Sea-surface image super-resolution based on sparse representation

Learning-based super-resolution (SR) is a popular SR technique that uses application-specific priors to recover missing high-frequency components in low resolution (LR) images. In this paper, we propose a novel approach for obtaining high-resolution (HR) image with solely a single low-resolution input sea-surface image. It is based on sparse representation via dictionary learning. As the image patch can be well represented through a sparse linear combination of elements from the training over-complete dictionary, this paper proposes a two-step statistical approach integrating the global model and a local patch model. During the training process, we divide the corresponding training images into patches and take the schismatic hierarchical clustering algorithm to get the idiosyncratic patches aimed at the background of sea-surface, using the jointly training method generating two over-complete dictionaries for the LR and HR images. In the reconstructed process, we infer the HR patch for each LR patch by the sparse prior in the local model, and recover the HR image via the reconstruction constraint in the global model. For our particular applications of sea-surface image SR, the proposed method has a more effective performance than other SR algorithms.