一种基于金字塔策略的人脸超分辨率方法

An Approach to Hallucinating Faces Based on Pyramid Strategy

高分辨率预期图像与低分辨率输入图像分别位于高斯金字塔的第0层和第i层.超分辨率过程据此被分成i-1个子过程,逐层进行,即将第k层低分辨率图像的超分辨率估计图像(位于第k-1层)作为第k-2层高分辨率预期图像的输入,直至产生第0层高分辨率估计图像为止.在每个超分辨率子过程中,采用基于patch的搜索策略.文中对这种基于金字塔策略的人脸超分辨率方法进行了实验,结果表明该方法能够达到很好的超分辨率效果.

High-resolution prospective image and low-resolution input image are respectively the 0^th and i^th level of the Gaussian pyramid. Thus, super-resolution is divided into i-1 processes and implemented level by level. ＂That is to say the estimate image of the k^th-level low-resolution image which located in the k-1^th level will be regarded as the input of the k-2^th-level high-resolution prospective image. The repetition ends till the 0^th level high-resolution estimate image is hallucinated. The searching strategy based on patch is used in each level. We demonstrate this pyramid based approach can achieve high quality hallucination.