Image super-resolution via a novel cascaded convolutional neural network framework

Due to excellent self-learning ability, deep convolutional neural networks (CNNs) are successfully employed in the field of single image super-resolution (SR) compared with interpolation methods and sparse coding methods. To implement the multi-scale image SR task with a single trained model and to further improve the performance of image SR, we propose a novel cascaded CNN framework with three stages, which are feature extraction, detail prediction and reconstruction. At the stage of feature extraction, we propose a scheme of multi-scale feature mapping to extract the inherent features via the low-resolution image. At the stage of detail prediction, the lost details of the original low-resolution image are predicted via the network cascade. Finally, a high-resolution image is achieved by the scheme of residual learning, which is implemented by superimposing the lost details and the original low-resolution image. To avoid gradient explosions, we use gradient clipping to train the proposed cascaded CNN framework. Comparison results indicate that our proposed cascaded CNN framework for image SR is superior to many state-of-the-art methods.