The Design And Application Of Perceptual Loss Function In Deep Learning

With the rapid development of the neural network architecture and loss function,convolution neural network models have been applied widely in the field of image processing.In high level visual tasks,on the one hand,deepening and widening neural network are limited by gradient back-propagation and computing resources,on the oth-er hand,The more parameters the neural network has,the more training data it needs.In the low level task such as image denoising and image super-resolution,high level features are not necessarily,suitable for dealing with low level ta,sks.Recently,more and more researchers realize that the design of loss function is also the most important link in image processing.A good loss function can contain more supervised infor-mation which is conductive to train the neural network,for example,in the field of population counting,population density map instead of population size contains more information.Not only but also,a good loss function can also make network learn more discriminative features,compared to softmax,L-softmax makes the distance between different categories of features larger.Perceptual loss function which is based on high-level features can guide the network to generate high-quality images.In this paper,we focus on the perceptual loss function and its application in image super-resolution and domain adaptation.Specifically,we make preliminary contributions from the follow-ing aspects:Firstly,in the field of image super-resolution,we proposed a denoising auto-encoder with symmetric skip connection(SDAE)and the output features of the encoder are used as the perceptual loss function.(1)SADE has shown the extremely good ef-fectiveness on learning features for CNN pre-training on classification and detection task.(2)The use of skip layer connections is conducive to train the network,and the output of the encoder contains the features of different layers.Compared to the stack-ing network structure,skip layer connection helps to reduce the correlation between adjacent features.(3)image denoising belongs to the image restoration task,compared to the classification task,it is better to recover the image from the mid feature,and less image information lost in features.Experiments demonstrate that the the images generated by the proposed method have both better visual quality and higher PSNR and SSIM than the start-of-art methods.Secondly,many of the results from our perceptual loss models have grid-like arti-facts at the pixel level compared to baseline methods,following the method proposed by SRGAN,we also train a discriminator to distinguish the generated images and real images.The image recovery network is aimed to generate the images with a mini-mum perceptual distance that can cheat the discriminator simultaneously.Experiments shown that the image recovery network generate more real high-resolution images.Thirdly,we extend the perceptual loss from measuring the difference between the two images to the measurement of two feature distributions.Different from the tradi-tional maximum mean difference(MMD),our function.f(·)is a pre-trained classifier,and eventually transforms the two complex feature distributions into simple multinomi-al distributions(KL-MMD).the function f(·)in MMD is a data independent function such as gaussian kernel.In our domain adaptation,we combine the Matching Gate with Attention Mechanism and put forward to Matching Attention to learn feature vectors.To our knowledge,our method achieves state-of-the-art digit recognition performance on three unsupervised adaptation results.