零樹小波視訊編碼之錯誤偵測與隱藏

摘要

我們都知道,在以零樹編碼為主軸的系統中,在所編出來的符號間均具有絕對的對映關係;所以在傳輸位元流時,一但在位元流中有一個錯誤發生時,則因其後之所有資料均會失去對映性而導致所有的資料均無法為系統重建影像時所使用。如EZW、ZTE即是。由此可知;零樹編碼對於通道錯誤是相當敏感的。
 

所以在本論文中,我們提出一套“零樹小波視訊編碼之錯誤偵測與隱藏“系統來針對零樹編碼對於通道錯誤的敏感性做偵測與補藏。而在此我們所用的錯誤偵測技術,可分為三部份來談。1) 序列數值碼 (Sequence Number Code) 錯誤偵測技術。2) 零樹小波視訊編碼之錯誤偵測技術。3) 移動向量之錯誤偵測。
 

其中序列數值碼錯誤偵測技術,主要是利用在數個區塊位元流間加入一個序列數值碼,然後在解碼端再以這些數值碼間的碼值連續性關係來偵測出在位元流中是否有錯誤發生及錯誤所發生的位址。而零樹小波視訊編碼之錯誤偵測技術,則是利用零樹編碼架構中,各符號間的數值對映性來判斷是否有錯誤發生。另外,移動向量之錯誤偵測部份亦是利用移動向量的數值特性來作偵測判斷。利用上述的三重錯誤偵測技術再加上錯誤隱藏後,我們發現在我們所提出的系統中,當錯誤率在高於3E-4之後,系統所得的重建影像品質會比H.263的原始架構所得的結果來的好。


 
 

Error Detection and Concealment of Zerotree Wavelet Video Coding

Abstract


 

In the zerotree wavelet video coding system, there always exist direct correspondences between the coded symbols. Therefore, if one error occurs in the received bitstream, the rest of decoded data will lose their corresponding relationships and will not be able to reconstruct the original image. From this, we observe that the zerotree coding is very sensitive to transmission errors.
 

In this thesis, we discuss how to improve the performance of the zerotree coding by using the error detection and concealment techniques at the decoding end in error-prone environments. The proposed error detection technology can be categorized into three parts: 1) sequence number code error detection, 2) zerotree mapping error detection, 3) motion vector error detection.
 

The sequence number code error detection adds a sequence number code between an integer number of blocks. Accordingly, we can detect errors by examining the consecutiveness of these sequence number codes. As for the zerotree mapping error detection, it uses zerotree’s coding structure to detect errors through the code correspondence. Furthermore, the motion vector error detection checks the valid range of motion vectors to identify errors. With the above mentioned error detection strategies and the additional error concealment, e.g., the temporal replacement in the wavelet domain, the reconstructed image quality is better than the default H.263 at the error rate above 3e-4.