基于边缘增强生成对抗网络的人脸超分辨率重建

Face hallucination based on edge enhanced generative adversarial network

针对对抗生成神经网络在人脸轮廓细节恢复上不够完善的问题,利用人脸图像的结构先验信息提出了一种边缘增强的生成对抗网络人脸超分辨率的重建算法.首先,利用人脸图像及其边缘图像的一致性关系设计一种并行网络提取面部和边缘细节特征;然后,通过特征融合网络获得高分辨率的生成图像;最后,利用判别网络判别生成图像的真伪.在人脸图像数据库上进行的人脸超分辨率重建实验结果表明:提出的边缘增强生成对抗网络能够提升面部细节重建能力,主观和客观评价指标均优于现有的人脸超分辨率算法.

Aiming at the imperfection of the countermeasure generation neural network in the restoration of facial contour details,an edge enhancement generation countermeasure network was proposed to enhance the super-resolution reconstruction performance of human face based on the prior structural information of face images.First,a parallel network was designed by using the consistency relationship between face images and their edge images.The network extracted facial and edge detail features,and then high-resolution generated images were obtained by feature fusion network.Finally,the authenticity of generated images was distinguished by discriminant network.Experimental results of face super-resolution reconstruction on face image database by the proposed algorithm show that the proposed edge enhancement generates confrontation network can improve the ability of facial detail reconstruction,and the subjective and objective evaluation indexes are superior to the existing frontier face super-resolution algorithms.