傳統上,一組編碼簿在經長久使用後,有些圓形會因為代表性的不足,使得還原影像變得嚴重失真,而解決代表性不足的方法是透過不同類型、更多的訓練向量來訓練編碼簿,以找出代表性較佳的碼向量。經研究後,我們發現化簡後的影像,其影像區塊有許多相似之處,籍由所記錄下來的像素平均值與偏差,我們僅需用同一碼向量便能有效的代表這些不同的原始影像區塊。如此就可免除編碼簿需大量訓練的問題。藉由以上概念,我們設計出以“機率"為基礎的編碼簿,試著由合理的機率分配,找出能代表大部分影像化簡區塊的編碼簿,進而解決編碼簿需不斷訓練的問題。 After being used for a long time, some images in the codebooks are incapable of returning to the original ones owing to the representative insufficiency. Traditionally, to solve the problem
is to find representative code words by other kind of training vectors. In this paper, we find out that blocks of simplified images among different original ones are similar. Therefore, by calculating the average of pixels and deviation, we can use few code words to represent variety
of original image blocks. Meanwhile, the problem that code words need to be trained again and again can be solved. In accordance with the above concepts, we design codebooks based on "probability". With reasonable distribution, we try to discover a codebook, which can represent
most of simplified image blocks, and further, solve the problem of codebook training.
關聯:
第十三屆國際資訊管理學術研討會論文集(I)=Proceedings of the 13th International Conference on Information Management(I),頁527-532