关键词: 人脸图像, 超分辨率, 级联神经网络, 面部先验, 关键点

Abstract:

A two-layer cascade neural network is proposed for face super-resolution to solve the problems of the insufficient use of facial prior, facial features shifting and edge blurring in common deep learning based super-resolution methods. A facial prior estimation module is used in the net to capture the landmark information of the input and constrain the spatial consistency of the target image with the reconstructed image. Extensive experiments over CelebA and Helen datasets demonstrate that the proposed method is capable of accurately reconstructing facial features on frontal faces, and is also robust to different facial variations, such as side and occlusion face.

Key words: face image, super-resolution, cascade neural network, facial prior, landmark