Merge pull request #73 from vais-ral/NLTV

Nonlocal TV method added (CPU version)

8 files changed, 404 insertions, 10 deletions

diff --git a/Wrappers/Matlab/demos/demoMatlab_denoise.m b/Wrappers/Matlab/demos/demoMatlab_denoise.m
index d11bc63..54b8bac 100644
--- a/Wrappers/Matlab/demos/demoMatlab_denoise.m
+++ b/Wrappers/Matlab/demos/demoMatlab_denoise.m
@@ -135,7 +135,21 @@ figure; imshow(u_diff4, [0 1]); title('Diffusion 4thO denoised image (CPU)');
 % tic; u_diff4_g = Diffusion_4thO_GPU(single(u0), lambda_regDiff, sigmaPar, iter_diff, tau_param); toc; 
 % figure; imshow(u_diff4_g, [0 1]); title('Diffusion 4thO denoised image (GPU)');
 %%
-
+fprintf('Weights pre-calculation for Non-local TV (takes time on CPU) \n');
+SearchingWindow = 7;
+PatchWindow = 2;
+NeighboursNumber = 15; % the number of neibours to include
+h = 0.23; % edge related parameter for NLM
+[H_i, H_j, Weights] = PatchSelect(single(u0), SearchingWindow, PatchWindow, NeighboursNumber, h);
+%%
+fprintf('Denoise using Non-local Total Variation (CPU) \n');
+iter_nltv = 2; % number of nltv iterations
+lambda_nltv = 0.085; % regularisation parameter for nltv
+tic; u_nltv = Nonlocal_TV(single(u0), H_i, H_j, 0, Weights, lambda_nltv, iter_nltv); toc; 
+rmse_nltv = (RMSE(u_nltv(:),Im(:)));
+fprintf('%s %f \n', 'RMSE error for Non-local Total Variation is:', rmse_nltv);
+figure; imshow(u_nltv, [0 1]); title('Non-local Total Variation denoised image (CPU)');
+%%
 %>>>>>>>>>>>>>> MULTI-CHANNEL priors <<<<<<<<<<<<<<< %
 
 fprintf('Denoise using the FGP-dTV model (CPU) \n');
diff --git a/Wrappers/Matlab/mex_compile/compileCPU_mex_Linux.m b/Wrappers/Matlab/mex_compile/compileCPU_mex_Linux.m
index 767d811..72a828e 100644
--- a/Wrappers/Matlab/mex_compile/compileCPU_mex_Linux.m
+++ b/Wrappers/Matlab/mex_compile/compileCPU_mex_Linux.m
@@ -52,6 +52,12 @@ fprintf('%s \n', 'Compiling ROF-LLT...');
 mex LLT_ROF.c LLT_ROF_core.c utils.c CFLAGS="\$CFLAGS -fopenmp -Wall -std=c99" LDFLAGS="\$LDFLAGS -fopenmp"
 movefile('LLT_ROF.mex*',Pathmove);
 
+fprintf('%s \n', 'Compiling NonLocal-TV...');
+mex PatchSelect.c PatchSelect_core.c utils.c CFLAGS="\$CFLAGS -fopenmp -Wall -std=c99" LDFLAGS="\$LDFLAGS -fopenmp"
+mex Nonlocal_TV.c Nonlocal_TV_core.c utils.c CFLAGS="\$CFLAGS -fopenmp -Wall -std=c99" LDFLAGS="\$LDFLAGS -fopenmp"
+movefile('Nonlocal_TV.mex*',Pathmove);
+movefile('PatchSelect.mex*',Pathmove);
+
 fprintf('%s \n', 'Compiling additional tools...');
 mex TV_energy.c utils.c CFLAGS="\$CFLAGS -fopenmp -Wall -std=c99" LDFLAGS="\$LDFLAGS -fopenmp"
 movefile('TV_energy.mex*',Pathmove);
@@ -66,6 +72,7 @@ mex NonlocalMarching_Inpaint.c NonlocalMarching_Inpaint_core.c utils.c CFLAGS="\
 movefile('NonlocalMarching_Inpaint.mex*',Pathmove);
 
