Face Super Resolution Based On Convolutional Neural Networks

Face super-resolution reconstruction,also known as face hallucination,refers to the technique of transforming a compressed low-resolution face image to a high-resolution clear face.It is capable of reconstructing the missing facial details in the process and removing ^tnhoet_{lcaorng ve}^nc _olmuati^oairsrey _{g o} ^aonudt_{nnifail c nateiuo}^bslpuer _rna^ciin.^a_{l nT}^ltt _eht^{hree siemaracghe}_{ewreofrokrse.},^s._T^{o n}TT _{thhii ss}s ^hfe _ppaa^{a ctre a}_{pp eerr}^{rdeit c} _hmasa^ioonns _ianclyh^{atlr uscu} _iestvued^tpieorn-,_d^r _{i etsh}^{easno dl}_ett ^u_{hf eo}^tci _{l lfoacwei}^{aonnn roet} _n^c _sgu^roe _pmear^ncsotrn _jroer^{usctrt iuocnt}_sroelsu ^talhge_utiltosn:^o _base^friatche _d^{m sw}¨_ùonⁱdthodtho For face reconstruction with multiple magnifications,this paper proposes a face super-resolution reconstruction method based on very deep convolutional neural networks.First,a neural network with 20 convolutional layers is used.Each layer of the network has a convolution kernel composed of several small filters.These kernels are cascaded to learn a relationship between low-resolution facial images and high resolution facial images.Secondly,by using residual learning,the network only needs to learn the detail information in the reconstruction process,avoiding the gradient disappearance and gradient explosion in training.When training the model,the face images with different magnifications are added to the training set,so that the modified model can effectively solve the face reconstruction problem with different magnifications.The experimental results show that the network has a significant improvement compared with the traditional method and the simple convolutionnetworks reconstruction method,and also have better reconstruction effect under larger magnification.However,the above method is mainly studied on the black and white face images in the laboratory environment.In order to study the reconstruction of the colorful face in the more complicated natural scene,a pyramid structure super resolution reconstruction network with attention mechanism is proposed.Firstly,a pyramid structure with sub-pixel up-sampling network is designed.Each up-sampling network has three residual modules,and each residual module is composed of a 20-layer network with attention mechanism.Then,the extracted features are reconstructed layer by layer into high-resolution images,and connected with the image that is directly up-sampled by 4 times.The method makes the model learn the detail information in the reconstruction process.The network is capable of achieving two times or four times super-resolution by selecting up-sampling of one or two times.Finally,the experimental results validated the effectiveness of the proposed method.