We propose methodologies to train highly accurate and efficient deepconvolutional neural networks (CNNs) for image super resolution (SR). A cascadetraining approach to deep learning is proposed to improve the accuracy of theneural networks while gradually increasing the number of network layers. Next,we explore how to improve the SR efficiency by making the network slimmer. Twomethodologies, the one-shot trimming and the cascade trimming, are proposed.With the cascade trimming, the network's size is gradually reduced layer bylayer, without significant loss on its discriminative ability. Experiments onbenchmark image datasets show that our proposed SR network achieves thestate-of-the-art super resolution accuracy, while being more than 4 timesfaster compared to existing deep super resolution networks.
展开▼