以四分樹預做切割的區塊壓縮法

Abstract

在本論文中，我們提出以四分樹預做切割，再用向量量化區塊壓縮來做影 像壓縮的方法。首先，我們將影像先切成許多不重疊，大小為 32*32 的 子區塊。對於每一個子區塊，根據四分樹切割法我們決定其是否要切成更 小的區塊。因為傳統四分樹結構需要大量記憶體。為解決此問題，我們提 出三種四分樹的編碼方法：位置四分樹、位元對映四分樹和向量量化位置 四分樹，來改善四分樹對記憶體的需求。事實證明，我們提出的三種方法 比傳統四分樹結構節省了大量的記憶體。此外，在對切割後的區塊做編碼 時，32*32，16*16 或 8*8 的區塊必然歸為平滑區塊，而 4*4 的區塊就 細分成平滑區塊、雜亂區塊和邊緣區塊。平滑區塊是以此區塊的平均值當 灰度代表值來記錄，雜亂區塊是以二灰度代表值的向量量化區塊壓縮方法 來編碼，而邊緣區塊則以三灰度代表值的向量量化區塊壓縮方法來編碼以 便改善邊的視覺品質。實驗結果顯示，復原影像的視覺品質不錯而壓縮誤 差也很小。以我們提出的壓縮方法，每個影像點約只需用 0.7-0.9 位元 來記錄。 A quadtree-based BTC-VQ coding method is proposed in the thesis for image compression. First, the image is partitioned into nonoverlapping 32*32 subimages. For each subimage, according to the smoothness of the gray values, the subimage is partioned using a quadtree segmentation. In order to improve the efficiency of the conventional quadtree structure, we propose three new methods, called position quadtree, bitmap quadtree, and position-VQ quadtree, to encode the segmentation quadtree. The number of bits needed to encode the segmentation quadtree is reduced significantly by using these three new methods. While coding the segmented blocks, the blocks with the size of 32*32, 16*16, and 8*8 are all coded using the mean value because they are identified as smooth blocks. On the other hand, the blocks with the size of 4*4 are classified into three categories: the smooth blocks, the texture blocks, and the edge blocks. A smooth block is coded using the mean value of the block. A texture block is coded with a 2-level BTC-VQ, while each edge block is coded using a 3-level BTC with VQ technique to improve the perceived quality. Simulation results show that good visual quality and low MSE are obtained. In addition, the bit rate is in the range of 0.7-0.9 bpp when the proposed QTBTC- VQ is used.