用於幀間預測之以相似度最大化為基礎之混合式匹配方法

Abstract

在使用區塊匹配演算法的雙向預測中，區塊匹配演算法被用來找尋和目標區塊最接近的二個參考區塊，這二個參考被融合成一個區塊，而這個區塊會比由匹配演算法找到的二個區塊還要像目標區塊。因此由多個參考區塊融合而成的區塊會比單個參考區塊還要更接近目標區塊。 在這篇論文裡, 我們也使用多個參考區塊來產生目標區塊的預測值。我們使用模板匹配演算法來尋找多個參考模板和參考區塊，並基於奇異值和最小的方法以這些參考模板和參考區塊的資訊來產生目標區塊的預測值。因為參考模板和參考區塊是由模板匹配演算法找到的，移動向量的資訊可以不被包含在位元串流中。

In the bi-direction inter prediction of block matching algorithm, block matching algorithms are used to find two block predictors which are similar with the target block and these two block predictors are combined as one block predictor. This predictor can be more similar with the target block than the two predictors found by block matching algorithm. Therefore, the predictor generated by multiple candidates can be better than the predictor generated by only one candidate. In this paper, we generate the predictor from multiple predictor candidates too. We find several template candidates and block candidates using template matching algorithm. Then we generate the target block predictor according to the information of template candidates, block candidates and target template based on minimization of the sum of the singular values. Because the block candidates and template candidates are found using template matching algorithm, the information of motion vector does not need be encoded in the bit-stream.