距离选择在超分辨率图像字典优化中的应用

he quality of Learning-based super-resolution (SR) largely depends on whether the dictionary includes rich details which are strongly similar to the input image. In this paper, for each input image, we optimize the dictionary by selecting the most similar training images based on the SIFT feature. A rich set of candidate distance metrics, such as Median distance, Euclidean distance, Harmonic distance and Geometric distance, are explored to match the SIFT feature, which can provide more accurate matching result than only using Euclidean distance. The optimal distance of each target image can be selected by comparing the four super-resolution images, which are recovered respectively by using the element of distance set to select the training image. According the optimal distance metric, all target images are divided into four groups. We propose the bag-of-words (BOW) model for quantize the input image and target images, and then select the optimal distance of input image by finding the most similar target image. Experimental results demonstrate that our method can obtain the high hit rate and the select distance is more adaptive for input image than the traditional Euclidean distance.