Image Super Resolution Using Deep Learning

Abstract

Image upscaling has been applied in many applications in the image processing field. This paper shows a model which is able to perform image upscaling by 4 times using a series of convolutional filters and trained using the generative adversarial network (GAN) training scheme. The GAN training process involves a generator network, which will perform the image upscaling. The results of the generator network will be evaluated by a discriminator network for the realistic score which will be feedback to the generator network for training. The chosen GAN type is the GAN with a relativistic discriminator which calculates how realistic is the generated image compared to the real image. The network also utilizes different structures of dilated convolution filter, inception module and residue connection between the filters to enhance the feature extraction capability. The high-definition image dataset DIV2K is used for the training.