適用於可調式小波視訊編碼之訊源機率模型與位元率-失真最佳化方法

Abstract

本研究主題可分為兩個主要項目：動態補償(motion-compensated) 差值訊號之訊源模型(source model)以及位元率-失真(rate-distortion)最佳化參數選擇，在第一個項目中，我們發展了零值廣義高斯分佈(ρ-GGD)訊源模型，可準確模擬可調式小波編碼(scalable wavelet coding)中的訊號機率分佈。我們提出了分段式線性方法可有效率估測在零值廣義高斯模型中的型態參數，並且藉由改善零值機率的估測而提高機率模型之精準度。在第二個項目中，我們提出了一個位元率-失真(rate-distortion)模型用以描述可調式小波視訊編碼中的動態預測效率。可調式編碼架構為開放式迴圈(open-loop)並且同時有多種位元率的編碼需求，與傳統非可調式編碼有很大的不同，也因此傳統上廣泛使用的拉格朗日(Lagrangian)最佳化方法無法良好應用於可調式小波視訊編碼上。為了找到在動態資訊與殘存訊號間最好的位元率分配方法，我們提出了動態資訊增益(MIG)做為量測動態預測效率指標。基於這項指標，新的代價函式一同被提出。相較於傳統拉格朗日最佳化作法，我們的實驗結果顯示了所提出的模式決定方法可在SNR與畫面率可調式條件下，擁有較佳的PSNR表現。

There are two key elements in this study, namely, the source modeling of the motion-compensated prediction error signals, and the coding parameter selection to minimize the rate-distortion criterion. For the first item, we develop an accurate ρ-GGD (Generalized Gaussian Distribution) source model for approximating the signal probability distribution in scalable wavelet coding. An efficient piecewise linear expression is designed to estimate the shape parameter of the ρ-GGD. We also improve the model accuracy in matching the real data by modifying the ρ parameter estimation formula. For the second item, a rate-distortion model for describing the motion prediction efficiency in scalable wavelet video coding is proposed. Different from the conventional non-scalable video coding, the scalable wavelet video coding needs to operate under multiple bitrate conditions and it has an open-loop structure. The conventional Lagrangian multiplier, which is widely used to solve the rate-distortion optimization problems in video coding, does not fit well into the scalable wavelet structure. In order to find the rate-distortion trade-off due to different bits allocated to motion and textual residual information, we suggest a motion information gain (MIG) metric to measure the motion prediction efficiency. Based on this metric, a new cost function for mode decision is proposed. Compared with the conventional Lagrangian optimization, our experimental results show that the new mode decision method generally improves the PSNR performance in the combined SNR and temporal scalability cases.