Unaligned Face Hallucination Based On Hierarchical Clustering Regression

Face hallucination refers to the process of reconstructing the corresponding high-resolution image from a low-resolution face image.Existing face super-resolution reconstruction algorithms usually assume that the input image is aligned and noiseless.When the input face image is not aligned,the performance of super-resolution reconstruction will be reduced.In this thesis,super-resolution reconstruction of unaligned single face images is studied.The main research results and contributions are as follows:(1).Unaligned face super-resolution reconstruction based on anchored neighborhood and hierarchical clustering regression is proposed.The size of the training sample is unified into a small size face image,which is used to train the face image dictionary.The dictionary atoms of the dictionary are clustering centers.The original face images are clustered and the face image clusters in each subspace are obtained.During the training,only one global dictionary needs to be trained.Face images in each subspace share these dictionary atoms.In each cluster,the neighborhood of each anchored is searched to generate the corresponding neighborhood subspace.Then,by learning the mapping relationship between low resolution and high resolution sample features,the regression model of each subspace can be obtained.The core of the method is that all face image subspace share a global dictionary,but for the same anchored,in different face image clusters,the neighborhood samples are different,which can more accurately learn the local mapping relationship.Compared with similar comparison algorithms,the peak signal-to-noise ratio(PSNR)of this method can be increased by at least 0.1165 dB and the structural similarity can be increased by 0.1782 on the unaligned face samples in Vggface 2 and CelebA databases.(2).Unaligned face super-resolution reconstruction based on hierarchical clustering neural network regression.In each face image cluster,the mapping relationship between low-resolution and high-resolution samples is trained by Laplacian pyramid network,and the regression model of each subspace can be obtained.In the reconstruction stage,the low-resolution face image is input and the corresponding face image cluster is selected.Then the high-resolution image is predicted based on the regression model of the image cluster.The network reconstructs high-resolution images step by step through multiple pyramid layers.It can extract features of low-resolution input images directly without the need for bicubic interpolation of low-resolution face images.The experimental results show that the peak signal-to-noise ratio(PSNR)of this method is 0.4634 dB higher than that of the previous one,and it has the advantages of reconstruction quality and time complexity over other leading algorithms.