Image super-resolution using local learnable kernel regression

In this paper, we address the problem of learning-based image super-resolution and propose a novel approach called Local Learnable Kernel Regression (LLKR). The proposed model employs a local metric learning method to improve the kernel regression for reconstructing high resolution images. We formulate the learning problem as seeking multiple optimal Mahalanobis metrics to minimize the total kernel regression errors on the training images. Through learning local metrics in the space of low resolution image patches, our method is capable to build a precise data-adaptive kernel regression model in the space of high resolution patches. Since the local metrics split the whole data set into several subspaces and the training process can be executed off-line, our method is very efficient at runtime. We demonstrate that the new developed method is comparable or even outperforms other super-resolution algorithms on benchmark test images. The experimental results also show that our algorithm can still achieve a good performance even with a large magnification factor.