Abstract
The main reason for adopting different techniques in vector quantizers (VQ) is to design an optimal quantizer. Since the actual probability distributions of image or voice data is not known, the vector quantizers are trained with long sequences of image or voice data known as a training sequence. Various distortion measures are present for quantizing the data. This paper aims at suggesting the implementation of three types of distortions (squared error distortion, linear distortion and piece-wise linear distortion), measuring and comparing the efficiency of each of the techniques in terms of average distortion and time for each measure. Quantization is of two types: scalar quantization and vector quantization. Scalar quantization is a phenomenon on which, each pixel is quantized either by reducing the number of gray levels or by reducing the resolution. Vector quantization is one in which the whole image is divided into several blocks and a code vector replaces each block. The main principle involved in the vector quantization is the clustering technique, which is applied to generate an optimal codebook for the given training sequence. After quantizing the training sequence the authors obtain a codebook which optimally represents the probability distribution of intensities in images. This work mainly concentrates on vector quantization, based on a long training sequence of data. Methods are discussed to reduce computational complexity of the vector quantizer algorithm, thereby making the codebook design faster. The quantizer designed in the authors' work reduces the psychovisual redundancy present in an image data by quantizing the data. This module can be used in the quantizer step of the overall image compressor.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call