Face Super-resolution Via Optimized Deep Learning Method

Face super-resolution research is a hot research field in computer graphics and vision.It has a wide range of applications in security surveillance,unmanned aerial vehicle imaging,compression reconstruction and so on.The resolution of the face image represents the recognitiability of the image,while high-resolution images tend to have richer information.However,in practical applications,it is usually difficult to clearly identify the specific features of the object due to the factors such as equipment defects and media compression,which leads to the low quantity video and low-resolution face images.At the same time,improving the quality of hardware equipment alone will lead to higher costs.Therefore,it is of great theoretical value and broad application prospects to improve the resolution of face images through software algorithms and to restore and enhance the visual effects of images.In recent years,deep learning technology has been rapid development,with the coming of the big data era,convolutional neural network also brings new ideas for the super-resolution of face images.Compared with the interpolation and popular learning-based face super-resolution algorithm,deep learning face-super-resolution algorithm can automatically learn the feature extraction process and enhance the generalization ability of the super-resolution model,with stronger potential and applicability.In this paper,a systematic study on the super-resolution of face images based on convolutional neural networks is carried out.Combining with the characteristics of strong structural and feature differentiation of face images,the structure and training of convolutional neural networks are optimized.From the two aspects of feature point detection and fusion,progressive convolution neural network optimization to explore new methods,the main research contents and contributions of this paper are follows:(1)We present a feature representation method based on key points fusion,which can enrich the semantic features of the face image and enhance the nonlinear fitting ability of neural networks.The proposed method is also robust to face images in unrestricted scenes.(2)We present a feature extraction algorithm based on multichannel convolution,proposed to improve the end-to-end training process of traditional neural networks.The super-resolution of traditional large multiples is divided into multiple small-scale super-resolution reconstruction processes.In order to reconstruct the final result,the parameters of the neural network model are reduced,the bustling capacity of the model is enhanced,the overfitting is avoided better,the calculation amount is reduced,and the super-resolution performance is improved.In summary,this article will study the semantic features of face images and fusion,as well as progressive convolutional neural network optimization,and provide new theories and methods for super-resolution based on single face images.A preliminary face super resolution reconstruction system was formed.Secondly,this article considers the structure and method of progressive networks and will continue to study them in future work.The method in this article can also be used in other image enhancement tasks such as grayscale image coloring,image denoising,etc.