向量量化法應用於影像壓縮之研究

Abstract

在此論文中我們將提出四種快速向量量化搜尋方法以及一個新的影像壓縮 方法。第一種快速向量量化搜尋方法稱為特徵向量法 (Eigenvector Method)。在此法中，我們首先計算此組訓練向量的特徵向量，然後將每 一向量轉換到此新的座標系統中，最後應用部分去除搜尋法來加速去除不 可能對應之編碼向量。第二種 (Mean-or-Variance Method)及第三種( Mean-and-Variance Method)利用每一向量區塊之二個特徵：平均值及變 異數，當二個區塊之平均值或變異數相差太大時，可以藉由導出的不等式 將許多不可能對應之編碼向量去除。最後一種快速向量量化搜尋方法( Mean Pyramid Method)是對於每一個向量區塊先建立其平均值的金字塔資 料結構。快速搜尋則是由此金字塔資料結構的第一層開始，逐層去除不可 能對應之編碼向量。實驗結果顯示，所提出之方法非常有效而且不會增加 影像之失真程度。在所提出的新的影像壓縮方法中，對於較平滑之區域採 用較大的向量區塊以增加影像壓縮倍率。而對於變化較大之區域則採用較 小的向量區塊，且以去掉平均值的分類向量量化法來對這些向量區塊編碼 。在此論文中我們根據邊緣方向性提出一種新的影像區塊分類法。同時， 我們也將提出新的預測方法以對所有向量區塊的平均值及邊緣區塊的向量 量化編碼位址做更有效的編碼。實驗結果顯示，所提出之方法可以得到相 當良好的影像品質。 In this dissertation, we will propose four fast VQ codebook search methods (the eigenvector method, the mean-or-variance method, the mean-and-variance method, and mean pyramid method) and a novel image compression approach. In the eigenvector method (EVM), each vector is represented by a new coordinate system formed by the eigenvectors which are obtained through KLT. The partial distortion elimination (PDE) method is then applied to quickly discard those unmatched codewords. The mean- or-variance (MOV) and mean- and-variance (MAV) algorithms use two significant features of a vector, mean value and variance, to reject the unmatched codewords. The mean pyramid method (MPM) uses the mean pyramids of codewords to reject codewords. The mean pyramid for each vector is first established. Fast search is then performed from the top level to the bottom level of the mean pyramid. Many codewords can be eliminated in the search process. In the proposed image compression approach, variable block size is adopted, using quadtree decomposition, to code low- activity regions which usually occupy large areas in an image with fewer bits of code. Mean-removed classified vector quantization is then used to code high-activity regions. A new edge-oriented classifier is proposed for block classification. Additionally, the mean values of all blocks and addresses of edge blocks were efficiently coded with a novel predictive noiseless coding (NPNC) method. Experiment results show that excellent reconstructed images and higher PSNR have been obtained.