Single Image Super-Resolution using Spatial LSTM

Abstract

In this thesis, a single image super-resolution method (SR) using spatial long short term memory (SR-SLSTM) is proposed. For a highly ill-posed problem like SR, the receptive field size determining the amount of contextual information is the most important factor to solve the problem. While convolutional layer is limited to enlarge the receptive field size to gather more surrounding information and yields local features, spatial LSTM efficiently increases the receptive field to an image-level within a single layer and yields global image-level features. SR-SLSTM utilizes both local and global features using spatial LSTM residual unit at the same time to predict highly accurate high-resolution (HR) image from low-resolution image (LR). SR-SLSTM shows state -of-the-art results for SR benchmarks and especially yields better results from images with structural objects which benefit from contextual information. That result convince our proposed method using spatial LSTM efficiently extracts contextual information to handle a ill-posed problem like SR.