【專題演講】109/9/10(四)16:30-17:20 林家祥教授


Sentinel-2 satellite, launched by the European Space Agency, plays a critical role in various Earth observation missions. However, the spatial resolutions of Sentinel-2 images are different across its 12 spectral bands, meaning that there is no pixel in such imagery. To facilitate the analysis of such multi-resolution images, super-resolving (SR) of the low-resolution bands to a higher resolution is desired. Without relying on big data, we computationally achieve this SR task from a single dataset. As in many image restoration inverse problems, we exploit image self-similarity, a commonly observed property in natural images. However, the design of self-similarity regularization in non-diagonal inverse problems is challenging; often, a self-similarity based denoiser is plugged into the algorithmic iterations, without a guarantee of convergence in general. For the first time, we explicitly define the concept of self-similarity as a convex function, built explicitly on a self-similarity graph that can be directly learned from the Sentinel-2 images. Remarkably, unlike widely used sparsity or total-variation regularization schemes, the proposed convex function is scene-adapted. We then develop a fast algorithm, termed Sentinel-2 super-resolution via scene-adapted self-similarity (SSSS), which efficiently and exactly solves three involved different types of very large-scale matrix inversions. We experimentally show the superiority of SSSS over four commonly observed scenes, indicating the potential usage of our newly introduced convex self-similarity regularization in other ill-posed imaging inverse problems.