Generalized Adaptive Weighted Recursive Least Squares Dictionary Learning for Retinal Vessel Inpainting

摘要

Inpainting is an ill-posed inverse problem that uses information from the available parts of the image to fill in the missing parts. Among the approaches used to solve this problem, recent works tend toward sparse representation algorithms. In particular dictionary learning for sparse representation of signals is preferred in this application. A recent work utilized recursive least squares dictionary learning (RLS-DL) to improve the inpainting of retinal vessels and has shown superiority over the existing approaches. In this paper, we propose to use generalized adaptive weighted recursive least squares (GAW-RLS) dictionary learning to inpaint the retinal vessels. The proposed GAW-RLS applies a correction weight to adaptively control the relative consistency of the training data with the existing estimate of the dictionary per each iteration. This enables GAW-RLS to outperform RLS-DL in vessel inpainting. Our simulation results indicate advantages of GAW-RLS over RLS-DL by reducing the recovery error for the missing pixels.