Large receptive field networks for accurate high-scale image super-resolution

This thesis presents a novel convolutional neural network architecture for high-scale image super-resolution. In particular, we introduce two separate modifications that can be made to the convolutional layers in the network: one-dimensional kernels and dilated kernels. We show how both of these methods can be used to expand the receptive field and performance of super-resolution networks, without increasing the number of trainable parameters or network depth. We show that these modifications can easily be integrated into any convolutional neural network to improve performance. Our methods are especially effective for the challenging high scale super-resolution due to the expanded network receptive field. We conduct extensive empirical evaluations to demonstrate the effectiveness of our methods, showing strong improvements over the state-of-the-art.