動態影像編碼法之率差關係估計與次最佳量化器控制

Abstract

在一般傳輸速率受限的狀況下作影像與視訊編碼，量化步階的決定在編碼 後資料率與編碼品質的控制上扮演一個重要的角色。在論文中，我們分析 了若干影音光碟(Video CD)中所包含的量化步階訊息，並將其與一些習知 的演算法則比較。此外，我們探討如何作量化步階的決定以求好的編碼品 質。於此，我們採用一為人熟知的演算法為基礎，即MPEG2之Test Model ，並在其上作改進。我們知道巨集方塊(macroblock)的率差關係曲線完整 表現了資料率與失真之間的關係，故可用以作量化步階選擇之用。因為計 算實際率差曲線的高複雜度，我們發展了一個區段近似模型以估計實際曲 線。在決定合適的量化步階時，我們先根據緩衝器(buffer)狀態決定一個 參考步階值以防止緩衝器發生溢滿(overflow)及空乏(underflow)現象。 接下來，人眼視覺與率差曲線之斜率二因素被合併考慮以調整此步階值以 決定次最佳化之步階值。依模擬實驗結果顯示照此改進的法則，在編碼品 質上，我們不僅可以提高PSNR值且可以改善視覺品質。 For image and video coding under transmission constraints, determination of quantization steps at encoder plays an important role in controlling the generated bit-rate and coding performance. In this thesis, we analyze the quantization information from some Video CD titles and compare them with several known rate control algorithms. In addition, we investigate how to determine quantization steps for good coding quality. We approach this by improving on a well-known quantizer control algorithm, namely, Test Model of MPEG2. We know that the rate-distortion curve of a macroblock fully characterize the relationship between rate and distortion and can thus be used in determination of good quantization steps. Becasue of the heavy computational load required for calculating the real R-D curves, we develop a piecewise approximation model to predict the real curves. In deciding an appropriate quantization step, we first calculate a reference step on the buffer status to prevent the buffer from overflow and underflow. Next, the human vision factor and the slope of the estimated rate-distortion curve are incorporated into adjustment of this quantization step to obtain a suboptimal quantization step. Simulation reesults show that, we can attain not only higher PSNR values but also higher human visual quality in coded video.