The Research On Generation Adversarial Methods For Ultra-Low Resolution Face Hallucination

With the continuous construction of security projects such as "intelligent transportation","safe city" and "smart city",video surveillance plays an increasingly prominent role in public security.However,an undeniable fact is that in large outdoor scenes and unrestricted environment,the face images collected by surveillance cameras are often of low resolution and unable to extract useful information.Although face image super-resolution technology,a.k.a face hallucination,can estimate high resolution face image from low resolution face image,existing face hallucination methods can not achieve ideal reconstruction effect when achieving 8 times magnification of small face image.Therefore,how to reconstruct high resolution images from low resolution images with only a dozen or fewer pixels in length and width has become a technical problem to be solved urgently at home and abroad.In recent years,thanks to the development of big data and cloud computing,deep learning has achieved unprecedented accuracy in computer vision tasks.Therefore,based on the deep neural network and generative adversarial networks,this paper makes the following three contributions to the 8 times amplification of ultra-low resolution face:Firstly,aiming at the problem of insufficient reconstruction effect of current face hallucination methods,a new method Tiny Face Hallucination via Relative Adversarial Generative Network was proposed.In this method,residual blocks,dense blocks and deep separable convolution operators are introduced into the generated network to reduce the number of parameters while ensuring the depth of the network.In order to give full play to the ability limit of the relative generative adversarial networks in the problem of small face hallucination,VGG128 was refined by removing batch normalization layer and adding full connection layer successively.Experimental results show that the proposed method can output the phantom face with higher resolution from deep network.However,like most deep neural networks,the face hallucination model still lacks some interpretability.Second,aiming at the problem of poor interpretability of current deep neural networks,the linear multistep method for solving differential equations is introduced into the generative model of the former method,i.e.,Tiny Face Hallucination via Relative Adversarial Generative Network,so as to explain the principle of deep neural network from the perspective of mathematical analysis.In specific,a new model of Tiny Face Hallucination via Linear Multistep Method is further proposed.The experimental results show that the linear multistep method can be used to construct the face hallucination model with small amount of parameters and high precision.However,under the constraint of pixel loss term,the realism brought by the relative generative adversarial networks for the phantom face is still insufficient.Thirdly,in order to enhance the reality of face hallucination and reduce the complexity of network training,a new method Tiny Face Hallucination via Contextual Loss was proposed.The method uses the contextual loss to train a feedforward neural network as a face hallucination model.As a perceptual loss,the contextual loss makes use of all the features extracted in the middle of the convolution network as a comparative object.It is more targeted to high frequency detail features of the face,so as to improve the reality of the generated samples.The experimental results show that the proposed method can output more realistic hallucinated faces from deep network.