Instituting denoise branch to work on improvements to RT denoise. Most definitely a work in progress, but the patch is becoming quite complicated so it is definitely worth having it backed up on Google code.
This commit is contained in:
Emil Martinec
486
rtengine/dirpyrLab_denoise.cc
Normal file
486
rtengine/dirpyrLab_denoise.cc
Normal file
@@ -0,0 +1,486 @@
|
||||
/*
|
||||
* This file is part of RawTherapee.
|
||||
*
|
||||
* RawTherapee is free software: you can redistribute it and/or modify
|
||||
* it under the terms of the GNU General Public License as published by
|
||||
* the Free Software Foundation, either version 3 of the License, or
|
||||
* (at your option) any later version.
|
||||
*
|
||||
* RawTherapee is distributed in the hope that it will be useful,
|
||||
* but WITHOUT ANY WARRANTY; without even the implied warranty of
|
||||
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
||||
* GNU General Public License for more details.
|
||||
*
|
||||
* You should have received a copy of the GNU General Public License
|
||||
* along with RawTherapee. If not, see <http://www.gnu.org/licenses/>.
|
||||
*
|
||||
* <20> 2010 Emil Martinec <ejmartin@uchicago.edu>
|
||||
*
|
||||
*/
|
||||
|
||||
//#include "rtengine.h"
|
||||
#include <cstddef>
|
||||
#include <cmath>
|
||||
#include "curves.h"
|
||||
#include "labimage.h"
|
||||
#include "improcfun.h"
|
||||
#include "array2D.h"
|
||||
|
||||
#ifdef _OPENMP
|
||||
#include <omp.h>
|
||||
#endif
|
||||
|
||||
#define SQR(x) ((x)*(x))
|
||||
#define CLIPTO(a,b,c) ((a)>(b)?((a)<(c)?(a):(c)):(b))
|
||||
#define CLIPC(a) ((a)>-32000?((a)<32000?(a):32000):-32000)
|
||||
#define CLIP(a) (CLIPTO(a,0,65535))
|
||||
|
||||
|
||||
#define DIRWT_L(i1,j1,i,j) ( rangefn_L[abs(data_fine->L[i1][j1]-data_fine->L[i][j])] )
|
||||
|
||||
#define DIRWT_AB(i1,j1,i,j) ( rangefn_ab[/*abs(data_fine->L[i1][j1]-data_fine->L[i][j])*/0 + \
|
||||
abs(data_fine->a[i1][j1]-data_fine->a[i][j]) + \
|
||||
abs(data_fine->b[i1][j1]-data_fine->b[i][j])] )
|
||||
|
||||
|
||||
#define NRWT_AB (nrwt_ab[abs(hipass[1])] * nrwt_ab[abs(hipass[2])])
|
||||
|
||||
|
||||
#define med3(a,b,c) (a<b ? (b<c ? b : (a<c ? c : a)) : (a<c ? a : (b<c ? c : b)))
|
||||
|
||||
#define hmf(a11,a12,a13,a21,a22,a23,a31,a32,a33) \
|
||||
(med3(a22,med3(a22,med3(a12,a22,a32),med3(a21,a22,a23)), \
|
||||
med3(a22,med3(a11,a22,a33),med3(a13,a22,a31))) )
|
||||
|
||||
#define PIX_SORT(a,b) { if ((a)>(b)) {temp=(a);(a)=(b);(b)=temp;} }
|
||||
|
||||
#define med3x3(a0,a1,a2,a3,a4,a5,a6,a7,a8,median) { \
|
||||
p[0]=a0; p[1]=a1; p[2]=a2; p[3]=a3; p[4]=a4; p[5]=a5; p[6]=a6; p[7]=a7; p[8]=a8; \
|
||||
PIX_SORT(p[1],p[2]); PIX_SORT(p[4],p[5]); PIX_SORT(p[7],p[8]); \
|
||||
PIX_SORT(p[0],p[1]); PIX_SORT(p[3],p[4]); PIX_SORT(p[6],p[7]); \
|
||||
PIX_SORT(p[1],p[2]); PIX_SORT(p[4],p[5]); PIX_SORT(p[7],p[8]); \
|
||||
PIX_SORT(p[0],p[3]); PIX_SORT(p[5],p[8]); PIX_SORT(p[4],p[7]); \
|
||||
