ベタトル量子化と多様体学習を用いた超解像技術

摘要

近年,画質改善の方法として超解像技術が注目されている.その中でも単一の低解像度画像から高解像度 画像を生成するSingle-Frame SRは，入力する低解像度画像とは別に，大規模なデータベースが必要である.しかし，従来まではデータベース内の学習データが冗長となる問題点や，パッチを探索する計算時間が膨大となる問題点があった.そのため,本研究ではk-means法を用いたベタトル量子化によりデータベースの冗長性を低滅し，計算時間の高速化を実現した.%Image Super-Resolution (SR) is to recover the lost high-frequency information from several or only one available image. Single-Frame SR, one of hot topics in SR research fields, can generate a high-resolution image from only one low-resolution image by using the prior prepared database. Therein, the example-based and neighborhood embedding-based SR are the very popular single-frame SRs to infer the lost information in the LR input with the known corresponding relations between LR and HR images in database which has to be prepared in large-scale for having most varieties of image, and then take alot of computational time for inferring. Therefore, this study proposes to first obtain some prototypes from the prepared LR and HR images using vector quantization such as k-means clustering method, and the achieved prototypes are as the training database for inferring the lost information of any LR input. Then, the amount of corresponding LR and HR data in training database can be greatly reduced, which guarantee much less computational time. Experimental results also show that our proposed strategy can achieve higher quality high-resolution image and lower computational time than conventional methods.