Face Hallucination Based On Cluster Model Regression

In fact,because of the constraints of camera equipment and impact of various realities,the resolution of captured face images are very low and can't be used directly.Therefore,it is necessary to reconstruct a low-resolution(LR)face image into a high-resolution(HR)face image by face hallucination technique.Face hallucination uses machine learning method to learn the mapping relationship between LR and HR face images through training samples,and reconstructs the input LR image into HR image by the mapping relationship.This paper aims to the application of face hallucination technology in real scene,and focuses on the reconstruction of noisy face images and free perspective face images.The main research results and contributions are as follows:1.Face hallucination based on Needle position cluster and multiple regression.Needle is a structure that connects the corresponding positions of each image in the image of Pyramid together.In order to enhance the reconstruction effect of the noise image,this paper uses the Needle structure to replace the traditional image patch because the cluster of Needle structure is more accurate than patch.After clustering,the patch is extracted and projected into a common manifold by Partial Least Squares.The mapping relationship between LR and HR images is learned in the common manifold.The proposed method and compared algorithms use FEI and CAS_PEAL_R1 face database to experiment with different degrees of noise.Experiment results show that the proposed method can increase about 0.2~0.5dB in PSNR compared with the algorithm using patch matching.2.Face hallucination for free perspective face images based on subspace nonlinear regression.In order to resolve the multi-pose problem of face imge in real scene,it is necessary to reconstruct the free perspective face image.Training a sparse dictionary by using LR patches is the first step.Each atom of a dictionary is regarded as a subspace base vector.All the training patches are classified according to the representation coefficient in the subspace base vector.Afterward the proposed method uses Extreme Learning Regression model to learn the nonlinear regression relationship between LR and HR images in each cluster.In the prediction phase,once the cluster of the LR patch is gotten,the corresponding learned regression function can be used to estimate HR patch.This paper makes experiments using the FEI face database and the free perspective face images collected from the Internet.The experimental results show that the proposed method can increase about 0.6~0.8dB in PSNR compared with A+.