Nonlocal B-spline representation of tensor decomposition for hyperspectral image inpainting

•Tight frame transform using B-spline is adopted in our method in order to obtain relatively tighter low-rank properties.•To employ low-rank and spatial non-local similarity properties, we use subspace clustering technique for completion problem.•We formulate the model by fully-connected tensor network decomposition to fully use spatial-spectral information.•The proximal alternating minimization is employed to optimize the variables and offers a principled and closed-form solution.