| 摘 要: 针对大多数人脸超分算法的预设退化模型与真实图像的退化方式差距大，导致人脸重建的效果不理想的问题，提出了一种针对真实图像退化的人脸超分辨率重建算法。首先设计了一种混合退化模型，通过对运动模糊、高斯噪声等多种退化形式进行合成用于模拟真实图像退化空间，生成接近现实场景的低分辨率图像。然后采用基于小波域的超分辨重建网络预测得到高分辨率图像的小波系数，并经过小波逆变换得到超分辨率图像。在高清人脸数据集(FFHQ)和真实人脸数据集(RealSR)上的实验结果表明，研究提出的算法不仅能有效提升重建效果，而且适用于真实场景下的人脸超分辨率重建。 |
| 关键词: 退化模型；人脸图像；超分辨率重建；小波变换；卷积神经网络 |
|
中图分类号: TP391
文献标识码: A
|
| 基金项目: 湖南省教育科学“十三五”规划2019年度立项课题一般资助课题(XJK19BXX009). |
|
| Face Super-resolution Reconstruction Algorithm based on Real Image Degradation |
|
LI Jieqin, XIE Dingfeng
|
(School of Information Engineering, Hunan Industry Polytechnic, Changsha 410000, China )
Lijieqin4568@163.com; coolboyxie@163.com
|
| Abstract: A large gap between the preset degradation model of most face hyper-resolution algorithms and the degradation of real images leads to the poor effect of face reconstruction. In view of this problem, this paper proposes a face super-resolution reconstruction algorithm for real image degradation. Firstly, a hybrid degradation model is designed, which combines motion blur, Gaussian noise and other degradation forms to simulate the real image degradation space and generate low-resolution images close to the real scene. Then, the wavelet coefficients of the high-resolution image are predicted by the super-resolution reconstruction network based on wavelet domain, and the super-resolution image is obtained by inverse wavelet transform. Experimental results on FFHQ (Flickr-Faces-HQ) and RealSR datasets show that the proposed algorithm not only effectively improves the reconstruction effect, but also is suitable for face super-resolution reconstruction in real scenes. |
| Keywords: degradation model; face image; super-resolution reconstruction; wavelet transform; convolution neural network |
