全域運動估測及訊源與通道結合編碼影響之研究

Abstract

雖然Shannon證明訊源編碼與通道編碼可以在不影響整個系統的效益下分別處理。但是在實際應用上，由於無法建立正確的訊號源與通道模組，導致無法求得最佳的訊源編碼以及無錯誤(error-free)的通道編碼。因此，上述Shannon的證明就無法適用。本論文主要是研究訊源編碼與通道編碼的雜訊影響。吾人先分析通道雜訊(channel noise)對視訊(video)編碼架構的影響，並發現若通道編碼無法維持無錯誤的環境，會使得VLC(Variable Length Code)codec產生連續的錯誤(burst error)，進而導致額外的量化損失(quantization distortion)。為了降低通道雜訊所造成的損失，吾人亦提出一適用於雜訊通道之訊源與通道結合編碼的方法。本論文分為以下三大部份. 在論文的第一部份中，吾人提出一全域運動預估的方法用來降低影像中時序(tempral)上的冗餘部份。此種預估法也可用來將影像中移動部份區分出來。此種全域運動預估的方法是利用物體投影在攝影機的幾何關係，推導出其數學模型。此外，由於攝影機變焦(zooming)動作會導致影像中的物體變形(deformation)，吾人提出可變區塊運動預估法來預估影像中有變焦動作的運動向量。 在論文的第二部份中，根據線性區塊編碼(linear block code)的量分布(weight distribution)以及傳送字碼與接收字碼間的Hamming distance，吾人可推導出此區塊編碼在二位元對稱通道(binary symmetric channel)上位元錯誤機率的上限值。一般量化器設計中，往往忽略掉經過通道編碼後通道的特性，以致有額外的量化損失。為了改善這種現象，吾人根據之前所求得的位元錯誤機率的上限值位元，提出了一個結合線性錯誤控制碼的量化器設計。 在論文的第三部份中，吾人提出一個適用於衰減通道(fading channel)的量化器設計。首先，吾人推導出此衰減通道經PSK或FSK調變之後的錯誤機率模組，再根據此模組設計出適合於衰減通道的量化器。在和傳統的訊源與通道分散編碼架構比較下，吾人所提的訊源與通道結合編碼架構有較佳的表現。此外，吾人亦將所設計的量化器應用在轉換編碼的架構中。實驗的結果也顯示出吾人所提出的轉換編碼架構可得到較佳的效益。

In Shannon's theory, source coding and channel coding can be treated separately without sacrificing the overall optimality. In a practical system, it may not be to identify the source and the channel models perfectly; thus, the theory may not be valid if either of the following two situations occur: (a) the source coder is sub-optimal, or (b) the channel coder cannot achieve the error-free condition. In this thesis, we study the noise effects on the combined source and channel coding and present a few combined coding algorithms for noisy channels. This thesis is divided into three parts. The first part describes a global motion parameter estimation method. This method can be used to segment an image sequence into objects of different motion. For any two image pixels belonging to the same moving object, constrained by the image projection geometry, their global motion components are bounded by a fixed relationship. Therefore, by examining the measured motion vectors we are able to group pixels into objects and, at the same time, identify the global motion parameters. Furthermore, because the block shape is distorted due to camera zooming, a deformable block motion estimation scheme is suggested to recover the object local motion vectors.\\ \indent In the second part, given the weight distribution of a linear block code and the weight of the Hamming distance between a transmitted codeword $v_t$ and a decoded codeword $v_d$, we derive the error probability that the transmitted codeword $v_t$ is decoded to $v_d$. Our new method can estimate the upper bound of the bit error probability in the case of the linear block code used for the binary symmetric channel. Most existing quantizer designs do not take into account the channel characteristics. Based on the estimated upper bound, we propose a combined quantizer and linear error control code design for noisy channels.\\ \indent In the third part, we present a quantizer that achieves the best overall performance when its outputs are transmitted over a fading channel. First, a probalistic model describing a fading channel with binary PSK or FSK modulation is derived. Then, we propose a procedure of designing the optimal quantizer for the slow fading channel by extending a previous work on the combined source/channel quantizer design. Next, we look into the structure of our quantizer to find the theoretical grounds behind its superior performance. We also compare this combined source/channel coder against the conventional separated source/channel coder and identify the preferred operating regions of these two systems. A transform image coding system over a fading channel is designed based on the preceding principles. Simulations indicate that our quantizer outperforms the channel-error-specific optimal quantizer, particularly when the channel error probability is not precisely known. Chapter 2. Global Motion Estimation Chapter 3. Combined Quantizer and Error Control Code Design for Noisy Channels Chapter 4. Optimum Quantizer Design for a Slow Fading Frequency Nonselective Channel Chapter 5. A Transform Image Coder over Noisy Channels Chapter 6. Conclusions