 delete SB_TV_core* ROF_TV_core* FGP_TV_core* FGP_dTV_core* TNV_core* utils* Diffusion_core* Diffus4th_order_core* TGV_core* LLT_ROF_core* CCPiDefines.h
+delete PatchSelect_core* Nonlocal_TV_core*
 delete Diffusion_Inpaint_core* NonlocalMarching_Inpaint_core*
 fprintf('%s \n', '<<<<<<< Regularisers successfully compiled! >>>>>>>');
 
diff --git a/Wrappers/Matlab/mex_compile/compileCPU_mex_WINDOWS.m b/Wrappers/Matlab/mex_compile/compileCPU_mex_WINDOWS.m
index 1b59dc2..6f7541c 100644
--- a/Wrappers/Matlab/mex_compile/compileCPU_mex_WINDOWS.m
+++ b/Wrappers/Matlab/mex_compile/compileCPU_mex_WINDOWS.m
@@ -60,6 +60,12 @@ fprintf('%s \n', 'Compiling ROF-LLT...');
 mex LLT_ROF.c LLT_ROF_core.c utils.c COMPFLAGS="\$COMPFLAGS -fopenmp -Wall -std=c99"
 movefile('LLT_ROF.mex*',Pathmove);
 
+fprintf('%s \n', 'Compiling NonLocal-TV...');
+mex PatchSelect.c PatchSelect_core.c utils.c COMPFLAGS="\$COMPFLAGS -fopenmp -Wall -std=c99"
+mex Nonlocal_TV.c Nonlocal_TV_core.c utils.c COMPFLAGS="\$COMPFLAGS -fopenmp -Wall -std=c99"
+movefile('Nonlocal_TV.mex*',Pathmove);
+movefile('PatchSelect.mex*',Pathmove);
+
 fprintf('%s \n', 'Compiling additional tools...');
 mex TV_energy.c utils.c COMPFLAGS="\$COMPFLAGS -fopenmp -Wall -std=c99"
 movefile('TV_energy.mex*',Pathmove);
@@ -73,9 +79,7 @@ fprintf('%s \n', 'Compiling Nonlocal marching method for inpaiting...');
 mex NonlocalMarching_Inpaint.c NonlocalMarching_Inpaint_core.c utils.c COMPFLAGS="\$COMPFLAGS -fopenmp -Wall -std=c99"
 movefile('NonlocalMarching_Inpaint.mex*',Pathmove);
 
-delete SB_TV_core* ROF_TV_core* FGP_TV_core* FGP_dTV_core* TNV_core* utils* Diffusion_core* Diffus4th_order_core* TGV_core* CCPiDefines.h
-delete Diffusion_Inpaint_core* NonlocalMarching_Inpaint_core*
-fprintf('%s \n', 'Regularisers successfully compiled!');
+
 %%
 %%% The second approach to compile using TDM-GCC which follows this
 %%% discussion:
@@ -105,15 +109,24 @@ fprintf('%s \n', 'Regularisers successfully compiled!');
 % movefile('TGV.mex*',Pathmove);
 % mex C:\TDMGCC\lib\gcc\x86_64-w64-mingw32\5.1.0\libgomp.a CXXFLAGS="$CXXFLAGS -std=c++11 -fopenmp" LLT_ROF.c LLT_ROF_core.c utils.c
 % movefile('LLT_ROF.mex*',Pathmove);
+% mex C:\TDMGCC\lib\gcc\x86_64-w64-mingw32\5.1.0\libgomp.a CXXFLAGS="$CXXFLAGS -std=c++11 -fopenmp" PatchSelect.c PatchSelect_core.c utils.c
+% mex C:\TDMGCC\lib\gcc\x86_64-w64-mingw32\5.1.0\libgomp.a CXXFLAGS="$CXXFLAGS -std=c++11 -fopenmp" Nonlocal_TV.c Nonlocal_TV_core.c utils.c
+% movefile('Nonlocal_TV.mex*',Pathmove);
+% movefile('PatchSelect.mex*',Pathmove);
 % mex C:\TDMGCC\lib\gcc\x86_64-w64-mingw32\5.1.0\libgomp.a CXXFLAGS="$CXXFLAGS -std=c++11 -fopenmp" TV_energy.c utils.c
 % movefile('TV_energy.mex*',Pathmove);
 % mex C:\TDMGCC\lib\gcc\x86_64-w64-mingw32\5.1.0\libgomp.a CXXFLAGS="$CXXFLAGS -std=c++11 -fopenmp" NonlDiff_Inp.c Diffusion_Inpaint_core.c utils.c
 % movefile('NonlDiff_Inp.mex*',Pathmove);
 % mex C:\TDMGCC\lib\gcc\x86_64-w64-mingw32\5.1.0\libgomp.a CXXFLAGS="$CXXFLAGS -std=c++11 -fopenmp" NonlocalMarching_Inpaint.c NonlocalMarching_Inpaint_core.c utils.c
 % movefile('NonlocalMarching_Inpaint.mex*',Pathmove);
-% delete SB_TV_core* ROF_TV_core* FGP_TV_core* FGP_dTV_core* TNV_core* utils* Diffusion_core* Diffus4th_order_core* TGV_core* CCPiDefines.h
-% delete Diffusion_Inpaint_core* NonlocalMarching_Inpaint_core*
-% fprintf('%s \n', 'Regularisers successfully compiled!');
+
+
+delete SB_TV_core* ROF_TV_core* FGP_TV_core* FGP_dTV_core* TNV_core* utils* Diffusion_core* Diffus4th_order_core* TGV_core* CCPiDefines.h
+delete PatchSelect_core* Nonlocal_TV_core*
+delete Diffusion_Inpaint_core* NonlocalMarching_Inpaint_core*
+fprintf('%s \n', 'Regularisers successfully compiled!');
+
+
 %%
 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
 
