A Combined Regularizer Based Image Inpainting

Abstract

As a significant topic in computer vision, image inpainting aims to recover the natural pattern of an image in which pixels have been partially removed or occluded. To achieve this task, the traditional method uses a convex total variation (TV) regularizer. However, as the TV regularizer has to be solved by iteration, image inpainting results obtained from classical methods have the over smoothness tendency and the staircase effect when iteration number increases. To solve the problem, in this paper we introduce a new regularizer to achieve better image inpainting. The new regularizer is a weighted combination of the classical TV regularizer and a nonconvex regularizer. To solve our new model efficiently, a new algorithm is also proposed which is based on the primal-dual algorithm and the iteratively reweighting method. Numerical results demonstrate that our proposed image inpainting model is more effective than classical TV regularizer based models.