融合参考图像的人脸超分辨率重构方法

Face Super-Resolution Reconstruction Method Fusing Reference Image

基于深度学习的图像超分辨率重构方法对低分辨率人脸图像进行超分辨率重构时,通常存在重构图像模糊和重构图像与真实图像差异较大等问题.基于此问题,文中提出融合参考图像的人脸超分辨率重构方法,可以实现对低分辨率人脸图像的有效重构.参考图像特征提取子网提取参考图像的多尺度特征,保留人脸神态和重点部位的细节特征信息,去除人脸轮廓和面部表情等冗余信息.基于提取的参考图像多尺度特征,逐级超分主网络对低分辨率人脸图像特征进行逐次填充,最终重构生成高分辨率的人脸图像.在数据集上的实验表明,文中方法可以实现对低分辨率人脸图像的有效重构,具有良好的鲁棒性.

While low-resolution face images are reconstructed via deep learning based super-resolution reconstruction method,some problems emerge,such as blurred reconstructed images and obvious difference between reconstructed images and real images.Aiming at these problems,a face super-resolution reconstruction method fusing reference image is proposed to reconstruct low-resolution human face images effectively.The multi-scale features of reference image are extracted by reference image feature extraction subnet to retain the detail information of key parts and remove the redundant information,such as facial contour and facial expression.Based on the multi-scale features of reference image,the step-by-step super-resolution main network fills the features to low-resolution face image step by step.Finally,the high-resolution face image is generated.Experiments on datasets indicate that the proposed method reconstructs low-resolution face images effectively with good robustness.