PIX_SORT(p[3],p[6]); PIX_SORT(p[1],p[4]); PIX_SORT(p[2],p[5]); \
|
||||
PIX_SORT(p[4],p[7]); PIX_SORT(p[4],p[2]); PIX_SORT(p[6],p[4]); \
|
||||
PIX_SORT(p[4],p[2]); median=p[4];} //a4 is the median
|
||||
|
||||
|
||||
namespace rtengine {
|
||||
|
||||
static const int maxlevel = 4;
|
||||
|
||||
//sequence of scales
|
||||
//static const int scales[8] = {1,2,4,8,16,32,64,128};
|
||||
//sequence of pitches
|
||||
//static const int pitches[8] = {1,1,1,1,1,1,1,1};
|
||||
|
||||
//sequence of scales
|
||||
//static const int scales[8] = {1,1,1,1,1,1,1,1};
|
||||
//sequence of pitches
|
||||
//static const int pitches[8] = {2,2,2,2,2,2,2,2};
|
||||
|
||||
//sequence of scales
|
||||
//static const int scales[8] = {1,1,2,2,4,4,8,8};
|
||||
//sequence of pitches
|
||||
//static const int pitches[8] = {2,1,2,1,2,1,2,1};
|
||||
|
||||
//sequence of scales
|
||||
static const int scales[8] = {1,1,2,4,8,16,32,64};
|
||||
//sequence of pitches
|
||||
static const int pitches[8] = {2,1,1,1,1,1,1,1};
|
||||
|
||||
//pitch is spacing of subsampling
|
||||
//scale is spacing of directional averaging weights
|
||||
//example 1: no subsampling at any level -- pitch=1, scale=2^n
|
||||
//example 2: subsampling by 2 every level -- pitch=2, scale=1 at each level
|
||||
//example 3: no subsampling at first level, subsampling by 2 thereafter --
|
||||
// pitch =1, scale=1 at first level; pitch=2, scale=2 thereafter
|
||||
|
||||
|
||||
|
||||
|
||||
void ImProcFunctions :: dirpyrLab_denoise(LabImage * src, LabImage * dst, const procparams::DirPyrDenoiseParams & dnparams )
|
||||
{
|
||||
|
||||
|
||||
//%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|
||||
|
||||
|
||||
|
||||
LUTf rangefn_L(65536);
|
||||
LUTf nrwt_l(1);
|
||||
|
||||
LUTf rangefn_ab(65536);
|
||||
LUTf nrwt_ab(65536);
|
||||
|
||||
//set up NR weight functions
|
||||
|
||||
float noise_L = 10.0*dnparams.luma;
|
||||
float noisevar_L = SQR(noise_L);
|
||||
|
||||
float noise_ab = 10.0*dnparams.chroma;
|
||||
float noisevar_ab = SQR(noise_ab);
|
||||
|
||||
|
||||
//set up range functions
|
||||
for (int i=0; i<65536; i++)
|
||||
rangefn_L[i] = exp(-(double)i / (1.0f+noise_L));// * (1.0+noisevar_L)/((double)(i*i) + noisevar_L+1.0);
|
||||
for (int i=0; i<65536; i++)
|
||||
rangefn_ab[i]= exp(-SQR((double)i) / (1.0f+3*noisevar_ab));// * (1.0+noisevar_ab)/((double)(i*i) + noisevar_ab+1.0);
|
||||
|
||||
|
||||
for (int i=0; i<65536; i++)
|
||||
nrwt_ab[i] = ((1.0+abs(i-32768)/(1.0+8*noise_ab)) * exp(-(double)fabs(i-32768)/ (1.0+8*noise_ab) ) );
|
||||
|
||||
|
||||
//for (int i=0; i<65536; i+=100) printf("%d %d \n",i,gamcurve[i]);
|
||||
|
||||
|
||||
//%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|
||||
|
||||
|
||||
|
||||
int level;
|
||||
|
||||
LabImage * dirpyrLablo[maxlevel];
|
||||
int w = (int)((src->W-1)/pitches[0])+1;
|
||||
int h = (int)((src->H-1)/pitches[0])+1;
|
||||
dirpyrLablo[0] = new LabImage(w, h);
|
||||
for (level=1; level<maxlevel; level++) {
