臺灣博碩士論文加值系統

在本論文之研究中 ，一種新的擷取影像重現背景資訊(uncovered ba-
ckground information) 的方法被發展出來 。藉此方法 ，當影像傳送時
，其背景部份僅傳此資訊 ，用此資訊去更新上一個影像的背景部份 。吾
人利用重現式背景預測器(Uncovered Background Predictor或 U B P)來
達成重現背景資訊的擷取工作 。然後再將此 U B P 和傳統之移動補償預
測器(Motion-Compensating Predictor或 M C P)結合 。其中在 M C P
部份 ，其移動向量的估計採區塊匹配式(block matching)演算法 。而編
碼部份 ，先求出欲編碼影像的亮度特性分佈圖(histogram) ，然後再以
霍夫曼碼(Huffman code) ，完成編碼工作 。另外在 U B P 方面吾人使
用截斷式循序機率比檢定法(Truncated Se- quential Probability
Ratio Test或 T S P R T)之假設檢定(hypothesis test)技術決定影像中
那一部份屬於背景(background) ，那一部份屬於主體(object) 。在做
T S P R T 時，其中概率比(likelihood ratio)為主體和背景的機率密度
函數(Probability Density Functions或 P D Fs)的比值 。而主體和背
景的 P D F 則採機率性類神經網路(Probabilistic N- eural Networks
或P N N )架構之非參數估計器(nonparametric estimato- rs)予以估計
。最後將原始影像和經 U B P 和 M C P 重建的影像做差值脈衝編碼調
變(Differential Pulse Code Modulation或 D P C M ) ，然後編碼送
出 。由實驗結果中知本論文之重現背景影像編碼方法比現有的方法更可
準確的做主體與背景間的適應性判斷 ，並配合霍夫曼編碼可進一步的增
進動態影像編碼效益 。且其演算法可以 P N N 架構實現出來 ，而達到
快速運算的目的 。

A new scheme is developed to grab uncovered background infor-
mation. According to this scheme the uncovered background is
tra- nsmitted to the receiver to update the background memory
in visu- al communication. The uncovered background information
is obtained by using the uncovered background predictor(UBP),
which is then combined with the motion-compensating predictor(
MCP). The estimate of the move- ment for the MCP is obtained by
employing the block matching alg- orithm. In the encoding
procedure, the histogram of the gray lev- els is derived for
the implementation of the Huffman codes. For the UBP, the
hypothesis testing technique with Truncated sequential
probability ratio test(TSPRT) is employed to decide t- he
background or object information in a image scene. In the TSP-
RT, the likelihood ratio function is formed with the
probability density functions(PDFs) of the object and the
background, in whi- ch the PDFs are estimated via probabilistic
neural network(PNN). Then, the differential pulse code
modulation(DPCM) is employed f- or encoding based on the
original image and reconstructed on with UBP and MCP.
Experimental results illustrate that the developed scheme for
uncovered background video image coding can adaptively
discrimin- ate between the object and the background signals of
consecutive image frames more accurately than the existing
methods. Furtherm- ore, the discrimination algorithm can be
implemented by using the PNN architecture, with which the
scheme for speedy video image c- oding scheme can be developed.