cvpr2013_IRL1_denoising.rar_CVPR 图像去噪_PSNR MATLAB代码_cvpr2013_w

Matlab code for iteratively reweighted l1 optimization for non-smooth non-convex optimization Copyright (c) 2013 Peter Ochs ------------------------------ Terms of use ------------------------------ This program is provided for research purposes only. Any commercial use is prohibited. If you are interested in a commercial use, please contact the copyright holder. If you used this program or part of this program in your research work, you should cite the following publication: Peter Ochs, Alexey Dosovitskiy, Thomas Brox, Thomas Pock An iterated L1 Algorithm for Non-smooth Non-convex Optimization in Computer Vision IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Jun 2013. This program is distributed WITHOUT ANY WARRANTY. -------------------------------------------------------------- Example Code for denoising with non-convex regularization term -------------------------------------------------------------- test_denoising.m: This file executes the denoising models with quadratic penalty on the data term and different convex and non-convex regularization terms: - primal_dual_TVL2.m: total variation penalty - primal_dual_TGVL2.m: total generalized variation penalty (of order 2) - ncLogTVL2_IRL1.m: penalty prototype log( 1 + beta*|grad u| ) - ncLogTGVL2_IRL1.m: penalty prototype log( 1 + beta*|grad u - v| ) + log( 1 + beta*|grad v| ) - ncpnormTVL2_IRL1.m: penalty prototype ( |grad u|+eps )^p with 0<p<1 - ncpnormTGVL2_IRL1.m: penalty prototype ( |grad u - v|+eps )^p + ( |grad v|+eps )^p with 0<p<1 Comments: --------- - the total generalized variation type implementation is the asymmetric total generalized variation - the parameters are not optimized with respect to the psnr, they are adjusted to visualize the effect of using non-convex penalty terms; feel free to play around with the parameters in order to better understand the results - there is no special reason for downsampling the input image (in test_denoising.m); it is just to get a quick result - in the inner iterations we perform at least 10 iterations - we perform a fixed number of 5000 iterations; a good "visual convergence" is achieved much earlier ------------------------------ Bugs ------------------------------ Please report any bugs to Peter Ochs: [email protected]