Algorithm 1 Multi-view Image Denoising

Input: Multi-view images Is,t, maximum candidate disparity value dmax, pre-trained MVCNN, target image number k.

Output: Denoised target image Iest.

Initialize: Denoised target image Iest = zeros(size(Is,t)), weight matrix W = zeros(size(Is,t)).

1: for d = 1:dmax

2:  Construct 3D focus image stacks Fd using Equation (3);

3:  Obtain denoised image stacks ${\widehat{F}}^d$ by applying MVCNN to Fd;

4: end

5: Estimate the disparity map for the target image using Equations (4)-(6);

6: for each pixel (x, y)

7:  Find its disparity d(x, y);

8:  Obtain a patch P centered at (x, y) in the kth image of image stack ${\widehat{F}}^d\left(x,y\right)$, and compute

its weight w.r.t. the reference patch Pref as $w=e^{-{\left(P - P_{ref}\right)}^2}$;

9:  Update Iest = Iest + w·P;

10:  Update W = W + w;

11: end

12: Compute the denoised target image Iest = Iest/W;

13: Detect and handle occlusion using Algorithm 2.