融合五官轮廓注意力的人脸超分模型

Face super-resolution model incorporating various facial components attention

现有超分辨率人脸重建方法无法有效利用不同五官的结构信息,导致重建图像五官区分性差、缺乏各自细节特征。为解决这一问题,提出了一种融合五官轮廓注意力的人脸超分模型。搭建两阶段式神经网络:第一阶段利用残差网络提取深层复杂特征并对输入图像进行初步重建;第二阶段通过注意力融合网络将注意力集中在不同的面部成分(包括五官和面部轮廓),分别重建各成分的细节特征再进行融合。通过在CASIA WebFace上的定量与定性实验,证明了该两阶段式结构能够逐步提取特征、增强图片清晰度,恢复更多的图像细节。

The existing face super-resolution(FSR)methods cannot effectively use the structure information of features from different facial parts,resulting in the lack of various details of the facial features in the reconstructed image.In order to solve the problem,a face SR model incorporating various facial components attention is proposed.The model builds a two-stage neural network.At the first phase the residual network is used to extract deep complex features and to perform preliminary reconstruction of the input image.For the second phase,the attention fusion network focuses attention on different facial components(including facial features and facial contours),reconstructs the detailed features of each component and then merges them.Experiments on the CASIA Web Face show that the two-stage Model can gradually extract features,generate face images with clear outlines and rich facial features.