利用位置字典对的人脸图像超分辨率方法

摘要

A novel method using position-dictionary pairs is proposed to reconstruct a high resolution frontal face image from a single low resolution version. Since the contents and structures in the same position of face image patches are similar and are more likely to be better represented by the combination of the same dictionary atoms, face images used for training are blocked into overlapped position-patches. Then a small position-dictionary pair is trained for each position. A basic high resolution image can be recovered for each low resolution input by using these position-dictionary pairs. Finally, the visual quality of the reconstructed image is improved by using a residual compensation method. Experimental results show that face images reconstructed by the proposed method have better visual effect. A comparison with the algorithm of image super-resolution via raw image patches based sparse representation show that the SSIM of the proposed method is 0. 082 higher; while the training time is shortened about 5 times.%针对基于稀疏表示的人脸超分辨率算法存在的字典尺寸大、训练时间长等问题,提出一种基于位置字典对的超分辨率重建方法.由于同一位置的人脸图像块具有相似的结构和内容,更有可能用相同的字典原子进行线性组合表示,因此把训练人脸图像按位置分块,首先为每个位置训练一个位置字典对,利用获得的多个位置字典对,对低分辨率测试人脸图像进行基本重建,然后应用残差补偿方法对位置块进行补偿.实验结果表明,由所提方法重建的人脸图像具有更好的视觉效果,与应用原始图像块进行稀疏表示的图像超分辨率算法相比,平均图像结构相似度指标值提高了0.082,同时字典训练时间缩短了约5倍.