【作者】 王伟；

【导师】 金璟璇；

【作者基本信息】 延边大学 ， 计算机技术（专业学位）， 2021， 硕士

【摘要】 改善现实场景中获取到的人脸图像的清晰度和分辨率,可以显著地提升人脸识别准确率。人脸图像超分辨率重建的目标是利用输入的低分辨率(Low Resolution,LR)的人脸图像生成相应的高分辨率(High Resolution,HR)的人脸图像。基于深度卷积神经网络的方法已经成为人脸图像超分辨率重建技术领域的主流方法。经研究表明,增加卷积神经网络的层数能够提高网络模型的非线性拟合能力,但是网络层数的不断增多,网络模型也更难于训练,不利于人脸特征信息的提取。此外,低分辨率人脸图像中包含的大量低频信息在通道中被平等地处理,这种处理方式阻碍了卷积神经网络的表征能力。超分辨率生成对抗网络(Super resolution Generative Adversarial Networks,SRGAN)能够生成更真实的图像纹理,但是超分重建的人脸图像细节处常常会出现幻影或模糊现象。为解决上述问题,本文提出一种基于注意力机制和生成对抗网络的人脸超分辨率重建方法,并将其应用于视频中人脸图像的重建。首先,本文改进了SRGAN网络,通过在生成网络中引入残差网络构建残差组结构来改进网络模型,增强网络的特征提取能力。在残差网络中,将密集连接结构组合起来,以充分利用先前网络所学到的高频信息,并将小感知野的特征传输到网络的下一层以完成特征提取的任务。其次,考虑到人脸图像不同通道之间的依赖性,本文引入了一种用于自适应重新估计通道特性的注意力机制,以区别对待不同的通道,从而提高网络的表征能力,使网络模型更关注于人脸的五官特征信息。最后,将改进后的SRGAN网络模型和通道注意力机制相结合,构建基于通道注意力机制(Channel Attention,CA)和生成对抗网络(Generative Adversarial Networks,GAN)的人脸超分重建网络(CA-SRGAN),并将其应用于视频中的人脸图像的超分辨率重建。并结合人脸检测与跟踪、人脸图像优选技术,设计并实现了人脸图像超分辨率重建模块。为验证本文所构建的基于CA-SRGAN的人脸超分重建模型的性能,本文与其它算法模型做了对比实验。实验结果表明,通道注意力机制可以使网络更关注于人脸的五官特征信息,有利于人脸图像的超分重建;改进的残差组结构可以提高重建的人脸图像的质量和模型的重建效率。本文提出的模型能够更好的重建人脸图像,峰值信噪比达到31.68d B,结构相似性达到0.8819。更多还原

【Abstract】 Improving the clarity and resolution of face images acquired in real scenes can significantly improve the accuracy of face recognition.The goal of face image super-resolution reconstruction is to use the input low resolution(LR)face image to generate the corresponding high resolution(HR)face image.The method based on deep convolutional neural network has become the mainstream method in the field of face image super-resolution reconstruction technology.Studies have shown that increasing the number of layers of a convolutional neural network can improve the nonlinear fitting ability of the network model,but the increasing number of network layers makes the network model more difficult to train,which is not conducive to the extraction of facial feature information.In addition,a large amount of low-frequency information contained in low-resolution face images is processed equally in the channel,which hinders the representation ability of convolutional neural networks.Super resolution Generative Adversarial Networks(SRGAN)can generate more realistic image textures,but the face image details of super-resolution reconstruction often show ghosting or blurring.In order to solve the above problems,this thesis proposes a face super-resolution reconstruction method based on attention mechanism and generative confrontation network,and applies it to the reconstruction of face images in videos.Firstly,this thesis improves the SRGAN network,and improves the network model by introducing the residual network into the generation network to construct the residual group structure,and enhances the feature extraction ability of the network.In the residual network,the densely connected structure is combined to make full use of the high-frequency information learned by the previous network,and the features of the small perceptual field are transmitted to the next layer of the network to complete the task of feature extraction.Secondly,taking into account the dependence between different channels of face images,this thesis introduces an attention mechanism for adaptive re-estimation of channel characteristics to treat different channels differently,thereby improving the characterization ability of the network and making the network model pay more attention to the facial features of the face.Finally,the improved SRGAN network model and the channel attention mechanism are combined to construct a face super-resolution reconstruction network(CA)and Generative Adversarial Networks(GAN),and applied it to super-resolution reconstruction of face images in videos.And combined with face detection and tracking,face image optimization technology,design and realize the face image super-resolution reconstruction module.In order to verify the performance of CA-SRGAN constructed in this thesis,this thesis has done a comparative experiment with other algorithm models.The experimental results show that the channel attention mechanism can make the network pay more attention to the facial features information of the face,which is conducive to the super-resolution reconstruction of the face image;the improved residual group structure can improve the quality of the reconstructed face image and the reconstruction of the model effectiveness.The model proposed in this thesis can better reconstruct the face image,the peak signal-to-noise ratio reaches31.68 d B,and the structural similarity reaches 0.8819更多还原