Research On Learning-based Face Hallucination

In recent years,as a branch of Super-Resolution(SR),face hallucination has been the research focus for researchers in facial image area as well as the neighborhood area.With the help of information technology,face hallucination extends the resolution of target face by searching the relevant information in Low-Resolution(LR)face images.This technique provides an effective method for obtaining High-Resolution(HR)face images from low quality image systems,which is flexible as well as cheap.In the process of face image reconstruction,details of facial components can not be intricately reconstructed and the target HR face image may be generated with heavy distortion.To solve the problems above,two methods are designed in this paper.One is the sparse representation and facial components face hallucination,and the other is the residual convolution neural network.In the Sparse representation and facial components joint face hallucination(SrfcFH),firstly,the eight direction gradient graph with pixel neighborhood correlation and low dimension is used as the training feature of the sparse dictionary,and the texture gradient is extracted from the HR image obtained from the dictionary.Then,gradient map of facial components is composed of a high resolution gradient of the most similar eyebrows,the eyes,the nose and the of the mouth is combined with the contour gradient map of the HR edge gradient estimation together to form the target HR gradient map.Finally,the target HR image is calculated by iterative reflection algorithm.In the Residual convolution neural network(RCNN),a network model is built with the idea of training face residual image.The network model is determined by training the mapping relationship between LR image and HR tag,and the low resolution face is reconstruct according to the network model.In order to verify the effectiveness of the proposed method in face reconstruction,the and performs a simulation experiment on the performance of the algorithm.The two methods are evaluated and analyzed by the comprehensive evaluation index of the performance of the algorithm,which is composed of the subjective visual quality,peak signal-to-noise ratio(PSNR)and structural similarity index measurement(SSIM).For the face hallucination via CNN,two types network models of 5-layer network structure and 10-layer network structure is established.And 6 network models are trained according to the Adam optimization algorithm,the AdamDelta optimization algorithm,the fixed learning rate,the inv learning strategy and the poly learning strategy.Finally,the experiment of the performance comparison is carried out.The experiment shows that,in terms of the PSNR index,the SrfcFH method in this paper achieves the average level of 31.32 dB,which is 0.25 dB higher than the pre-improved ScSR.And the SrfcFH has a great improvement on the recovery effect of the facial components on the subjective visual quality.In terms of the PSNR index,the RCNN reaches an average level of 32.06 dB,which is 0.7dB higher than the SrfcFH.While in terms of the SSIM index,the average level is 0.865,which is 0.01 higher than that of SRfcFH.But in terms of the subjective visual quality,RCNN method is superior to the traditional method,while inferior to the SrfcFH.In this paper,the high frequency information of similar HR facial components and the depth mapping relationship between LR face and HR face are separately used in the sparse representation and facial components for face hallucination as well as the face hallucination via residual convolution neural network.And the two methods improve the matters that the traditional face hallucination is deficient in the information of human face component recovery and the mapping relations between high and low resolution faces are not clear.So that the visual quality and fidelity of the reconstructed HR face are improved.