|
||||
w = (int)((w-1)/pitches[level])+1;
|
||||
h = (int)((h-1)/pitches[level])+1;
|
||||
dirpyrLablo[level] = new LabImage(w, h);
|
||||
};
|
||||
|
||||
|
||||
//////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
level = 0;
|
||||
|
||||
int scale = scales[level];
|
||||
int pitch = pitches[level];
|
||||
//int thresh = 10 * c[8];
|
||||
//impulse_nr (src, src, m_w1, m_h1, thresh, noisevar);
|
||||
|
||||
dirpyr(src, dirpyrLablo[0], 0, rangefn_L, rangefn_ab, pitch, scale, dnparams.luma, dnparams.chroma );
|
||||
|
||||
level = 1;
|
||||
|
||||
while(level < maxlevel)
|
||||
{
|
||||
scale = scales[level];
|
||||
pitch = pitches[level];
|
||||
|
||||
dirpyr(dirpyrLablo[level-1], dirpyrLablo[level], level, rangefn_L, rangefn_ab, pitch, scale, dnparams.luma, dnparams.chroma );
|
||||
|
||||
level ++;
|
||||
}
|
||||
|
||||
|
||||
//%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|
||||
|
||||
|
||||
|
||||
for(int level = maxlevel - 1; level > 0; level--)
|
||||
{
|
||||
|
||||
int scale = scales[level];
|
||||
int pitch = pitches[level];
|
||||
idirpyr(dirpyrLablo[level], dirpyrLablo[level-1], level, rangefn_L, nrwt_l, nrwt_ab, pitch, scale, dnparams.luma, dnparams.chroma/*, Lcurve, abcurve*/ );
|
||||
}
|
||||
|
||||
|
||||
scale = scales[0];
|
||||
pitch = pitches[0];
|
||||
// freeing as much memory as possible since the next call to idirpyr will need lots
|
||||
for(int i = 1; i < maxlevel; i++) {
|
||||
delete dirpyrLablo[i];
|
||||
}
|
||||
|
||||
idirpyr(dirpyrLablo[0], dst, 0, rangefn_L, nrwt_l, nrwt_ab, pitch, scale, dnparams.luma, dnparams.chroma/*, Lcurve, abcurve*/ );
|
||||
|
||||
// freeing the last bunch of memory
|
||||
delete dirpyrLablo[0];
|
||||
|
||||
//%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|
||||
|
||||
|
||||
};
|
||||
|
||||
void ImProcFunctions::dirpyr(LabImage* data_fine, LabImage* data_coarse, int level, \
|
||||
LUTf & rangefn_L, LUTf & rangefn_ab, int pitch, int scale, \
|
||||
const int luma, const int chroma )
|
||||
{
|
||||
|
||||
//pitch is spacing of subsampling
|
||||
//scale is spacing of directional averaging weights
|
||||
//example 1: no subsampling at any level -- pitch=1, scale=2^n
|
||||
//example 2: subsampling by 2 every level -- pitch=2, scale=1 at each level
|
||||
//example 3: no subsampling at first level, subsampling by 2 thereafter --
|
||||
// pitch =1, scale=1 at first level; pitch=2, scale=2 thereafter
|
||||
|
||||
|
||||
//%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|
||||
// calculate weights, compute directionally weighted average
|
||||
|
||||
int width = data_fine->W;
|
||||
int height = data_fine->H;
|
||||
|
||||
//generate domain kernel
|
||||
int halfwin = 3;//MIN(ceil(2*sig),3);
|
||||
int scalewin = halfwin*scale;
|
||||
|
||||
#ifdef _OPENMP
|
||||
#pragma omp parallel for
|
||||
#endif
|
||||
|
||||
for(int i = 0; i < height; i+=pitch ) { int i1=i/pitch;
|
||||
