective memory. From computer simulation with two-dimensional rotated patterns, the LVC2 is shown to be better than the LVC1 in learning and generalization abilities, and both are better than the conventional Kohonen self-organizing feature map (SOFM) and the learning vector quantization (LVQ1). Furthermore, the incorporation of the firing number into the weight change equation is shown to be efficient for both the LVC1 and the LVC2 to achieve higher learning and generalization abilities. The theoretical analysis given here is not only for rotation invariant pattern classification, but it is also applicable to other WTA networks for learning vector quantization.

In order to construct fuzzy systems automatically, there are many studies on combining fuzzy inference with neural networks. In these studies, fuzzy models using self-organization and vector quantization have been proposed. It is well known that these models construct fuzzy inference rules effectively representing distribution of input data, and not affected by increment of input dimensions. In this paper, we propose a destructive fuzzy modeling using neural gas network and demonstrate the validity of a proposed method by performing some numerical examples.

A new compression algorithm for fingerprint images is introduced. A modified wavelet packet scheme which uses a fixed decomposition structure, matched to the statistics of fingerprint images, is used. Based on statistical studies of the subbands, different compression techniques are chosen for different subbands. The decision is based on the effect of each subband on reconstructed image, taking into account the characteristics of the Human Visual System (HVS). A noise shaping bit allocation procedure which considers the HVS, is then used to assign the bit rate among subbands. Using Lattice Vector Quantization (LVQ), a new technique for determining the largest radius of the Lattice and its scaling factor is presented. The design is based on obtaining the smallest possible Expected Total Distortion (ETD) measure, using the given bit budget. At low bit rates, for the coefficients with high-frequency content, we propose the Positive-Negative Mean (PNM) algorithm to improve the resolution of the reconstructed image. Furthermore, for the coefficients with low-frequency content, a lossless predictive compression scheme is developed. The proposed algorithm results in a high compression ratio and a high reconstructed image quality with a low computational load compared to other available algorithms.

We propose a memory-based processor called a Functional Memory Type Parallel Processor for vector quantization (FMPP-VQ). The FMPP-VQ is intended for low bit-rate image compression using vector quantization. It accelerates the nearest neighbor search on vector quantization. In the nearest neighbor search, we look for a vector nearest to an input one among a large number of code vectors. The FMPP-VQ has as many PEs (processing elements, also called "blocks") as code vectors. Thus distances between an input vector and code vectors are computed simultaneously in every PE. The minimum value of all the distances is searched in parallel, as in conventional CAMs. The computation time does not depend on the number of code vectors. In this paper, we explain the detail of the architecture of the FMPP-VQ, its performance and its layout density. We designed and fabricated an LSI including four PEs. The test results and performance estimation of the LSI are also reported.

This paper describes an adaptive neural vector quantization algorithm with a deleting approach of weight (reference) vectors. We call the algorithm an adaptive learning and self-deleting algorithm. At the beginning, we introduce an improved topological neighborhood and an adaptive vector quantization algorithm with little depending on initial values of weight vectors. Then we present the adaptive learning and self-deleting algorithm. The algorithm is represented as the following descriptions: At first, many weight vectors are prepared, and the algorithm is processed with Kohonen's self-organizing feature map. Next, weight vectors are deleted sequentially to the fixed number of them, and the algorithm processed with competitive learning. At the end, we discuss algorithms with neighborhood relations compared with the proposed one. The proposed algorithm is also good in the case of a poor initialization of weight vectors. Experimental results are given to show the effectiveness of the proposed algorithm.

A multiple classifier system can be a powerful solution for robust pattern recognition. It is expected that the appropriate combination of multiple classifiers may reduce errors, provide robustness, and achieve higher performance. In this paper, high performance Chinese syllable recognition is presented using combinations of multiple classifiers. Chinese syllable recognition is divided into base syllable recognition (disregarding the tones) and recognition of 4 tones. For base syllable recognition, we used a combination of two multisegment vector quantization (MSVQ) classifiers based on different features (instantaneous and transitional features of speech). For tone recognition, vector quantization (VQ) classifier was first used, and was comparable to multilayer perceptron (MLP) classifier. To get robust or better performance, a combination of distortion-based classifier (VQ) and discriminant-based classifier (MLP) is proposed. The evaluations have been carried out using standard syllable database CRDB in China, and experimental results have shown that combination of multiple classifiers with different features or different methodologies can improve recognition performance. Recognition accuracy for base syllable, tone, and tonal syllable is 96.79%, 99.82% and 96.24% respectively. Since these results were evaluated on a standard database, they can be used as a benchmark that allows direct comparison against other approaches.

In DCT transform coding it is usually necessary to discard some of the ac coefficients obtained after the transform operation for data compression reasons. Although most of the energy is usually compacted in the few coefficients that are transmitted, there are many instances where the discarded coefficients contain significant information. The absence of these coefficients at the decoder can lead to visible degradation of the reconstructed image especially around slow moving objects. We propose a simple but effective method which uses an indirect form of vector quantization to supplement scalar quantization in the transform domain. The distribution pattern of coefficients that fall below a fixed threshold is vector quantized and an index of the pattern chosen from a codebook is transmitted together with two averages; one for the positive coefficients and the other for negative coefficients. In the reconstruction, the average values are used instead of setting the corresponding coefficients to zero. This is tantamount to quantizing the mid range frequency coefficients with 1 bit but the resulting bit-rate is much less. We aim to propose an alternative to using traditional vector quantization which entails computational complexities and large time and memory requirements.