diff --git a/Wrappers/Matlab/mex_compile/regularisers_CPU/Nonlocal_TV.c b/Wrappers/Matlab/mex_compile/regularisers_CPU/Nonlocal_TV.c
new file mode 100644
index 0000000..014c0a0
--- /dev/null
+++ b/Wrappers/Matlab/mex_compile/regularisers_CPU/Nonlocal_TV.c
@@ -0,0 +1,88 @@
+/*
+ * This work is part of the Core Imaging Library developed by
+ * Visual Analytics and Imaging System Group of the Science Technology
+ * Facilities Council, STFC and Diamond Light Source Ltd. 
+ *
+ * Copyright 2017 Daniil Kazantsev
+ * Copyright 2017 Srikanth Nagella, Edoardo Pasca
+ * Copyright 2018 Diamond Light Source Ltd. 
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ * http://www.apache.org/licenses/LICENSE-2.0
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "matrix.h"
+#include "mex.h"
+#include "Nonlocal_TV_core.h"
+
+#define EPS 1.0000e-9
+
+/* Matlab wrapper for C-OMP implementation of non-local regulariser
+ * Weights and associated indices must be given as an input.
+ * Gauss-Seidel fixed point iteration requires ~ 3 iterations, so the main effort
+ * goes in pre-calculation of weights and selection of patches
+ *
+ *
+ * Input Parameters:
+ * 1. 2D/3D grayscale image/volume
+ * 2. AR_i - indeces of i neighbours
+ * 3. AR_j - indeces of j neighbours
+ * 4. AR_k - indeces of k neighbours (0 - for 2D case)
+ * 5. Weights_ij(k) - associated weights 
+ * 6. regularisation parameter
+ * 7. iterations number 
+ 
+ * Output:
+ * 1. denoised image/volume 	
+ * Elmoataz, Abderrahim, Olivier Lezoray, and Sébastien Bougleux. "Nonlocal discrete regularization on weighted graphs: a framework for image and manifold processing." IEEE Trans. Image Processing 17, no. 7 (2008): 1047-1060.
+ */
+
+void mexFunction(
+        int nlhs, mxArray *plhs[],
+        int nrhs, const mxArray *prhs[])
+{
+    long number_of_dims,  dimX, dimY, dimZ;
+    int IterNumb, NumNeighb = 0;
+    unsigned short *H_i, *H_j, *H_k;
+    const int  *dim_array;
+    const int  *dim_array2;
+    float *A_orig, *Output=NULL, *Weights, lambda;
+    
+    dim_array = mxGetDimensions(prhs[0]);
+    dim_array2 = mxGetDimensions(prhs[1]);
+    number_of_dims = mxGetNumberOfDimensions(prhs[0]);
+    
+    /*Handling Matlab input data*/
+    A_orig  = (float *) mxGetData(prhs[0]); /* a 2D image or a set of 2D images (3D stack) */
+    H_i  = (unsigned short *) mxGetData(prhs[1]); /* indeces of i neighbours */
+    H_j  = (unsigned short *) mxGetData(prhs[2]); /* indeces of j neighbours */
+    H_k  = (unsigned short *) mxGetData(prhs[3]); /* indeces of k neighbours */
+    Weights = (float *) mxGetData(prhs[4]); /* weights for patches */
+    lambda = (float) mxGetScalar(prhs[5]); /* regularisation parameter */
+    IterNumb = (int) mxGetScalar(prhs[6]); /* the number of iterations */
+ 
+    dimX = dim_array[0]; dimY = dim_array[1]; dimZ = dim_array[2];   
+         
+    /*****2D INPUT *****/
+    if (number_of_dims == 2) {
+        dimZ = 0;   
+        NumNeighb = dim_array2[2];
+        Output = (float*)mxGetPr(plhs[0] = mxCreateNumericArray(2, dim_array, mxSINGLE_CLASS, mxREAL));  
+        }
+    /*****3D INPUT *****/
+    /****************************************************/
+    if (number_of_dims == 3) {
+        NumNeighb = dim_array2[3];
+        Output = (float*)mxGetPr(plhs[0] = mxCreateNumericArray(3, dim_array, mxSINGLE_CLASS, mxREAL));
+    }
+    
+    /* run the main function here */
+    Nonlocal_TV_CPU_main(A_orig, Output, H_i, H_j, H_k, Weights, dimX, dimY, dimZ, NumNeighb, lambda, IterNumb);
+}
diff --git a/Wrappers/Matlab/mex_compile/regularisers_CPU/PatchSelect.c b/Wrappers/Matlab/mex_compile/regularisers_CPU/PatchSelect.c
new file mode 100644
index 0000000..f942539
--- /dev/null
+++ b/Wrappers/Matlab/mex_compile/regularisers_CPU/PatchSelect.c
@@ -0,0 +1,92 @@
+/*
+ * This work is part of the Core Imaging Library developed by
+ * Visual Analytics and Imaging System Group of the Science Technology
+ * Facilities Council, STFC and Diamond Light Source Ltd. 
