An Novel Face Image Super-Resolution Based on Sparse Representation and Wasserstein Distance

Face super-resolution has attracted much attention in recent years. Many algorithms have been proposed. However, these approaches only perform well under the condition that the input is noiseless or has small noise. These existing approaches think the measure of between LR and HR is Euclidean distance. Euclidean distance cannot measure well the condition that the image's pixel values tend to have random fluctuations of varying magnitudes or the image is represented as sparse high-dimensional data. In this paper, we propose a novel sparse representation and Wasserstein distance based face super-resolution approach that seem the measure of between LR and HR as Wasserstein distance. Specifically, we introduce the fused Lasso to the least squares representation of the input LR image in order to obtain a stable sparse representation, especially when the noise level of the input LR image is high. Experiments are carried out on the benchmark FERET face dataset. Visual and quantitative comparisons show that the proposed face super-resolution method achieves comparable performance to the state-of-the-art methods under noiseless condition, and yields superior super-resolution results when the input LR face image is contaminated by strong noise.