3.Global Image Denoising
H.Talebi and P.Milanfar,IEEE transactions on Image Processing, Vol 23, no.2,pp.755-768, Feb.2014
Image denoising algorithms are mostly patch based, like BM3D although seemed to be good, as the image size grows bigger in size the performance is not up to the mark. So, in this paper each pixel value is estimated from all other pixels in the image, which means denoising is done globally and not locally to achieve good results.
The noisy image is first prefiltered by using the available denoising methods such as bilateral filter or NLM(Non-Local Means). Eigen vectors are decomposed and approximated for the pre-filtered noisy image. If the eigen values for the full image is decomposed, then it is very complicated and needs very large space for storage and the process is slow. This problem is overcome by breaking or splitting the images or the matrix , the eigen vectores are approximated only for few points by using Nystrom and sinkhorn approximation. The eigen vectors are then shrunk to get the eigen values, followed by the filter optimization. Filters like iterative filter and truncated filter are used to get the denoised image.
This global denoising image method can be effectively used with any existing denoising filters to get better results.
Software : https://users.soe.ucsc.edu/~htalebi/GLIDE.php#
The noisy image is first prefiltered by using the available denoising methods such as bilateral filter or NLM(Non-Local Means). Eigen vectors are decomposed and approximated for the pre-filtered noisy image. If the eigen values for the full image is decomposed, then it is very complicated and needs very large space for storage and the process is slow. This problem is overcome by breaking or splitting the images or the matrix , the eigen vectores are approximated only for few points by using Nystrom and sinkhorn approximation. The eigen vectors are then shrunk to get the eigen values, followed by the filter optimization. Filters like iterative filter and truncated filter are used to get the denoised image.
This global denoising image method can be effectively used with any existing denoising filters to get better results.
Software : https://users.soe.ucsc.edu/~htalebi/GLIDE.php#
Results for MATLAB Code
Elapsed time= 1001.06 sec
Eigen-decomposition Approximation..
Sampling percentage=1.03
Elapsed time=379.12 sec
MSE minimization..
Elapsed time=267.85 sec
GLIDE PSNR = 25.58db
Eigen-decomposition Approximation..
Sampling percentage=1.03
Elapsed time=379.12 sec
MSE minimization..
Elapsed time=267.85 sec
GLIDE PSNR = 25.58db