+ *
+ * Copyright 2017 Daniil Kazantsev
+ * Copyright 2017 Srikanth Nagella, Edoardo Pasca
+ * Copyright 2018 Diamond Light Source Ltd. 
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ * http://www.apache.org/licenses/LICENSE-2.0
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "matrix.h"
+#include "mex.h"
+#include "PatchSelect_core.h"
+
+/* C-OMP implementation of non-local weight pre-calculation for non-local priors
+ * Weights and associated indices are stored into pre-allocated arrays and passed
+ * to the regulariser
+ *
+ *
+ * Input Parameters:
+ * 1. 2D/3D grayscale image/volume
+ * 2. Searching window (half-size of the main bigger searching window, e.g. 11)
+ * 3. Similarity window (half-size of the patch window, e.g. 2)
+ * 4. The number of neighbours to take (the most prominent after sorting neighbours will be taken)
+ * 5. noise-related parameter to calculate non-local weights
+ *
+ * Output [2D]:
+ * 1. AR_i - indeces of i neighbours
+ * 2. AR_j - indeces of j neighbours
+ * 3. Weights_ij - associated weights
+ *
+ * Output [3D]:
+ * 1. AR_i - indeces of i neighbours
+ * 2. AR_j - indeces of j neighbours
+ * 3. AR_k - indeces of j neighbours
+ * 4. Weights_ijk - associated weights
+ */
+/**************************************************/
+void mexFunction(
+        int nlhs, mxArray *plhs[],
+        int nrhs, const mxArray *prhs[])
+{
+    int number_of_dims,  SearchWindow, SimilarWin, NumNeighb;
+    mwSize dimX, dimY, dimZ;
+    unsigned short *H_i=NULL, *H_j=NULL, *H_k=NULL;
+    const int  *dim_array;
+    float *A, *Weights = NULL, h;
+    int dim_array2[3]; /* for 2D data */
+    int dim_array3[4]; /* for 3D data */
+    
+    dim_array = mxGetDimensions(prhs[0]);
+    number_of_dims = mxGetNumberOfDimensions(prhs[0]);
+    
+    /*Handling Matlab input data*/
+    A  = (float *) mxGetData(prhs[0]); /* a 2D or 3D image/volume */
+    SearchWindow = (int) mxGetScalar(prhs[1]);    /* Large Searching window */
+    SimilarWin = (int) mxGetScalar(prhs[2]);    /* Similarity window (patch-search)*/
+    NumNeighb = (int) mxGetScalar(prhs[3]); /* the total number of neighbours to take */
+    h = (float) mxGetScalar(prhs[4]); /* NLM parameter */
+
+    dimX = dim_array[0]; dimY = dim_array[1]; dimZ = dim_array[2];
+    dim_array2[0] = dimX; dim_array2[1] = dimY; dim_array2[2] = NumNeighb;  /* 2D case */
+    dim_array3[0] = dimX; dim_array3[1] = dimY; dim_array3[2] = dimZ; dim_array3[3] = NumNeighb;  /* 3D case */
+    
+    /****************2D INPUT ***************/
+    if (number_of_dims == 2) {
+        dimZ = 0;               
+        H_i = (unsigned short*)mxGetPr(plhs[0] = mxCreateNumericArray(3, dim_array2, mxUINT16_CLASS, mxREAL));
+        H_j = (unsigned short*)mxGetPr(plhs[1] = mxCreateNumericArray(3, dim_array2, mxUINT16_CLASS, mxREAL));
+        Weights = (float*)mxGetPr(plhs[2] = mxCreateNumericArray(3, dim_array2, mxSINGLE_CLASS, mxREAL));
+        }
+    /****************3D INPUT ***************/
+    if (number_of_dims == 3) {        
+        H_i = (unsigned short*)mxGetPr(plhs[0] = mxCreateNumericArray(4, dim_array3, mxUINT16_CLASS, mxREAL));
+        H_j = (unsigned short*)mxGetPr(plhs[1] = mxCreateNumericArray(4, dim_array3, mxUINT16_CLASS, mxREAL));
+        H_k = (unsigned short*)mxGetPr(plhs[2] = mxCreateNumericArray(4, dim_array3, mxUINT16_CLASS, mxREAL));
+        Weights = (float*)mxGetPr(plhs[3] = mxCreateNumericArray(4, dim_array3, mxSINGLE_CLASS, mxREAL));        
+    }
+    
+    PatchSelect_CPU_main(A, H_i, H_j, H_k, Weights, (long)(dimX), (long)(dimY), (long)(dimZ), SearchWindow, SimilarWin, NumNeighb, h, 0); 
+    
+ }
diff --git a/Wrappers/Python/ccpi/filters/regularisers.py b/Wrappers/Python/ccpi/filters/regularisers.py
index c7ae808..bf7e23c 100644
--- a/Wrappers/Python/ccpi/filters/regularisers.py
+++ b/Wrappers/Python/ccpi/filters/regularisers.py
@@ -2,7 +2,7 @@
 script which assigns a proper device core function based on a flag ('cpu' or 'gpu')
 """
 
