Research On Face Hallucination

Video surveillance has been widely used in more occasions for the past few years.People in the surveillance video are often under unconstrained environment.The obtained face images caused by the pixel problem of the device and the long distance between the monitor and target often have small size and low quality problems.Low-resolution face images contain little useful information and much noise,which seriously affects the performance of low-resolution face recognition system in the real world.In order to restore the high-frequency detail information of the low-resolution face images and reconstruct more high-quality face images,this thesis mainly focuses on the learning-based face image super-resolution,that is,face hallucination.The major contributions of this paper are summarized below:(1)In contrast to existing face hallucination algorithm,a novel two-stage local details restoration framework for face hallucination is proposed in this thesis.Our framework pays more attention to the further local details enhancement.The local position-patch face hallucination method with locality constraint is introduced as the basic method of the two stages.The KNN algorithm and the contextual information are incorporated into the two stages respectively which bring more precise local details for the face image.Experiments demonstrate the proposed algorithm can recover the precise details and obtain a more high-quality face image.(2)In order to solve the problem of neglecting the non-local correlation when dealing with face hallucination problems,a non-local convolutional neural network is proposed in this paper.Compared with the existing face hallucination algorithms,non-local unit is introduced into the neural network.This unit makes the neural network pay more attention to image features which are more effective for reconstructing face images.Through the ability of CNN to acquire local information and the ability of non-local units to acquire global information,the representation of the neural network is effectively improved.Experiments show that the proposed algorithm is superior to other algorithms in subjective and objective evaluation.(3)Generally,mean square error loss function is often used in convolutional neural networks for face hallucination.This network design results in blurred ima-es.Therefore,a face hallucination algorithm based on conditional generative adversarial network is proposed in this paper.In the design of generator.Encoder-Decoder network structure with skip connection is proposed,which makes the network obtain more rich image hierarchy features.In the discriminator,the conditional constraint information is introduced to make the training of the generative adversarial network more stable.In addition,a fusion loss function is proposed,which includes pixel space loss,perceptional loss,adversarial loss and total variation loss.The experimental results show that the proposed algorithm can synthesize more abundant high-frequency detail information and obtain better visual effect.