基于生成对抗网络的人脸超分辨率重建方法  

Facial super-resolution reconstruction method based on generative adversarial networks

作　　者：张曦[1] 库少平[1] ZHANG Xi;KU Shao-ping(School of Computer Science and Artificial Intelligence,Wuhan University of Technology,Wuhan 430070,China)

机构地区：[1]武汉理工大学计算机与人工智能学院,武汉430070

出　　处：《吉林大学学报(工学版)》2025年第1期333-338,共6页Journal of Jilin University:Engineering and Technology Edition

基　　金：工业与信息化部重大专项项目(工信部装函[2017]614号)。

摘　　要：针对人脸重建图像峰值信噪比和结构相似度低的问题,提出了一种在生成对抗网络基础上的人脸超分辨率重建方法。创建一个两分支生成器网络,运用网络内的特征提取模块得到图像浅层特征,重建图像高频分支,利用深层残差网络训练低分辨率图像高频数据,映射人脸图像数据;停止更新固定生成器参数,使用相对判别器网络辨别人脸图像重建真伪,引入注意力机制优化生成对抗网络重建性能,通过感知损失函数评估重建可靠性,实现人脸超分辨率重建。实验结果表明:本文方法能有效提升面部细节重建能力,主客观评价指标都较为优秀,其中主观评价结果显示该方法重建的人脸图像具备更清晰的细节数据,人脸轮廓更为立体,失真率较低,拥有更为自然和真实的外在特征;客观评价结果显示该方法的峰值信噪比达到44.6 dB,结构相似度达到0.74,充分说明该方法具有较强的鲁棒性,可为优化人脸识别应用性能提供有效借鉴。In order to solve the problem of low peak signal-to-noise ratio and structural similarity of reconstructed face images,a super-resolution reconstruction method based on generative adversarynetwork is proposed.A two-branch generator network is created,the shallow features of the image are obtained by using the feature extraction module in the network,the high-frequency branches of the image are reconstructed,the high-frequency data of the low-resolution image is trained by the deep residual network,and the face image data is mapped.Stop updating fixed generator parameters,use relative discriminator network to identify the authenticity of face image reconstruction,introduce attention mechanism to optimize the reconstruction performance of generative adversarial network,evaluate the reconstruction reliability by perception loss function,and realize the super-resolution reconstruction of face.The experimental results show that the proposed method can effectively improve the ability of facial detail reconstruction,and the subjective and objective evaluation indexes are excellent.The subjective evaluation results show that the reconstructed facial images with the proposed method have clearer detailed data,more three-dimensional facial contours,less distortion rate,and more natural and real external features.The objective evaluation results show that the peak signal-to-noise ratio of the proposed method reaches 44.6d B and the structural similarity reaches 0.74,which fully demonstrates the robustness of the proposed method and provides an effective reference for optimizing the application performance of face recognition.

关 键 词：生成对抗网络 高频分量 超分辨率 判别器 损失函数 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]