-from ccpi.filters.cpu_regularisers import TV_ROF_CPU, TV_FGP_CPU, TV_SB_CPU, dTV_FGP_CPU, TNV_CPU, NDF_CPU, Diff4th_CPU, TGV_CPU, LLT_ROF_CPU
+from ccpi.filters.cpu_regularisers import TV_ROF_CPU, TV_FGP_CPU, TV_SB_CPU, dTV_FGP_CPU, TNV_CPU, NDF_CPU, Diff4th_CPU, TGV_CPU, LLT_ROF_CPU, PATCHSEL_CPU, NLTV_CPU
 try:
     from ccpi.filters.gpu_regularisers import TV_ROF_GPU, TV_FGP_GPU, TV_SB_GPU, dTV_FGP_GPU, NDF_GPU, Diff4th_GPU, TGV_GPU, LLT_ROF_GPU
     gpu_enabled = True
@@ -144,6 +144,31 @@ def DIFF4th(inputData, regularisation_parameter, edge_parameter, iterations,
     	    raise ValueError ('GPU is not available')
         raise ValueError('Unknown device {0}. Expecting gpu or cpu'\
                          .format(device))
+        
+def PatchSelect(inputData, searchwindow, patchwindow, neighbours, edge_parameter, device='cpu'):
+    if device == 'cpu':
+        return PATCHSEL_CPU(inputData,
+                     searchwindow,
+                     patchwindow,
+                     neighbours, 
+                     edge_parameter)
+    elif device == 'gpu' and gpu_enabled:
+        return 1
+    else:
+        if not gpu_enabled and device == 'gpu':
+    	    raise ValueError ('GPU is not available')
+        raise ValueError('Unknown device {0}. Expecting gpu or cpu'\
+                         .format(device))
+
+def NLTV(inputData, H_i, H_j, H_k, Weights, regularisation_parameter, iterations):
+    return NLTV_CPU(inputData,
+                     H_i,
+                     H_j,
+                     H_k, 
+                     Weights,
+                     regularisation_parameter,
+                     iterations)
+
 def TGV(inputData, regularisation_parameter, alpha1, alpha0, iterations,
                      LipshitzConst, device='cpu'):
     if device == 'cpu':
diff --git a/Wrappers/Python/demos/demo_cpu_regularisers.py b/Wrappers/Python/demos/demo_cpu_regularisers.py
index e99b271..31e4cad 100644
--- a/Wrappers/Python/demos/demo_cpu_regularisers.py
+++ b/Wrappers/Python/demos/demo_cpu_regularisers.py
@@ -13,6 +13,7 @@ import numpy as np
 import os
 import timeit
 from ccpi.filters.regularisers import ROF_TV, FGP_TV, SB_TV, TGV, LLT_ROF, FGP_dTV, TNV, NDF, DIFF4th
+from ccpi.filters.regularisers import PatchSelect, NLTV
 from qualitymetrics import rmse
 ###############################################################################
 def printParametersToString(pars):
@@ -350,7 +351,7 @@ a.text(0.15, 0.25, txtstr, transform=a.transAxes, fontsize=14,
 imgplot = plt.imshow(ndf_cpu, cmap="gray")
 plt.title('{}'.format('CPU results'))
 
