融合注意力机制的高分辨人脸识别图像重建

High-Resolution Face Recognition Image Reconstruction Combined with Attention Mechanism

针对由于人脸姿势、光照不均、拍摄环境、拍摄设备等内外部因素造成图像分辨率低的问题,提出融合注意力机制的高分辨人脸识别图像重建模型。首先以低分辨率人脸图像对作为两个生成器输入,通过残差块和注意力模块堆叠网络提取人脸特征信息,进而生成高分辨率人脸图像。训练中使用一个鉴别器来监督两个生成器的训练过程。利用Adam算法对鉴别器、生成器以及对抗损失函数进行迭代优化来提升网络模型性能。模型在CASIA-WebFace数据集上进行训练,在CASIA-WebFace、CelebA、LFW数据集上进行测试。实验表明,注意力模块可有效弥补浅层卷积神经网络提取特征时因偏向局部关系建模而缺乏对全局信息学习的不足,能学习利于重建高分辨率图像的特征信息。本模型重建人脸有较好的视觉效果,并且具有保留人脸身份信息的特性。

Aiming at the problem of low image resolution caused by internal and external factors such as face posture,uneven lighting,shooting environment,and shooting equipment,this paper presents a high-resolution face recognition image model combined with attention mechanism.Firstly,low-resolution face image pairs are used as input to the two generators,and facial feature information is extracted through the residual block and attention module stacking network,and then high-resolution people are generated.Face image.A discriminator is used in training to supervise the training process of the two generators.The Adam algorithm is used to iteratively optimize the discriminator,generator and counter loss function to improve the performance of the network model.The model is trained on the CASIA-WebFace dataset and tested on the CASIA-WebFace,CelebA,and LFW datasets.Experiments show that the attention module can effectively compensate for the lack of global information learning due to the partial relationship modeling when the shallow convolutional neural network extracts features,and it can learn the feature information that is conducive to the reconstruction of high-resolution images.This model has a good visual effect in reconstructing the face,and has the characteristics of retaining the identity information of the face.