Adaptive image super-resolution with neural networks

Abstract

A framework for image super resolution using Multilayer Perceptron (MLP) was developed. In this paper, we focus on verifying the influence of training images to the performance of Multilayer Perceptron. The training images were collected from various categories, i.e. flowers, buildings, animals, vehicles, human and cuisine. The neural network trained with single image category has better performance compared to other methods. Multiple categories of training images made the MLP more robust towards different test scenarios. Common image degradation, i.e. motion blur and noise, can be reduced when the MLP is provided with proper training samples. © 2014 Springer Science+Business Media Singapore.