Abstract:

To solve the problem that information of large singular values may get lost when small singular values are filtered in the matrix completion (MC) problem where all singular values of trace norm are thresholded by the same constant in low rank based MC model, a weighted low rank based MC model is proposed. The singular values are assigned with different weights, which can avoid singular values thresholded by the same constant. The approximation gradient algorithm is applied for solving weighted MC based low rank model. Finally, experimental results in image inpainting show that the proposed weighted MC outperforms traditional unweighted MC models in peak signal to noise ratio value and has obvious advantages.