本論文針對Chang et al. [1] 所提之基於鄰居內嵌法的超解析演算法做了改進。基於鄰居內嵌法的超解析法在大量的訓練樣本中找K個最近鄰居是非常耗時的，因此我們對訓練樣本先進行分群。此後對輸入區塊只要找尋最相似之群中心再至該群中找出K個最近鄰居，如此便可節省大量的計算時間。不同於Chang et al.的方法是以歐式距離來找K個最近鄰居，我們是以自行定義的相似度Similarity來找K個最相似鄰居，再用LLE (Local linear embedding) [2]的方法求出最佳組合係數，最後利用此組合係數對K個低解析區塊(即K個最相似鄰居)之相對高頻資訊區塊求得其線性組合，再將此組合的高頻資訊區塊加在對輸入區塊升頻取樣所得的放大區塊上此即為所求之高解析區塊。實驗證明本論文所提之分群機制的確能在不太影響超解析效果之下節省大量計算時間。 In this paper, we propose an improved version of the neighbor embedding super-resolution (SR) algorithm proposed by Chang et al. [1]. Because neighbor embedding SR algorithm requires intensive computational time when finding the K nearest neighbors for the input patch in a huge set of training samples. We tackle this problem by clustering the training sample into a number of clusters, with which we first find for the input patch the nearest cluster center, and then find the K nearest neighbors in the corresponding cluster. Different from Chang’s method that uses the Euclidean distance to find the K nearest neighbors of a low-resolution patch, we define a similarity function and use it to find the K most similar neighbors of a low-resolution patch, then use LLE (Local linear embedding) [2] to find optimal coefficients, with which the linear combination of the K most similar neighbors best approaches the input patch. These coefficients are then used to form a linear combination of the K high-frequency patches corresponding to the K respective low-resolution patches (or the K most similar neighbors), the resulting high-frequency patch is then added to the enlarged (or up-sampled) version of the input patch. Experimental results show that the proposed clustering scheme efficiently reduces computational time without significantly affecting the performance.