-
+#%%
 print ("%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%")
 print ("___Anisotropic Diffusion 4th Order (2D)____")
 print ("%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%")
@@ -395,7 +396,71 @@ a.text(0.15, 0.25, txtstr, transform=a.transAxes, fontsize=14,
 imgplot = plt.imshow(diff4_cpu, cmap="gray")
 plt.title('{}'.format('CPU results'))
 
+#%%
+print ("%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%")
+print ("___Nonlocal patches pre-calculation____")
+print ("%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%")
+# set parameters
+pars = {'algorithm' : PatchSelect, \
+        'input' : u0,\
+        'searchwindow': 7, \
+        'patchwindow': 2,\
+        'neighbours' : 15 ,\
+        'edge_parameter':0.23}
+
+H_i, H_j, Weights = PatchSelect(pars['input'], 
+              pars['searchwindow'],
+              pars['patchwindow'], 
+              pars['neighbours'],
+              pars['edge_parameter'],'cpu')
+
+#%%
+print ("%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%")
+print ("___Nonlocal Total Variation penalty____")
+print ("%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%")
+## plot 
+fig = plt.figure()
+plt.suptitle('Performance of NLTV regulariser using the CPU')
+a=fig.add_subplot(1,2,1)
+a.set_title('Noisy Image')
+imgplot = plt.imshow(u0,cmap="gray")
+
+pars2 = {'algorithm' : NLTV, \
+        'input' : u0,\
+        'H_i': H_i, \
+        'H_j': H_j,\
+        'H_k' : 0,\
+        'Weights' : Weights,\
+        'regularisation_parameter': 0.085,\
+        'iterations': 2
+        }
+#%%
+start_time = timeit.default_timer()
+nltv_cpu = NLTV(pars2['input'], 
+              pars2['H_i'],
+              pars2['H_j'], 
+              pars2['H_k'],
+              pars2['Weights'],
+              pars2['regularisation_parameter'],
+              pars2['iterations'])
+
+rms = rmse(Im, nltv_cpu)
+pars['rmse'] = rms
+
+txtstr = printParametersToString(pars)
+txtstr += "%s = %.3fs" % ('elapsed time',timeit.default_timer() - start_time)
+print (txtstr)
+a=fig.add_subplot(1,2,2)
 
