Wyniki wyszukiwania - Biblioteka Nauki

1

Can learning vector quantization be an alternative to SVM and deep learning? : recent trends and advanced variants of learning vector quantization for classification learning

100%

Villmann T. , Bohnsack A. , Kaden M.

|

tom Vol. 7, No. 1

65--81

EN

Learning vector quantization (LVQ) is one of the most powerful approaches for prototype based classification of vector data, intuitively introduced by Kohonen. The prototype adaptation scheme relies on its attraction and repulsion during the learning providing an easy geometric interpretability of the learning as well as of the classification decision scheme. Although deep learning architectures and support vector classifiers frequently achieve comparable or even better results, LVQ models are smart alternatives with low complexity and computational costs making them attractive for many industrial applications like intelligent sensor systems or advanced driver assistance systems. Nowadays, the mathematical theory developed for LVQ delivers sufficient justification of the algorithm making it an appealing alternative to other approaches like support vector machines and deep learning techniques. This review article reports current developments and extensions of LVQ starting from the generalized LVQ (GLVQ), which is known as the most powerful cost function based realization of the original LVQ. The cost function minimized in GLVQ is an soft-approximation of the standard classification error allowing gradient descent learning techniques. The GLVQ variants considered in this contribution, cover many aspects like bordersensitive learning, application of non-Euclidean metrics like kernel distances or divergences, relevance learning as well as optimization of advanced statistical classification quality measures beyond the accuracy including sensitivity and specificity or area under the ROC-curve. According to these topics, the paper highlights the basic motivation for these variants and extensions together with the mathematical prerequisites and treatments for integration into the standard GLVQ scheme and compares them to other machine learning approaches. For detailed description and mathematical theory behind all, the reader is referred to the respective original articles. Thus, the intention of the paper is to provide a comprehensive overview of the stateof- the-art serving as a starting point to search for an appropriate LVQ variant in case of a given specific classification problem as well as a reference to recently developed variants and improvements of the basic GLVQ scheme.

2

New Bit Reduction of Vector Quantization Using Block Prediction and Relative Addressing

100%

Hu Y.-C. , Tsai P. , Lo C.-C.

|

2008

|

tom Vol. 87, nr 3-4

313-329

EN

A low bit rate image coding scheme based on vector quantization is proposed. In this scheme, the block prediction coding and the relative addressing techniques are employed to cut down the required bit rate of vector quantization. In block prediction coding, neighboring encoded blocks are taken to compress the current block if a high degree of similarity between them is existed. In the relative addressing technique, the redundancy among neighboring indices are exploited to reduce the bit rate. From the results, it is shown that the proposed scheme significantly reduces the bit rate of VQ while keeping good image quality of compressed images.

3

Audio Compression using a Modified Vector Quantization algorithm for Mastering Applications

88%

Prince S. , D B. , Kirubaraj A. A. , Immanuel S. J. , M S.

|

tom Vol. 69, No. 2

287--292

EN

Audio data compression is used to reduce the transmission bandwidth and storage requirements of audio data. It is the second stage in the audio mastering process with audio equalization being the first stage. Compression algorithms such as BSAC, MP3 and AAC are used as standards in this paper. The challenge faced in audio compression is compressing the signal at low bit rates. The previous algorithms which work well at low bit rates cannot be dominant at higher bit rates and vice-versa. This paper proposes an altered form of vector quantization algorithm which produces a scalable bit stream which has a number of fine layers of audio fidelity. This modified form of the vector quantization algorithm is used to generate a perceptually audio coder which is scalable and uses the quantization and encoding stages which are responsible for the psychoacoustic and arithmetical terminations that are actually detached as practically all the data detached during the prediction phases at the encoder side is supplemented towards the audio signal at decoder stage. Therefore, clearly the quantization phase which is modified to produce a bit stream which is scalable. This modified algorithm works well at both lower and higher bit rates. Subjective evaluations were done by audio professionals using the MUSHRA test and the mean normalized scores at various bit rates was noted and compared with the previous algorithms.

4

Speaker recognition based on telephone quality short Polish sequences with removed silence

63%

Marciniak. T. , Krzykowska A. , Weychan R.

|

tom R. 88, nr 6

42-46

EN

This paper presents the effectiveness of speaker identification based on short Polish sequences. An impact of automatic removal of silence on the speaker recognition accuracy is considered. Several methods to detect the beginnings and ends of the voice signal have been used. Experimental research was carried out in Matlab environment with the use of a specially prepared database of short speech sequences in Polish. The construction of speaker models was realized with two techniques: Vector Quantization (VQ) and Gaussian Mixture Models (GMM). We also tested the influence of the sampling rate reduction on the speaker recognition performance.

PL

Artykuł przedstawia badania efektywności rozpoznawania mówcy opartego na krótkich wypowiedziach w języku polskim. Sprawdzono wpływ automatycznego wykrywania i usuwania ciszy na jakość rozpoznawania mówcy. Przebadano kilka różnych metod wykrywania początku i końca fragmentów mowy w wypowiadanych sekwencjach. Eksperymenty zostały przeprowadzone z użyciem środowiska Matlab i specjalnie utworzonej bazy krótkich wypowiedzi w języku polskim. Do budowy modeli mówców wykorzystano kwantyzacja wektorowa (VQ) oraz Gaussian Mixture Models (GMM). Podczas badań sprawdzono także wpływ obniżenia szybkości próbkowania na skuteczność identyfikacji mówcy.

5

Wavelet based speaker recognition

63%

Jakuczun W.

Prace Instytutu Podstaw Informatyki Polskiej Akademii Nauk

|

2003

|

tom Nr 959

1-7

EN

In this article we are presenting wavelet-based method for designing speaker recognition features. The proposed method is compared to linear prediction method. As a classificator we used LBG algorithm, which is one of the vector quantization (VG) algorithms.

PL

W artykule prezentujemy metodę szukania cech opartych na falkach dla problemu rozpoznawania mówców. Aby lepiej ocenić zaproponowaną metodą porównano ją do liniowej predykcji. Jako klasyfikatora użyliśmy algorytmu wektorowej kwantyzacji. Słowa kluczowe: rozpoznawanie mówców, falki, liniowa predykacja, kwantyfikacja wektorowa.

6

Vector quantized pattern learning for neural network-based image restoration

63%

Bao P. , Nim-Tze L.

|

tom Vol. 6, no 3-4

61-73

EN

This paper presents a hybrid scheme for image restoration with edge-preserving regularization and artificial neural network based on vector quantized pattern learning. The edge information is extracted from the source image as a priori knowledge to recover the details and reduce the ringing artifact of the subband-coded image. The spatially independent vector patterns are generated from source images using vector quantization to de-correlate the image patterns for more effective and efficient pattern learning and to minimize the number of training patterns while retaining the representativeness of the training patterns. The vector-quantized patterns are then used to train the multilayer perceptron model for the restoration process. To evaluate the performance of the proposed scheme, a comparative study with the set partitioning in hierarchical tree (SPIHT) and the full pattern trained NN has been conducted using a set of gray-scale digital images.The experimental results have drown that the proposed scheme could result in better performances compared with SPIHT on both objective and subjective quality for lower compression ratio subband coded image.