for(int j = 0, j1=0; j < width; j+=pitch, j1++)
|
||||
{
|
||||
float dirwt_l, dirwt_ab, norm_l, norm_ab;
|
||||
//float Lmed,Lhmf;
|
||||
//float lops,aops,bops;
|
||||
float Lout, aout, bout;
|
||||
norm_l = norm_ab = 0;//if we do want to include the input pixel in the sum
|
||||
Lout = 0;
|
||||
aout = 0;
|
||||
bout = 0;
|
||||
|
||||
for(int inbr=MAX(0,(i-scalewin)); inbr<=MIN(height-1,(i+scalewin)); inbr+=scale) {
|
||||
for (int jnbr=MAX(0,(j-scalewin)); jnbr<=MIN(width-1,(j+scalewin)); jnbr+=scale) {
|
||||
/*for(int inbr=(i-scalewin); inbr<=(i+scalewin); inbr+=scale) {
|
||||
if (inbr<0 || inbr>height-1) continue;
|
||||
for (int jnbr=(j-scalewin); jnbr<=(j+scalewin); jnbr+=scale) {
|
||||
if (jnbr<0 || jnbr>width-1) continue;*/
|
||||
dirwt_l = DIRWT_L(inbr, jnbr, i, j);
|
||||
dirwt_ab = DIRWT_AB(inbr, jnbr, i, j);
|
||||
Lout += dirwt_l*data_fine->L[inbr][jnbr];
|
||||
aout += dirwt_l*dirwt_ab*data_fine->a[inbr][jnbr];
|
||||
bout += dirwt_l*dirwt_ab*data_fine->b[inbr][jnbr];
|
||||
norm_l += dirwt_l;
|
||||
norm_ab += dirwt_l*dirwt_ab;
|
||||
}
|
||||
}
|
||||
//lops = Lout/norm;//diagnostic
|
||||
//aops = aout/normab;//diagnostic
|
||||
//bops = bout/normab;//diagnostic
|
||||
|
||||
data_coarse->L[i1][j1]=Lout/norm_l;//low pass filter
|
||||
data_coarse->a[i1][j1]=aout/norm_ab;
|
||||
data_coarse->b[i1][j1]=bout/norm_ab;
|
||||
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
|
||||
|
||||
};
|
||||
|
||||
//%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|
||||
|
||||
void ImProcFunctions::idirpyr(LabImage* data_coarse, LabImage* data_fine, int level, LUTf &rangefn_L, LUTf & nrwt_l, LUTf & nrwt_ab, \
|
||||
int pitch, int scale, const int luma, const int chroma/*, LUTf & Lcurve, LUTf & abcurve*/ )
|
||||
{
|
||||
|
||||
int width = data_fine->W;
|
||||
int height = data_fine->H;
|
||||
|
||||
//array2D<float> nrfactorL (width,height);
|
||||
|
||||
float noisevar_L = 4*SQR(25.0 * luma);
|
||||
float noisevar_ab = 2*SQR(100.0 * chroma);
|
||||
float scalefactor = 1.0/pow(2.0,(level+1)*2);//change the last 2 to 1 for longer tail of higher scale NR
|
||||
|
||||
noisevar_L *= scalefactor;
|
||||
|
||||
// for coarsest level, take non-subsampled lopass image and subtract from lopass_fine to generate hipass image
|
||||
|
||||
// denoise hipass image, add back into lopass_fine to generate denoised image at fine scale
|
||||
|
||||
// now iterate:
|
||||
// (1) take denoised image at level n, expand and smooth using gradient weights from lopass image at level n-1
|
||||
// the result is the smoothed image at level n-1
|
||||
// (2) subtract smoothed image at level n-1 from lopass image at level n-1 to make hipass image at level n-1
|
||||
// (3) denoise the hipass image at level n-1
|
||||
// (4) add the denoised image at level n-1 to the smoothed image at level n-1 to make the denoised image at level n-1
|
||||
|
||||
// note that the coarsest level amounts to skipping step (1) and doing (2,3,4).