+# these are matplotlib.patch.Patch properties
+props = dict(boxstyle='round', facecolor='wheat', alpha=0.75)
+# place a text box in upper left in axes coords
+a.text(0.15, 0.25, txtstr, transform=a.transAxes, fontsize=14,
+         verticalalignment='top', bbox=props)
+imgplot = plt.imshow(nltv_cpu, cmap="gray")
+plt.title('{}'.format('CPU results'))
+
+#%%
 print ("%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%")
 print ("_____________FGP-dTV (2D)__________________")
 print ("%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%")
@@ -447,7 +512,7 @@ a.text(0.15, 0.25, txtstr, transform=a.transAxes, fontsize=14,
          verticalalignment='top', bbox=props)
 imgplot = plt.imshow(fgp_dtv_cpu, cmap="gray")
 plt.title('{}'.format('CPU results'))
-
+#%%
 print ("%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%")
 print ("__________Total nuclear Variation__________")
 print ("%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%")
diff --git a/Wrappers/Python/src/cpu_regularisers.pyx b/Wrappers/Python/src/cpu_regularisers.pyx
index bf9c861..e51e6d8 100644
--- a/Wrappers/Python/src/cpu_regularisers.pyx
+++ b/Wrappers/Python/src/cpu_regularisers.pyx
@@ -27,6 +27,8 @@ cdef extern float Diffusion_CPU_main(float *Input, float *Output, float lambdaPa
 cdef extern float Diffus4th_CPU_main(float *Input, float *Output, float lambdaPar, float sigmaPar, int iterationsNumb, float tau, int dimX, int dimY, int dimZ);
 cdef extern float TNV_CPU_main(float *Input, float *u, float lambdaPar, int maxIter, float tol, int dimX, int dimY, int dimZ);
 cdef extern float dTV_FGP_CPU_main(float *Input, float *InputRef, float *Output, float lambdaPar, int iterationsNumb, float epsil, float eta, int methodTV, int nonneg, int printM, int dimX, int dimY, int dimZ);
+cdef extern float PatchSelect_CPU_main(float *Input, unsigned short *H_i, unsigned short *H_j, unsigned short *H_k, float *Weights, int dimX, int dimY, int dimZ, int SearchWindow, int SimilarWin, int NumNeighb, float h, int switchM);
+cdef extern float Nonlocal_TV_CPU_main(float *A_orig, float *Output, unsigned short *H_i, unsigned short *H_j, unsigned short *H_k, float *Weights, int dimX, int dimY, int dimZ, int NumNeighb, float lambdaReg, int IterNumb);
 
 cdef extern float Diffusion_Inpaint_CPU_main(float *Input, unsigned char *Mask, float *Output, float lambdaPar, float sigmaPar, int iterationsNumb, float tau, int penaltytype, int dimX, int dimY, int dimZ);
 cdef extern float NonlocalMarching_Inpaint_main(float *Input, unsigned char *M, float *Output, unsigned char *M_upd, int SW_increment, int iterationsNumb, int trigger, int dimX, int dimY, int dimZ);
@@ -446,6 +448,94 @@ def Diff4th_3D(np.ndarray[np.float32_t, ndim=3, mode="c"] inputData,
     Diffus4th_CPU_main(&inputData[0,0,0], &outputData[0,0,0], regularisation_parameter, edge_parameter, iterationsNumb, time_marching_parameter, dims[2], dims[1], dims[0])
 
