Deep learning based image super-resolution

Abstract

Super-resolution is an algorithmic approach to increase the spatial resolution ofan image. In areas like medical imaging, satellite image processing, biometrics,robotics, text to speech conversion, optical character recognition etc., there is aneed for high-resolution images as they carry more details and finer grayscaletransitions in addition to the pleasing view.In this thesis, we propose a fast and robust method for single image superresolutionusing deep learning framework. Given the low spatial resolution testimage and a database consisting of low and high spatial resolution (LR - HR)images, we obtain super-resolution (SR) for the test image upto a magnificationfactor of 8. The novelty of our approach lies in the elimination of interpolationwhile learning. Our approach tries to learn the direct mapping between the LRand HR images. We use the idea proposed in [11] to represent the mapping by usinga deep convolutional neural network (CNN). CNN filters are learned by standardback-propagation and stochastic gradient descent method. We propose thata single trained network for a factor of 2 is sufficient to perform super-resolutionwith higher magnification factors. We have used grayscale images in all our experiments.Results have been compared with bicubic interpolation (digital zoom)and state-of-the art methods. Visual and quantitative comparisons confirm theefficacy of our proposed method.