Image Subband Coding using Block Adaptive Vector Quantization

研究了基于小波变换的高效图像编码方法。利用两级树形结构的滤波器组将原始图像分解为7个子带。研究了几种矢量量化的方法：(1)所有子带类型和逐级分辨率类型：前者由从所有子带中选取的15个样本组成，后者由两种类型的矢量组成，一种是3维的2级矢量，另一种是1维的1维的矢量。对于这两种类型，我们根据它们的功率将矢量分为三类，每一类都反映了它所对应的像素块的特征。每个向量通过其类别码本进行矢量量化。这两种类型的编码性能几乎相同，但后者在计算量上优于前者。(2)方向分组子带类型：所有子带被分成3组，即水平、垂直和对角。在每组中，形成5维矢量，并根据它们的矢量权将其分为3类。然后，每个组类向量由其自己的码本进行矢量量化。这种方法的结果与0.4-0.7 bpp之间的逐级类型的性能一样好。当码率大于0.7bpp时，方向分组型优于其他子带类型；当码率低于0.4bpp时，所有子带类型均优于其他子带类型。(3)显著样本类型：水平扫描时，由n个相应的显著样本组成一个矢量，其幅度大于某一阈值。然后，为了压缩这些重要样本的地址信息，我们以每个样本块而不是每个单个样本来决定是否有效。该方案在编码性能上优于上述方法。

High efficiency image coding using wavelet transform is investigated. The original image is decomposed into 7 subbands by 2-level tree-structured filter bank. Vector quantization is applied to the high frequency subbands, while the lowest subband is coded by 2D-DPCM.Several methods of vector quantization are investigated as follows :(1) All subband type and resolution level by level type : In the former, each vector consists of 15 samples picking from all subbands and the latter consists of two kinds of vectors, one is 3 dimensional for level 2 and the other 12 dimensional for level 1. For both types, we classify vectors into three classes by their powers, each of which reflects the feature of its corresponding pixel block. Each vector is vector-quantized by its class codebook. Both types have nearly the same coding performance, but the latter is superior to the former in having less amounts of computation.(2) Directionally grouped subband type : All subbands are grouped into 3 groups, i. e., horizontal vertical and diagonal. In each group, 5 dimensional vectors are formed and classified into 3 classes by their vector powers. Then each group-class vector is vector quantized by its own code book. This method results as good performance as level by level type between 0.4-0.7bpp. For the rate higher than 0.7bpp, directionally grouped type is superior and for the rate lower than 0.4bpp, all subband type is superior to others.(3) Significant sample type : A vector is formed with n consequtive significant samples in horizontal scanning, whose amplitudes are greater than a threshold. Then, in order to compress the address information of those significant samples, we do decision whether effective or not by every sample block instead of every single sample. This scheme is superior to the above described methods in coding performance.