一种用于联合低光增强和人脸超分的深度学习网络  

A deep learning network for joint low-light enhancement and face spuer-resolution

作　　者：丛维仪 郑卓然 贾修一[1] CONG Weiyi;ZHENG Zhuoran;JIA Xiuyi(School of Computer Science and Engineering,Nanjing University of Science and Technology,Nanjing 210094,China)

机构地区：[1]南京理工大学计算机科学与工程学院,江苏南京210094

出　　处：《智能系统学报》2025年第1期109-117,共9页CAAI Transactions on Intelligent Systems

基　　金：国家自然科学基金项目(62176123,62476130).

摘　　要：在低光环境下,人脸图像增强是许多任务的重要恢复方法。然而,现有的低光环境下人脸超分辨率方法通常依赖于低光增强和超分算法的序列建模。遗憾的是,由于优化目标之间的差异,使用这种方法来增强人脸图像很容易导致伪影或噪声。为了应对这一挑战,本文提出了一个端到端的低光人脸图像超分辨率网络(low-light face super resolution network,LFSRNet)。该网络由浅层特征提取、深层特征提取和特征过滤上采样3个模块组成。首先浅层特征模块将输入的低光、低分辨率人脸图像映射到特征空间。随后,深度特征提取模块对其进行亮度校正并细化结构。最后,特征过滤上采样模块处理提取到的特征并重建人脸图像。此外,为了更好地重建丢失的面部细节本文还设计了一个损失函数faceMaskLoss。大量实验证明了所提模型的有效性。In low-light environments,face image enhancement is used as a vital recovery method for many tasks.However,existing methods for face super-resolution in low-light environments usually relied on sequence modeling that combines low-light enhancement and super-resolution algorithms.Unfortunately,using this method to enhance a face image easily led to artifacts or noise because of the differences between the optimization objectives.To tackle this challenge,we proposed LFSRNet,an end-to-end low-light face image super-resolution network.Our network consisted of three modules:shallow feature extraction,deep feature extraction,and feature filtering upsampling.The shallow feature module initially mapped the input low-light,low-resolution face image into feature space.Subsequently,the deep feature extraction module performed luminance correction and refined the structure.Finally,the feature filtering upsampling module processed the extracted features and reconstructed the face image.Additionally,in order to better reconstruct the lost facial details,we also designed a loss function faceMaskLoss.Extensive experiments demonstrate the effectiveness of our proposed model.

关 键 词：人脸超分辨率 低光图像增强 监督学习 随机掩码 损失函数 深度学习 局部特征提取 全局特征提取 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]