Improvement of Image Super Resolution by Deep Neural Networks

The modern methods and architectures for image super resolution which are based on deep neural networks (DNNs) are considered. Several ways of their improvements were proposed and demonstrated. It was shown that the perception models built on MobileNet and EfficientNet families of DNNs turned out to be faster in training and have a better perception loss rate than previously used VGG family. In the more general context the usage of the smaller DNNs with the higher performance and lower size allow researchers to use and deploy them on devices with the limited computational resources for Edge Computing layer.