視訊編碼器之記憶體縮減與壓縮技術

Abstract

在此論文中，我們提出了多種記憶體最佳化演算法和實際可行的超大型積體電路架構可應用於系統單晶片多媒體系統。在單晶片系統設計中，許多需要大量簡單運算和資料傳輸的演算法通常實現在陣列處理器。針對這點，我們發展了一套系統化設計方法，可針對在有限的計算能力、晶片面積、和頻寬的限制之下，對於記憶體做最佳化。另外，影像壓縮在單晶片多媒體系統中扮演非常重要的角色，因為大量的視訊及影像資料如果不被壓縮的話，將會耗掉大部分的記憶體和頻寬資源。我們也發展了三種影像和貼圖壓縮演算法，可滿足不同的壓縮比和影像品質需求。 此論文主要分成三部分。第一部份我們針對完全搜尋型區塊比對演算法，提出了二種有效率的超大型積體電路架構以及記憶體最佳化演算法。從搜尋區域資料流分析中，導出systolic array 和semi-systolic array 架構。經由巧妙安排記憶體的結構，非但頻寬可被最小化，而且處理單元（processor element）的效率提升。另外，控制器結構非常簡單直接，簡化了設計複雜度且能符合需求。在另一方面，經由探討資料的相關圖（dependency graph），我們集中注意力在降低內部緩衝區大小在頻寬的限制下。我們發展的系統化設計方法可提供降低內部緩衝區非效率的使用。 在第二部分，我們提出了新穎的密碼簿最佳化演算法和貼圖壓縮演算法。向量量化演算法的優點在於解碼非常常簡單，但是在效能方面卻受到密碼簿設計很大的影響。大部分向量量化壓縮皆採用LBG法來產生密碼簿，但是LBG法仍不能產生最佳解。我們提出一種以基因演算法為基礎的密碼簿產生方式，可比LBG法提升影像品質0.01~0.14分貝。但是向量量化壓縮法的實用度仍受限於密碼簿大小、訓練集內容、以及較差的影像品質。所以我們提出一種只需要非常小的密碼簿的貼圖壓縮方法，而且密碼簿產生方式是依據被壓縮影像內容，而非訓練集。這個演算法非常適合於高品質的需求而且解壓縮非常簡單。 在第三部分，我們提出一種增進傳統小波編碼效能的演算法。這種演算法根據位元率-失真分佈可以對單一或一群組的小波係數產生最佳的位元平面擷取點。實驗結果證明在測試影像 ”Lena” 中，我們所提出的演算法比EZW高出1.09~1.1分貝，比SPIHT高出0.22~0.25分貝，比RDE高出0.02~0.35分貝，在位元率為0.125~1.0的情況下。

In this dissertation, we have proposed various memory-optimization algorithms and implemented practical VLSI architectures for SOC multimedia systems. In SOCs, the computation-intensive and memory-intensive algorithms are usually mapped to array processors. We have developed a systematic design procedure to optimize the on-chip memory size under computing power, area or I/O bandwidth constraints. Image compression algorithms are very important for SOC multimedia systems, because the huge amount of video and image data occupies very high memory size and bandwidth if the data has not been compressed. We have developed various image and texture compression algorithms, which can be used for different performance and compression ratio requirements. This dissertation is divided into three parts. The first part presents two efficient VLSI architectures and memory size optimization are proposed. Starting from an overlapped data flow of search area, both systolic array and semi-systolic array architectural solutions are derived. By means of exploiting stream memory banks, not only I/O bandwidth can be minimized but also processor element efficiency can be improved. In addition, the controller structure for both solutions are very straightforward, making them very suitable for VLSI implementation to meet computational requirements. On the other hand, by exploring the DG (dependency graph), we focus in the problem of reducing the internal buffer size under minimal I/O bandwidth constraint, this may offer the guideline on reducing redundant internal buffer and then achieve cost-effective VLSI architectures. In the second part, a novel codebook optimization algorithm and a novel texture compression algorithm are proposed. VQ-based algorithms are one of the popular types of compression algorithms, which are very simple in decoding. The performance of VQ-based algorithms is greatly affected by the codebooks. Most VQ-based algorithms incorporate the LBG algorithm as the codebook generation scheme. But the LBG algorithm only finds the sub-optimal solution for a given training set. We have proposed a genetic-based codebook generation algorithm, which outperforms the LBG algorithm with 0.01~0.14 dB in PSNR. The vector-quantization solutions of the image compression are usually limited in the high cost of the codebook size, training sets, and image quality for high-quality applications. This dissertation proposes an algorithm, where the small codebook entries are embedded into the data of compressed image block according to the context of this image block. In addition, an optimization procedure of finding optimal codebook entries based on genetic algorithms is proposed. The proposed algorithm is suitable for high-quality application and the cost of the decoder is very low and can effectively reduce the memory bandwidth for multimedia systems. In the third part, a novel wavelet image coder for quality improvement is presented. The proposal achieves better bit-plane truncation of wavelet coefficients by exploiting optimal truncation point for each group of coefficients in different levels depending on rate-distortion distribution. Simulation results show our proposed algorithms outperform EZW, SPIHT, and RDE by 1.09~1.1 dB, 0.22~0.25 dB, and 0.02~0.35 dB respectively at 0.125~1 bpp for image “Lena”.