基于学习的人脸图像超分辨率重建方法

Learning-based super-resolution reconstruction of face image

提出一种针对正面人脸图像的超分辨率重建方法,通过学习人脸图像梯度的空间分布特性,获取梯度先验知识;通过结合贝叶斯最大后验概率估计理论,采用最速下降优化方法得到高分辨率人脸图像。实验结果表明,该方法在仅输入2—3幅低分辨率图像的情况下即可重建出具有较佳高频细节的超分辨率图像。

Propose an algorithm that learned a prior on the spatial distribution of the images gradient for frontal images of faces,and then incorporated such a prior into the Maximum a Posterior Estimate （MAPE） algorithm.To get a single global minimum,use a gradient descent algorithm.Experimental results demonstrate that the proposed method is effective in reconstructing the high-frequency components of the images.Furthermore,the method can convert a small number（2-3） of low-resolution images of a face into a single high-resolution image.