基于小波变换和残差处理的人脸图像超分辨率技术研究

Research of super-resolution in facial image restoration based on wavelet transform and residuals processing

超分辨率图像复原技术常见的有插值法,POCS等,它们有细节模糊,运算复杂度大的弱点,针对以上问题,在基于学习的超分辨率图像复原中,提出了一种全新的小波系数特征向量匹配方法。算法分为两步:(1)采用基于补偿残差向量和多样本平均的低分辨率人脸图像的小波特征向量匹配及人脸图像复原。(2)用边缘提取和特定区域平滑的方法去除Gibbs效应等噪声。经实验和传统的插值法以及常规匹配方法比较,在细节复原和运算复杂度方面都有一定的提高。

Common super-resolution image restoration technologies are the interpolation methods, POCS and so on, their weaknesses are vague details, complex computing, based on Learning-based super-resolution, against the problems above, this paper proposes a new wavelet coefficients eigenvector matching method. This algorithm is divided into two steps： （1） Using residual vector compensation and sample average methods to match low-resolution facial images＇ wavelet eigenvector and recovery facial images. （2） Using edge detection and specific regional smoothing methods to remove the Gibbs effect and other noise. Comparing with the traditional interpolation methods and conventional matching methods, the details of the recoveries and the complexity of computing has indeed improved.