Efficient, Lightweight, Coordinate-Based Network for Image Super Resolution

Deep learning approaches have recently proven their effectiveness in the task of image Super Resolution (SR). In most cases, very deep structures have been adopted to increase the models’ performance, leading to neural networks with a high parameter count that require large computational resources. In this paper, we propose an efficient, lightweight model, which leverages the benefits of recursive architectures. The structure of our network is based on progressive reconstruction, which strengthens the information flow by taking advantage of dense and residual connections. Moreover, since SR is a problem that involves spatial representations and transformations, we exploit the pixel position information to reinforce the reconstruction task. To achieve that, we use the Coordinate Convolutional layer, which exploits coordinate information allowing the network to learn the translation dependency required by the SR task. We show that the proposed method performs favorably compared to lightweight state-of-the-art methods on public benchmark datasets.