|
||||
// in other words, skip step one if pitch=1
|
||||
|
||||
// step (1)
|
||||
|
||||
if (pitch==1) {
|
||||
|
||||
// step (1-2-3-4)
|
||||
|
||||
#ifdef _OPENMP
|
||||
#pragma omp parallel
|
||||
#endif
|
||||
{
|
||||
|
||||
#ifdef _OPENMP
|
||||
#pragma omp for
|
||||
#endif
|
||||
for(int i = 0; i < height; i++)
|
||||
for(int j = 0; j < width; j++) {
|
||||
double wtdsum[3], norm;
|
||||
float hipass[3], hpffluct[3], tonefactor, nrfactora, nrfactorb;
|
||||
|
||||
hipass[1] = data_fine->a[i][j]-data_coarse->a[i][j];
|
||||
hipass[2] = data_fine->b[i][j]-data_coarse->b[i][j];
|
||||
|
||||
//Wiener filter
|
||||
//chroma
|
||||
hpffluct[1]=SQR(hipass[1])+0.001;
|
||||
hpffluct[2]=SQR(hipass[2])+0.001;
|
||||
nrfactora = (hpffluct[1]) /((hpffluct[1]) + noisevar_ab * NRWT_AB);
|
||||
nrfactorb = (hpffluct[2]) /((hpffluct[2]) + noisevar_ab * NRWT_AB);
|
||||
|
||||
hipass[1] *= nrfactora;
|
||||
hipass[2] *= nrfactorb;
|
||||
|
||||
data_fine->a[i][j] = hipass[1]+data_coarse->a[i][j];
|
||||
data_fine->b[i][j] = hipass[2]+data_coarse->b[i][j];
|
||||
}
|
||||
|
||||
|
||||
}//end of pitch=1
|
||||
|
||||
} else {//pitch>1
|
||||
|
||||
LabImage* smooth;
|
||||
|
||||
smooth = new LabImage(width, height);
|
||||
#ifdef _OPENMP
|
||||
#pragma omp parallel
|
||||
#endif
|
||||
|
||||
{
|
||||
|
||||
#ifdef _OPENMP
|
||||
#pragma omp for
|
||||
#endif
|
||||
for(int i = 0; i < height; i+=pitch)
|
||||
{
|
||||
int ix=i/pitch;
|
||||
for(int j = 0, jx=0; j < width; j+=pitch, jx++) {
|
||||
|
||||
//copy common pixels
|
||||
smooth->L[i][j] = data_coarse->L[ix][jx];
|
||||
smooth->a[i][j] = data_coarse->a[ix][jx];
|
||||
smooth->b[i][j] = data_coarse->b[ix][jx];
|
||||
}
|
||||
}
|
||||
//if (pitch>1) {//pitch=2; step (1) expand coarse image, fill in missing data
|
||||
#ifdef _OPENMP
|
||||
#pragma omp for
|
||||
#endif
|
||||
for(int i = 0; i < height-1; i+=2)
|
||||
for(int j = 0; j < width-1; j+=2) {
|
||||
//do midpoint first
|
||||
double norm=0.0,wtdsum[3]={0.0,0.0,0.0};
|
||||
//wtdsum[0]=wtdsum[1]=wtdsum[2]=0.0;
|
||||
for(int ix=i; ix<MIN(height,i+3); ix+=2)
|
||||
for (int jx=j; jx<MIN(width,j+3); jx+=2) {
|
||||
wtdsum[0] += smooth->L[ix][jx];
|
||||
wtdsum[1] += smooth->a[ix][jx];
|
||||
wtdsum[2] += smooth->b[ix][jx];
|
||||
norm++;
|
||||
}
|
||||
norm = 1/norm;
|
||||
smooth->L[i+1][j+1]=wtdsum[0]*norm;
|
||||
smooth->a[i+1][j+1]=wtdsum[1]*norm;
|
||||
smooth->b[i+1][j+1]=wtdsum[2]*norm;
|
||||
}
|
||||
#ifdef _OPENMP
|
||||
#pragma omp for
|
||||
#endif
|
||||
|
||||
for(int i = 0; i < height-1; i+=2)
