使用有限狀態熵編碼之誤差補償式向量格化法作影像壓縮

Abstract

我們將在本論文中對向量格化 (vector quantization) 作一深入研究。 於向量格化過程中使用有限狀態的熵編碼 (finite state entropy coding), 誤差補償式還原(error-componsated reconstruction)，以及 採用逐級方式處理向量格化(progressive VQ) 是本論文中三個主要探討 的技術。在具有限狀態熵編碼的向量格化中， 所有狀態使用相同的共通 編碼簿 (universal codebook)，但是每一個狀態各擁有其可變長度通道 碼 (channel codeword) 的編碼簿，該通道碼的功能是傳送相對應的向量 格化之向量碼。 此有限狀態熵編碼法為一利用相鄰區段間的關聯性以降 低傳送所需位元數的簡易架構。我們可以在解碼端 (decoder) 使用誤差 補償式還原的後處理級 (post-processing stage) 來改善還原影像的品 質，該後處理級是利用邊緣連續 (edge continuity) 的特性來達到改善 影像品質的目的。不管是否為記憶式的向量格化， 此後處理級皆可適用 。經模擬證實此後處理級可在不傳送任何多餘的資訊下，訊噪比 (SNR) 仍比傳統的向量格化多出一分貝。 如果向量格化同時使用有限狀態的編 碼以及誤差補償式還原，則在傳送較少的位元數下仍可得到比傳統向量格 化品質更佳的還原影像。理論上，向量格化的性能會隨著向量長度的增加 而提升， 但其複雜度亦隨之增高，其複雜度增高的程度通常難以處理， 因此我們提出逐級式的向量格化， 希望在可接受的複雜度下開發較長向 量所擁有的潛在增益。 逐級式的向量格化可分五級運作於離散餘弦轉換 域 (DCT domain)在最初一級中， 我們使用具有限狀態編碼的純量格化在 做直流係數的編碼。至於其他四級，則分別對其子向量內的交流係數作向 量格化。 除了降低複雜度外， 逐級向量格化更可以根據每個區段的特性 分配位元數，讓還原影像的品質更均勻，逐級向量格化的前兩級是一定得 運作， 但是後三級則是根據前級的〞誤差平方和〞(mse) 和我們設定的 〞門檻值〞 (threshold value) 決定是否運作。 We consider vector quantization (VQ) for image coding. Three specific techniques are investigated, namely, VQ with finite- state entropy coding, error-compensated reconstruction, and progressive VQ. In VQ with finite-state entropy coding, all the states employ the same universal VQ codebook, but each state has its own variable-length codebook for mapping of the VQ code vectors into corresponding channel codewords. The finite-state entropy coding exploits the inter-block correlation among nearby image blocks (vectors) and it is found that even a simple scheme can reduce the bit rate significantly. The reconstructed image quality in the decoder can be enhanced with an error-compensating postprocessing stage which seeks to maintain edge continuity across block boundaries. This postprocessing function can be used with VQ encoders with or without memory. Simulation results show that it yields an improvement of 1 dB over ordinary VQ. Thus a combined use of finite-state entropy coding and error-compensated reconstruction in VQ-based image coding can lead to a better coded image quality at a lower bit rate than conventional VQ of the same size. The progressnive VQ is proposed to capitalize on the potential gain of large-vector VQ while keeping the complexity low. It has five stages and operates in the DCT domain. At the lowest stage, the dc coefficients are coded by scalar quantization with finite-state entropy coding to exploit the inter-block correlation. At the other four stages, ordinary VQ is employed on other subsets of the DCT coefficients.In addition to reducing complexity, the progressive VQ can be used to dynamically allocate bits to each block based on its characteristics to achieve a more uniform coded image quality.A particular method for doing so is discussed.In this method, the first two stages of the progressive VQ is always performed and the other three stages are selected based on a comparison of the sum of squared errors in each block and some threshold.