Face super-resolution based on online dictionary learning

Abstract

Recent learning-based face super-resolution methods, such as Yang's Sparse Coding Super-Resolution (SCSR) are promising with sharp edges visually. But it also leads to obvious artifacts. In order to eliminate the artifacts, Online Dictionary Learning (ODL) algorithm is introduced in the dictionary learning phase to generate accurate overcomplete dictionary. On the other hand, the reconstruction regularization parameter is tuned independently to acquire the optimal sparse coefficient. With the solved dictionary and sparse coefficient, the recovered high-resolution image patches are predicted to reconstruct the target output image. In the experiments, the PSNRs and SSIMs of the proposed method are much higher than some state-of-the-art super-resolution methods. The PSNR is 0.85dB higher than SCSR in average, while the SSIM is 0.0133 higher. The artifacts and noise along the edges and corners are both eliminated effectively. More details of the faces are added to make the recovered high-resolution image clear. Additionally, the PSNRs decreases more smoothly while the noise level of the input lowre-solution image increases. It means that the proposed algorithm has promising performance in super-resolution reconstruction with good denoising effect.