Gradient boosting for single image super-resolution

The learning-based single image super-resolution (SISR) algorithm aims at recovering a high-resolution (HR) image from low-resolution (LR) input. The quality of the HR output mainly depends on the strength of the learning algorithms. Observing that gradient boosting is powerful in dealing with learning problems, we propose a new SISR method based on the gradient boosting framework. First, the boosting framework is extended to the general form of multi-output regression. Then, an error correction approximation is used to sequentially train the boosting trees. The training data for each tree are the pairs of the features of the LR image patches and the negative gradients of the loss function. Meanwhile, shrinkage, a slow learning strategy, is exploited to reduce the risk of overfitting. Finally, all boosting trees are linearly combined to form an accurate predictor. The experimental results verify that our method can generate visually pleasant HR images and achieve accuracy on par with state-of-the-art methods in terms of quantitative evaluation.