Research On Schemes Of Human Face Super-resolution Based On Hybrid Algorithm

With the continuous development of information technology and the Internet,the Internet of Things has become a major trend in the data age.Image as the most direct medium to transmit information in people's daily life,plays an irreplaceable role in the information age when things are connected.Although the hardware platform has reached a high level,due to environmental and noise factors,the imaging effect in public security areas such as video surveillance is still unable to achieve accurate face recognition,and super-resolution reconstruction algorithm can make up for the hardware defects from the software aspect.At the same time,the super-resolution algorithm based on deep learning has potential advantages in solving image reconstruction problems.Therefore,this paper focuses on the research of super-resolution reconstruction of human face.The main work is as follows:1)Aiming at the problems of high complexity and long training time of existing learning-based super-resolution algorithms,an improved face super-resolution convolution neural network model(SRCNN)is proposed in this paper.By changing the size of convolution kernel and the number of convolution kernels and introducing pooling layer,the convolution layer of SRCNN model is deepened,the number of features is reduced,the training time is shortened and the effect of reconstruction is better.2)Aiming at the problem that the traditional learning-based image super-resolution algorithm can't recover the features that are not available in the training database,this paper introduces the prior information of the original image through the iterative back-projection algorithm,and proposes a novel face super-resolution reconstruction algorithm(SRCNN-IBP)based on convolution neural network and iterative back-projection.The composite algorithm consists of nine layers,including four Convolution layer,two pooling layers,two under-sampling layers and one difference layer.The first five layers of the model are mainly realized: patch extraction and expression,non-linear mapping and reconstruction;two lower sampling layers down-sampling the original image and reconstructed high-resolution image respectively;the difference layer makes a difference between the results of the two lower sampling layers as a prior guidance;the upper sampling layer reconstructs the difference results;the data update layer reconstructs the reconstructed results of the fifth layer and the upper sampling error.Layer is iteratively reconstructed until the result of difference layer is less than the set threshold and stop iteration to output the reconstructed result.The experimental results show that the proposed algorithm achieves good results in training speed and reconstruction effect.