Development of multidimensional variable rate sampling method and its application to high definition image cmmunications

采用可变速率采样的系统设计成功的关键是如何在采样和量化引起的失真之间进行权衡。由于量化器是利用自适应采样建立在编码器的递归闭环系统中，因此很难设计出同时对输入信号最优的采样器和量化器。在本项目中，我们利用基于广义压扩模型的矢量量化器在编码器闭环内来解决这一问题。本课题的研究成果总结如下：(1)在多种不同的采样率和量化码率条件下，对再现图像进行了客观评价测试。结果表明，图像的客观再现质量与主观质量指标SNR具有较好的一致性。(2)研究了SNR与采样率的关系、SNR与量化码率的关系以及这些参数之间的相互关系。结果表明，简单地最小化采样和量化引起的总偏移量可以获得良好的编码性能。(3)为了控制采样和量化引起的失真，提出了两种图像数据压缩方法。一种是基于分段双线性块的矢量量化方法，另一种是基于递归矢量量化的方法。在这两个系统中，采样率都是自适应控制的，以适应图像的局部特征。(4)提出了一种基于二次范数压扩模型描述的压扩矢量量化器的设计方法。(5)提出了矢量量化的快速编码算法。

The key to the successful design of the systems that utilize the variable rate sampling is the way of ajusting trade-offs between the distortions due to sampling and quantization. It is difficult to design the sampler and quantizer that are simultaneously optimum to the input signal since the quantizer is built in the recursive closed loop of the encoder using adaptive sampling. In this project, we utilize the vector quantizer based on the generalized companding model within the encoder closed loop for resolving this problem. The results of this project are summaried as follows.(1) Objective evaluation tests are conducted using reproduced images under many different conditions on sampling rate and quantizer bit rate. It is found that good coincidence is observed between the objective reproduction quality of image and the subjective quality measure, SNR.(2) Moreover, the investigations are done on the SNR vs.sampling rate and the SNR vs.the quantizer bit rate, and the mutual relations of these parameters. It is found that simply minimizing the total distotions due to sampling and quantization gives good coding performance.(3) In order to control the distortions due to sampling and quantization, two image data compression methods are developed. One of them is based on the vector quantization of the piece-wise bi-linear patch, and the other is based on the recursive vector quantization mothod. In both systems, the sampling rate is adaptively controlled so as to fit the local feature of the image.(4) A method for designing companded vector quantizers that is described by using the quadratic-norm based companding model is developed.(5) Fast encoding algorithms of vector quantization are developed.