|
||||
for(int j = 0; j < width-1; j+=2) {
|
||||
//now right neighbor
|
||||
if (j+1==width) continue;
|
||||
double norm=0.0,wtdsum[3]={0.0,0.0,0.0};
|
||||
|
||||
for (int jx=j; jx<MIN(width,j+3); jx+=2) {
|
||||
wtdsum[0] += smooth->L[i][jx];
|
||||
wtdsum[1] += smooth->a[i][jx];
|
||||
wtdsum[2] += smooth->b[i][jx];
|
||||
norm++;
|
||||
}
|
||||
for (int ix=MAX(0,i-1); ix<MIN(height,i+2); ix+=2) {
|
||||
wtdsum[0] += smooth->L[ix][j+1];
|
||||
wtdsum[1] += smooth->a[ix][j+1];
|
||||
wtdsum[2] += smooth->b[ix][j+1];
|
||||
norm++;
|
||||
}
|
||||
norm = 1/norm;
|
||||
smooth->L[i][j+1]=wtdsum[0]*norm;
|
||||
smooth->a[i][j+1]=wtdsum[1]*norm;
|
||||
smooth->b[i][j+1]=wtdsum[2]*norm;
|
||||
|
||||
//now down neighbor
|
||||
if (i+1==height) continue;
|
||||
norm=0.0;wtdsum[0]=wtdsum[1]=wtdsum[2]=0.0;
|
||||
for (int ix=i; ix<MIN(height,i+3); ix+=2) {
|
||||
wtdsum[0] += smooth->L[ix][j];
|
||||
wtdsum[1] += smooth->a[ix][j];
|
||||
wtdsum[2] += smooth->b[ix][j];
|
||||
norm++;
|
||||
}
|
||||
for (int jx=MAX(0,j-1); jx<MIN(width,j+2); jx+=2) {
|
||||
wtdsum[0] += smooth->L[i+1][jx];
|
||||
wtdsum[1] += smooth->a[i+1][jx];
|
||||
wtdsum[2] += smooth->b[i+1][jx];
|
||||
norm++;
|
||||
}
|
||||
norm=1/norm;
|
||||
smooth->L[i+1][j]=wtdsum[0]*norm;
|
||||
smooth->a[i+1][j]=wtdsum[1]*norm;
|
||||
smooth->b[i+1][j]=wtdsum[2]*norm;
|
||||
|
||||
}
|
||||
|
||||
#ifdef _OPENMP
|
||||
#pragma omp for
|
||||
#endif
|
||||
|
||||
// step (2-3-4)
|
||||
for( int i = 0; i < height; i++)
|
||||
for(int j = 0; j < width; j++) {
|
||||
|
||||
float hipass[3], hpffluct[3], nrfactora, nrfactorb;
|
||||
|
||||
hipass[1] = data_fine->a[i][j]-smooth->a[i][j];
|
||||
hipass[2] = data_fine->b[i][j]-smooth->b[i][j];
|
||||
|
||||
//Wiener filter
|
||||
//chroma
|
||||
hpffluct[1]=SQR(hipass[1])+0.001;
|
||||
hpffluct[2]=SQR(hipass[2])+0.001;
|
||||
nrfactora = (hpffluct[1]) /((hpffluct[1]) + noisevar_ab * NRWT_AB /* * abcurve[smooth->L[i][j]]*/);
|
||||
nrfactorb = (hpffluct[2]) /((hpffluct[2]) + noisevar_ab * NRWT_AB /* * abcurve[smooth->L[i][j]]*/);
|
||||
|
||||
hipass[1] *= nrfactora;
|
||||
hipass[2] *= nrfactorb;
|
||||
|
||||
data_fine->a[i][j] = hipass[1]+smooth->a[i][j];
|
||||
data_fine->b[i][j] = hipass[2]+smooth->b[i][j];
|
||||
}
|
||||
|
||||
} // end parallel
|
||||
delete smooth;
|
||||
}//end of pitch>1
|
||||
|
||||
};
|
||||
|
||||
|
||||
#undef DIRWT_L
|
||||
#undef DIRWT_AB
|
||||
|
||||
//#undef NRWT_L
|
||||
#undef NRWT_AB
|
||||
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user