In this paper, we propose a new approach to speaker independent isolated-word speech recognition using multisegment multiple vector quantization (VQ) codebooks. In this approach, words are recognized by means of multisegment multiple VQ codebooks, a separate multisegment multiple VQ codebooks are designed for each word in the recognition vocabulary by dividing equally the word into multiple segments which is correlative with number of syllables or phonemes of the word, and designing two individual VQ codebooks consisting of both instantaneous and transitional speech features for each segment. Using this approach, the influence of the within-word coarticulation can be minimized, the time-sequence information of speech can be used, and the word length differences in the vocabulary or speaking rates variations can be adapted automatically. Moreover, the mel-cepstral coefficients based on unbiased estimation of log spectrum (UELS) are used, and comparison experiment with LPC derived mel cepstral coefficients is made. Recognition experiments Using testing databases consisting of 100 Japanese words (Waseda database) and 216 phonetically balanced words (ATR database), confirmed the effectiveness of the new method and the new speech features. The approach is described, computational complexity as well as memory requirements are analyzed, the experimental results are presented.

This paper proposes a new pitch prediction method for 4 kbps CELP (Code Excited LPC) speech coding with 20 msec frame, for the future ITU-T 4 kbps speech coding standardization. In the conventional CELP speech coding, synthetic speech quality deteriorates rapidly at 4 kbps, especially for female and children's speech with short pitch period. The pitch prediction performance is significantly degraded for such speech. The important reason is that when the pitch period is shorter than the subframe length, the simple repetition of the past excitation signal based on the estimated lag, not the pitch prediction, is usually carried out in the adaptive codebook operation. The proposed pitch prediction method can carry out the pitch prediction without the above approximation by utilizing the current subframe excitation codevector signal, when the pitch prediction parameters are determined. To further improve the performance, a split vector synthesis and perceptually spectral weighting method, and a low-complexity perceptually harmonic and spectral weighting method have also been developed. The informal listening test result shows that the 4 kbps speech coder with 20 msec frame, utilizing all of the proposed improvements, achieves 0.2 MOS higher results than the coder without them.

In this paper, a new off-line handwritten word recognition method based on the explicit modeling of character junctures is presented. A handwritten word is regarded as a sequence of characters and junctures of four types. Hence both characters and junctures are explicitly modeled. A handwriting system employing hidden Markov models as the main statistical framework has been developed based on this scheme. An interconnection network of character and ligature models is constructed to model words of indefinite length. This model can ideally describe any form of hamdwritten words including discretely spaced words, pure cursive words, and unconstrained words of mixed styles. Also presented are efficient encoding and decoding schemes suitable for this model. The system has shown encouraging performance with a standard USPS database.

This paper presents the M-LCELP (Multi-mode Learned Code Excited LPC) speech coder, which has been developed for the next generation half-rate digital cellular telephone systems. M-LCELP develops the following techniques to achieve high-quality synthetic speech at 4kb/s with practically reasonable computation and memory requirements: (1) Multi-mode and multi-codebook coding to improve coding efficiency, (2) Pitch lag differential coding with pitch tracking to reduce lag transmission rate, (3) A two-stage joint design regular-pulse codebook with common phase structure in voiced frames, to drastically reduce computation and memory requirements, (4) An efficient vector quantization for LSP parameters, (5) An adaptive MA type comb filter to suppress excitation signal inter-harmonic noise. The MOS subjective test results demonstrate that 4.075kb/s M-LCELP synthetic speech quality is mostly equivalent to that for a North American full-rate standard VSELP coder. M-LCELP codec requires 18 MOPS computation amount. The codec has been implemented using 2 floating-point dsp chips.

We propose a recognition method for HMM using a simultaneous generative histogram. Proposed method uses the correlation between two features, which is expressed by a simultaneous generative histogram. Then output probabilities of integrated HMM are conditioned by the codeword of another feature. The proposed method is applied to isolated digit word recognition to confirm its validity.

This letter proposes an LSP quantizing method which uses interframe correlation of the parameters. The quantized parameters are represented as a moving average of code vectors. Using this method, LSP parameters are quantized efficiently and the degradation of decoded parameters caused by bit errors affects only a few following frames.

This study proposes a new training method for hidden Markov model with separate vector quantization (SVQ-HMM) in speech recognition. The proposed method uses the correlation of two different kinds of features: cepstrum and delta-cepstrum. The correlation is used to decrease the number of reestimation for two features thus the total computation time for training models decreases. The proposed method is applied to Japanese language isolated dgit recognition.

This paper discusses the bit-rate compression of super high definition still images with subband coding. Super high definition (SHD) images with more than 20482048 pixels or resolution are introduced as the next generation imaging system beyond HDTV. In order to develop bit-rate reduction algorithms, an image evaluation system for super high definition images is assembled. Signal characteristics are evaluated and the optimum subband analysis/synthesis system for the SHD images is clarified. A scalar quantization combined with run-length and Huffman coding is introduced as a conventional subband coding algorithm, and its coding performance is evaluated for SHD images. Finally, new coding algorithms based on block Huffman coding and entropy coded vector quantization are proposed. SNR improvement of 0.5 dB and 1.0 dB can be achieved with the proposed block Huffman coding and the vector quantization algorithm, respectively.