An adaptive boosting procedure for low-rank based image denoising

Abstract

Low-rank methods have shown great potential in many low-level vision tasks, especially in image denoising. Nevertheless, state-of-the-art algorithms still have limitations in some aspects, such as a heavy reliance on empirical values when setting the denoising parameters and the lack of consideration of the noise effect on the selection of similar patches. To address these problems, an adaptive boosting technique for improving the low-rank based image denoising results is proposed. In particular, different from adding the residual back to the image, we consider strengthening the signal by leveraging the availability of the previous denoised image and combine this boosting technique with dynamic parameters applied to the weighted nuclear norm minimization (WNNM) method, whose feasibility is derived and proved theoretically. In each iteration of the boosting scheme, the dynamic boosting parameters are motivated by an optimal solution analysis. In addition, in the process of image denoising, an adaptive patch search scheme is put forward to gain an effective index of similar patches. Furthermore, to adaptively determine the optimal number of iterations, a stopping criterion with the correlation coefficient is applied. The experimental results demonstrate that the proposed approach can preserve more detail information while removing noise and can outperform various state-of-the-art denoising algorithms in terms of both PSNR and SSIM.