| 摘要: |
| 针对人脸超分辨率算法中图像失真大、缺乏细节特征等问题，提出了一种基于先验知识的人脸超分辨率重建模型。通过在超分网络中加入纹理辅助分支，为重建过程提供额外纹理结构先验，以生成精细的面部纹理，恢复高分辨率纹理图。同时引入级联叠加模块对纹理辅助分支进行反馈。设计特征融合模块，将纹理特征图与超分分支特征图融合，获得更好的纹理细节；将纹理损失融入损失函数，以提高网络恢复纹理细节的能力。4倍放大因子下，该方法的峰值信噪比(Peak Signal-to-Noise Ratio，PSNR)、结构相似性指数(Structural Similarity Index，SSIM)比现有方法至少提升1.082 5 dB 和0.036，无参考图像质量评价(Natural Image Quality Evaluator，NIQE)至少降低1.690 2；8倍放大因子下，该方法的PSNR与SSIM值分别至少提升0.787 5 dB和0.046 85，NIQE值最小降低3.92。 |
| 关键词: 人脸超分辨率重建 先验知识 生成对抗网络（GAN） |
| DOI：10.20079/j.issn.1001-893x.220110003 |
|
| 基金项目:国家自然科学基金资助项目（62001033，62201066） |
|
| A face super-resolution reconstruction method based on prior generative adversarial network |
| DU Yansong,CAO Lin,DU Kangning,SONG Peiran,GUO Yanan |
| (School of Information and Communication Engineering,Beijing Information Science and Technology University,Beijing 100101,China;a.School of Information and Communication Engineering;b.Key Laboratory of the Ministry of Education for Optoelectronic Measurement Technology and Instrument,Beijing Information Science and Technology University,Beijing 100101,China) |
| Abstract: |
| To relieve the large image distortion and detailed feature missing problems in face super-resolution algorithms,a face super-resolution reconstruction model is proposed.This model adds an auxiliary texture branch to the super-resolution network and restores the high-resolution texture map through the texture branch,which provides additional texture structure priors for the reconstruction process to generate fine facial textures.A cascaded stacking module is introduced into the super-resolution network to feed back the auxiliary texture branch.In addition,a feature fusion module is designed to fuse the texture feature map with the super-resolution feature map to obtain better texture details.The texture loss is considered in the loss function to improve the texture detail reconstruction ability of the model.Experimental results show that the proposed methods peak signal-to-noise ratio(PSNR) and the structural similarity index(SSIM) is increased by 1.082 5 dB and 0.036,the natural image quality evaluator(NIQE) is decreased by 1.690 2 when the upscale factor is 4.The minimum PSNR value and SSIM value are increased by 0.787 5 dB and 0.046 85,respectively,and the NIQE value is decreased by 3.92 under the 8 upscale factor. |
| Key words: face super-resolution restruction prior knowledge generative adversarial network(GAN) |
