Jointly Optimized Regressors for Image Super-resolution

摘要: Learning regressors from low-resolution patches to high-resolution has shown promising results for image super-resolution. We observe that some are better at dealing with certain cases, and others different cases. In this paper, we jointly learn a collection of regressors, which collectively yield the smallest super-resolving error all training data. After training, each sample is associated label indicate its 'best' regressor, one yielding error. During testing, our method bases on concept 'adaptive selection' select most appropriate regressor input patch. assume similar can be super-resolved by same use fast, approximate kNN approach transfer labels test patches. The conceptually simple computationally efficient, yet very effective. Experiments four datasets show outperforms competing methods.