Abstract
在近年來, 數位影像序列的編碼技術已成為注目的焦點.由於移動補償式混合編碼器(MC-DPCM Hybrid Coder)能夠提供極高的壓縮比. 因此H.261, MPEG 及近來之全數位高畫質電視均採用移動補償式混合編碼器,可知其將為編碼器之主流. 由於移動補償式混和編碼器使用了可變長度編碼, 故其每個區塊的位元率相差很多, 必需加入一個緩衝器來吸收編碼器輸出位元率之變動. 故本篇論文所探討的主題即為此編碼器之緩衝器控制.緩衝器控制之主要目的為控制編碼器之輸出位元率, 避免緩衝器溢位(overflow)或不足(underflow),並使輸出位元率及重建的影像序列品質保持在一定的範圍, 使得人眼無法分辨出來重建的品質的跳動, 而看起來覺得舒服.本篇論文將提出一種緩衝器控制的方法, 首先利用模糊分類法則(fuzzy classification)把每個區塊分類, 再依其對人類視覺系統的靈敏度, 調整量化器, 控制分配多少位元給此區塊, 保持所產生的誤差能予以接受.本模糊分類法則是利用模糊集合(Fuzzy Set) 的觀念於叢集化(Clustering), 其中心部份為一模糊關聯記憶體(FAM)系統. 此系統之輸入為影像區塊之特徵(Features), 再將這些特徵的值轉到其對映的模糊集合之成員函數(membership function), 然後利用事先建好的IF-THEN 規則來判斷此一區塊可能屬於那一種類別, 再收集各規則之結果做解模糊化, 求出此一區塊之類別. 將其與緩衝器之狀況合併起來, 做判斷以決定此區塊應如何量化 (由量化器之量化因子控制),再以可變長度編碼器編碼.即可決定對於此一區塊應如何在位元率與品質的考慮下, 做最佳化編碼.Adaptive Quantization and Buffer Control are key components ofthe digital image compression technology where both storagespace and channel bandwidth are very limited. Because of thevariable length coding that a bit rate near its source entropyis applied to a constant rate channel, therefore a buffer isneeded to regulate the output bit rate of the coder. In thesimple buffer control scheme, the value of Qp is only dependenton the buffer fullness. On the other hand, we may combine thebuffer control with adaptive quantization so that betterperformance at a given bit rate can be provided. In thisadaptive coding system, a classifier is used to distinguish thesignificance of a block. We will use a classifier that islinked with the fuzzy set theory in the coder,i.e., a FuzzyAssociative Memory is used to perform classification byprescribed rules. In this thesis, we propose a modified featureR instead of the well-known feature L that represents the low-frequency ac energy L. The feature R can provide a good measureto the significance of the low-frequencies in a macroblock andit is less sensitive to the different video sequences. Afterclassification, a quantization-parameter string that contains astring of Qp to each class is used to encode a frame at a givenbit rate. To different fullness of buffer, we use one ofquantization-parameter strings to encode the blocks of a frame.The simulation results show that the performance of theproposed scheme is better than nonadaptive methods. Especially,for sequences with large amount of motion, the nonadaptivemethod is very poor, and the distinction between them isobvious.