基于学习超分辨率重建中的样本选择方法

Sample selection method for learning-based image super-resolution

提出一种人脸图像超分辨率重建(Super-Resolution Reconstruction,SRR)的自适应学习样本选择方法。利用局部保持投影(Locality Preserving Projections,LPP)算法的局部保持能力,在人脸图像局部流形上分析其非线性结构特征,并给出了LPP变换向量的数值解法。在LPP的特征空间中动态搜索学习样本,即选择出与输入图像块最为相似的像素块集合。利用选择出的样本通过基于像素块的特征变换法完成超分辨率重建。实验表明,自适应样本选择方法可以快速、有效地选择出少量学习样本,具有良好的图像高频信息复原能力。

This paper presents an adaptive learning sample selection method for face hallucination.The nonlinear structural features of face images are explored on facial local manifolds using Locality Preserving Projections（LPP） algorithm,and the efficient computation method of the transform vectors of LPP is presented.Learning samples are dynamically picked out in the eigen-space of LPP,i.e.,the patch set most similar to the input image patch.The selected samples are used for super-resolution reconstruction by the patch-based eigen-transformation method.Experimental results fully demonstrate that the proposed adaptive sample selection method can fast and efficiently select out a small amount of learning samples,with good reconstruction performance in terms of high-frequency information restoration.