Made available in DSpace on 2015-05-14T12:36:26Z (GMT). No. of bitstreams: 1 parte1.pdf: 2704137 bytes, checksum: 1bc9cc5c3099359131fb11fa1878c22f (MD5) Previous issue date: 2010-08-31 / Coordenação de Aperfeiçoamento de Pessoal de Nível Superior - CAPES / This work analyzes methods for textures images classification and segmentation using lossless data compression algorithms models. Two data compression algorithms are evaluated: the Prediction by Partial Matching (PPM) and the Lempel-Ziv-Welch (LZW) that had been applied in textures classification in previous works. The textures are pre-processed using histogram equalization. The classification method is divided into two stages. In the learning stage or training, the compression algorithm builds statistical models for the horizontal and the vertical structures of each class. In the classification stage, samples of textures to be classified are compressed using models built in the learning stage, sweeping the samples horizontally and vertically. A sample is assigned to the class that obtains the highest average compression. The classifier tests were made using the Brodatz textures album. The classifiers were tested for various contexts sizes (in the PPM case), samples number and training sets. For some combinations of these parameters, the classifiers achieved 100% of correct classifications. Texture segmentation process was made only with the PPM. Initially, the horizontal models are created using eight textures samples of size 32 x 32 pixels for each class, with the PPM context of a maximum size 1. The images to be segmented are compressed by the models of classes, initially in blocks of size 64 x 64 pixels. If none of the models achieve a compression ratio at a predetermined interval, the block is divided into four blocks of size 32 x 32. The process is repeated until a model reach a compression ratio in the range of the compression ratios set for the size of the block in question. If the block get the 4 x 4 size it is classified as belonging to the class of the model that reached the highest compression ratio. / Este trabalho se propõe a analisar métodos de classificação e segmentação de texturas de imagens digitais usando algoritmos de compressão de dados sem perdas. Dois algoritmos de compressão são avaliados: o Prediction by Partial Matching (PPM) e o Lempel-Ziv-Welch (LZW), que já havia sido aplicado na classificação de texturas em trabalhos anteriores. As texturas são pré-processadas utilizando equalização de histograma. O método de classificação divide-se em duas etapas. Na etapa de aprendizagem, ou treinamento, o algoritmo de compressão constrói modelos estatísticos para as estruturas horizontal e vertical de cada classe. Na etapa de classificação, amostras de texturas a serem classificadas são comprimidas utilizando modelos construídos na etapa de aprendizagem, varrendo-se as amostras na horizontal e na vertical. Uma amostra é atribuída à classe que obtiver a maior compressão média. Os testes dos classificadores foram feitos utilizando o álbum de texturas de Brodatz. Os classificadores foram testados para vários tamanhos de contexto (no caso do PPM), amostras e conjuntos de treinamento. Para algumas das combinações desses parâmetros, os classificadores alcançaram 100% de classificações corretas. A segmentação de texturas foi realizada apenas com o PPM. Inicialmente, são criados os modelos horizontais usados no processo de segmentação, utilizando-se oito amostras de texturas de tamanho 32 x 32 pixels para cada classe, com o contexto PPM de tamanho máximo 1. As imagens a serem segmentadas são comprimidas utilizando-se os modelos das classes, inicialmente, em blocos de tamanho 64 x 64 pixels. Se nenhum dos modelos conseguir uma razão de compressão em um intervalo pré-definido, o bloco é dividido em quatro blocos de tamanho 32 x 32. O processo se repete até que algum modelo consiga uma razão de compressão no intervalo de razões de compressão definido para o tamanho do bloco em questão, podendo chegar a blocos de tamanho 4 x 4 quando o bloco é classificado como pertencente à classe do modelo que atingiu a maior taxa de compressão.

Language Identification (LID) is the task of automatically identifying the language of speech signal uttered by an unknown speaker. An N language LID task is to classify an input speech utterance, spoken by an unknown speaker and of unknown text, as belonging to one of the N languages L1, L2, . . , LN. We present a new approach to spoken language modeling for language identification using the Lempel-Ziv-Welch (LZW) algorithm, with which we try to overcome the limitations of n-gram stochastic models by automatically identifying the valid set of variable length patterns from the training data. However, since several patterns in a language pattern table are also shared by other language pattern tables, confusability prevailed in the LID task. To overcome this, three pruning techniques are proposed to make these pattern tables more language specific. For LID with limited training data, we present another language modeling technique, which compensates for language specific patterns missing in the language specific LZW pattern table. We develop two new discriminative measures for LID based on the LZW algorithm, viz., (i) Compression Ratio Score (LZW-CRS) and (ii) Weighted Discriminant Score (LZW-WDS). It is shown that for a 6-language LID task of the OGI-TS database, the new model (LZW-WDS) significantly outperforms the conventional bigram approach. With regard to the front end of the LID system, we develop a modified technique to model for Acoustic Sub-Word Units (ASWU) and explore its effectiveness. The segmentation of speech signal is done using an acoustic criterion (ML-segmentation). However, we believe that consistency and discriminability among speech units is the key issue for the success of ASWU based speech processing. We develop a new procedure for clustering and modeling the segments using sub-word GMMs. Because of the flexibility in choosing the labels for the sub-word units, we do an iterative re-clustering and modeling of the segments. Using a consistency measure of labeling the acoustic segments, the convergence of iterations is demonstrated. We show that the performance of new ASWU based front-end and the new LZW based back-end for LID outperforms the earlier reported PSWR based LID.