Formulation of the modified akaike information criterion (AIC) for a one‐dimensional multipulse model and its application to image coding

Abstract

AbstractIn recent years, attention has been given to coding based on the structurally extracted component and the three dimensional object model. However, it has a problem in that the content of the processing and the required model are fairly complex compared to the conventional waveform coding.From such a viewpoint, the authors considered the one‐dimensionalmultipulse model, which was proposed as an analysis/synthesis technique for speech, and have been investigating the method using the model which can easily extract the structure of the image and provide the coding.In this paper, the multipulse model is regarded as a stochastic model. A method is proposed, where the number of pulses required to represent the features of the image is determined automatically using Akaike's information criterion and a similar criterion (called modified AIC).As the first step, the modified AIC for the multipulse model is derived, and the sequential computation procedure is shown. Then the relation between the penalty coefficient of the modified AIC and the optimal number of pulses is described, and a method is shown based on that result where the traditional information criterion is utilized selectively based on the constraint for the compression ratio of the number of pulses.Finally, the modified AIC for the multipulse model is applied to the hierarchical coding from the two‐tone image to the gray‐level image, and the usefulness is demonstrated. It is expected that the proposed method will be applied to the image coding, image feature extraction and speech feature extraction.