Research On Super-resolution Of Single Face Image Based On Deep Learning

Single face image super-resolution reconstruction aims to improve the resolution of face images and enhance the visual effects,which has been widely used in public safety,video surveillance and other fields.At the same time,compared to the multi-frame super-resolution reconstruction technique,the single image super-resolution problem provides less information of the image and is more difficult to solve.Therefore,the research of super-resolution technology for low resolution face images is very necessary.Our paper focuses on the sample-based learning approaches.From the perspective of deep learning,we adopt two deep convolutional neural networks with different architectures.After reviewing the classic single image super-resolution algorithms based on sample learning,the main works of this paper are as follows:We proposed a single face image super-resolution method,which is based on l1 norm loss function and is a deep residual learning network.The method applies an end-to-end deep residual network without the batch normalization(BN)layers and chooses the l1 norm based loss function MAE as the loss function to train the network.Experiments show that the proposed method performs better than the original deep residual network in restoring the face structure and fidelity.We also proposed a single face image super-resolution method based on the reconstructive dense convolutional network(DenseNet),which is also an end-to-end network.In this network's architecture,each convolution layer has a direct connection between itself and the above convolution layers.The features are concatenate and the transportation ability of features is enhanced.The feature information is utilized more efficiently.This architecture also increases the width of the network.Experiments show that the proposed method performs better than other methods both on the international public data sets and the low quality facial images in real scenes.