     return outputData
+
+#****************************************************************#
+#***************Patch-based weights calculation******************#
+#****************************************************************#
+def PATCHSEL_CPU(inputData, searchwindow, patchwindow, neighbours, edge_parameter):
+    if inputData.ndim == 2:
+        return PatchSel_2D(inputData, searchwindow, patchwindow, neighbours, edge_parameter)
+    elif inputData.ndim == 3:
+        return PatchSel_3D(inputData, searchwindow, patchwindow, neighbours, edge_parameter)
+def PatchSel_2D(np.ndarray[np.float32_t, ndim=2, mode="c"] inputData,
+                     int searchwindow,
+                     int patchwindow,
+                     int neighbours,
+                     float edge_parameter):
+    cdef long dims[3]
+    dims[0] = neighbours
+    dims[1] = inputData.shape[0]
+    dims[2] = inputData.shape[1]
+    
+    
+    cdef np.ndarray[np.float32_t, ndim=3, mode="c"] Weights = \
+            np.zeros([dims[0], dims[1],dims[2]], dtype='float32')
+    
+    cdef np.ndarray[np.uint16_t, ndim=3, mode="c"] H_i = \
+            np.zeros([dims[0], dims[1],dims[2]], dtype='uint16')
+            
+    cdef np.ndarray[np.uint16_t, ndim=3, mode="c"] H_j = \
+            np.zeros([dims[0], dims[1],dims[2]], dtype='uint16')
+
+    # Run patch-based weight selection function
+    PatchSelect_CPU_main(&inputData[0,0], &H_j[0,0,0], &H_i[0,0,0], &H_i[0,0,0], &Weights[0,0,0], dims[2], dims[1], 0, searchwindow, patchwindow,  neighbours,  edge_parameter, 1)
+    return H_i, H_j, Weights
+
+def PatchSel_3D(np.ndarray[np.float32_t, ndim=3, mode="c"] inputData,
+                     int searchwindow,
+                     int patchwindow,
+                     int neighbours,
+                     float edge_parameter):
+    cdef long dims[4]
+    dims[0] = inputData.shape[0]
+    dims[1] = inputData.shape[1]
+    dims[2] = inputData.shape[2]
+    dims[3] = neighbours
+    
+    cdef np.ndarray[np.float32_t, ndim=4, mode="c"] Weights = \
+            np.zeros([dims[3],dims[0],dims[1],dims[2]], dtype='float32')
+    
+    cdef np.ndarray[np.uint16_t, ndim=4, mode="c"] H_i = \
+            np.zeros([dims[3],dims[0],dims[1],dims[2]], dtype='uint16')
+            
+    cdef np.ndarray[np.uint16_t, ndim=4, mode="c"] H_j = \
+            np.zeros([dims[3],dims[0],dims[1],dims[2]], dtype='uint16')
+            
+    cdef np.ndarray[np.uint16_t, ndim=4, mode="c"] H_k = \
+            np.zeros([dims[3],dims[0],dims[1],dims[2]], dtype='uint16')
+
+    # Run patch-based weight selection function
+    PatchSelect_CPU_main(&inputData[0,0,0], &H_i[0,0,0,0], &H_j[0,0,0,0], &H_k[0,0,0,0], &Weights[0,0,0,0], dims[2], dims[1], dims[0], searchwindow, patchwindow,  neighbours, edge_parameter, 1)
+    return H_i, H_j, H_k, Weights
+
+
+#****************************************************************#
+#***************Non-local Total Variation******************#
+#****************************************************************#
+def NLTV_CPU(inputData, H_i, H_j, H_k, Weights, regularisation_parameter, iterations):
+    if inputData.ndim == 2:
+        return NLTV_2D(inputData, H_i, H_j, Weights, regularisation_parameter, iterations)
+    elif inputData.ndim == 3:
+        return 1
+def NLTV_2D(np.ndarray[np.float32_t, ndim=2, mode="c"] inputData,
+                     np.ndarray[np.uint16_t, ndim=3, mode="c"] H_i,
+                     np.ndarray[np.uint16_t, ndim=3, mode="c"] H_j,
+                     np.ndarray[np.float32_t, ndim=3, mode="c"] Weights,
+                     float regularisation_parameter,
+                     int iterations):
+
+    cdef long dims[2]
+    dims[0] = inputData.shape[0]
+    dims[1] = inputData.shape[1]
+    neighbours = H_i.shape[0]
+    
+    cdef np.ndarray[np.float32_t, ndim=2, mode="c"] outputData = \
+            np.zeros([dims[0],dims[1]], dtype='float32')
+    
+    # Run nonlocal TV regularisation
+    Nonlocal_TV_CPU_main(&inputData[0,0], &outputData[0,0], &H_i[0,0,0], &H_j[0,0,0], &H_i[0,0,0], &Weights[0,0,0], dims[1], dims[0], 0, neighbours, regularisation_parameter, iterations)
+    return outputData
+
 #*********************Inpainting WITH****************************#
 #***************Nonlinear (Isotropic) Diffusion******************#
 #****************************************************************#