轉換編碼影像之近乎最佳後級處理法

Abstract

後級處理是在低率傳輸中可增進高失真影像品質的方法之一。後級處理， 是建立在接收端的處理單元，純粹利用傳輸端所送來的影像資料，降低編 碼誤差，以產生較好的輸出品質。此外，由於一般影像標準大多採用轉換 編碼的方式，且編碼誤差出現轉換域中，故在此域中作影像還原，將可獲 得較好的效果。在本論文裡，我們提出兩種在轉換域的影質改善方法，即 維尼濾波法，和誤差樣本補償法。我們將轉換係數之量化誤差視為白色雜 訊，並利用方形區塊編碼相鄰區塊間轉換係數的相關性，設計出一適應性 維尼濾波器，藉以除去高頻率的頻帶外雜訊，以消除影像的方塊效應。誤 差樣本補償法則是針對區塊中轉換係數間的殘餘相關性而設計的。基於這 種相關性，我們便可藉由低頻係數之資訊預測出高頻的量化誤差，經補償 後，高頻係數可獲得還原。比較處理前、後的影像，在平滑區的方塊效應 ，邊緣區的鋸齒效應，以及紋路區的模糊現象均可降低，SNR 值亦有提升 。 When the channel capacity cannot afford high-rate transmission of data, a highly compressed image is severely degraded. In this situation, a post-processing at the receiving end, can be used to improve the image quality merely operating on the received data. Because the coding errors occur in the transform domain for the popular transform-based coding schemes, it is cost-saving to implement the post-processor in the transform domain. Two transform-domain post-processing schemes are proposed in this thesis; namely, Wiener Filtering and Error Pattern Compensation. In the former approach, we take the advantage of the inter-block correlation among DCT coefficients to design the linear minimal mean square error (LMMSE) filter to reduce blocking effect. In the latter, the residual correlation among the coefficients inside a block is utilized to estimate the quantization error patterns. Again the estimators are LMMSE filters. In addition, instead of one single filter, multiple filters are used to match the image local characteristics. We can achieve noticeable improvement in both subjective and objective measures, particularly on the highly compressed images.