rawTherapee/rtengine/impulse_denoise.cc
2019-11-03 17:03:40 +01:00

558 lines
18 KiB
C++

/*
* 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 widthITheightOUT ANY widthARRANTY; without even the implied warranty of
* MERCheightANTABILITY 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 <https://www.gnu.org/licenses/>.
*
* 2010 Emil Martinec <ejmartin@uchicago.edu>
*
*/
#include <cstddef>
#include "rt_math.h"
#include "labimage.h"
#include "improcfun.h"
#include "cieimage.h"
#include "sleef.h"
#include "opthelper.h"
#include "gauss.h"
using namespace std;
namespace rtengine
{
void ImProcFunctions::impulse_nr (LabImage* lab, double thresh)
{
// %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
// impulse noise removal
// local variables
int width = lab->W;
int height = lab->H;
// buffer for the lowpass image
float * lpf[height] ALIGNED16;
lpf[0] = new float [width * height];
// buffer for the highpass image
char * impish[height] ALIGNED16;
impish[0] = new char [width * height];
for (int i = 1; i < height; i++) {
lpf[i] = lpf[i - 1] + width;
impish[i] = impish[i - 1] + width;
}
//The cleaning algorithm starts here
//%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
// modified bilateral filter for lowpass image, omitting input pixel; or Gaussian blur
const float eps = 1.0;
#ifdef _OPENMP
#pragma omp parallel
#endif
{
gaussianBlur (lab->L, lpf, width, height, max(2.0, thresh - 1.0));
}
//%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
float impthr = max(1.0, 5.5 - thresh);
float impthrDiv24 = impthr / 24.0f; //Issue 1671: moved the Division outside the loop, impthr can be optimized out too, but I let in the code at the moment
#ifdef _OPENMP
#pragma omp parallel
#endif
{
int i1, j1, j;
float hpfabs, hfnbrave;
#ifdef __SSE2__
vfloat hfnbravev, hpfabsv;
vfloat impthrDiv24v = F2V( impthrDiv24 );
#endif
#ifdef _OPENMP
#pragma omp for
#endif
for (int i = 0; i < height; i++) {
for (j = 0; j < 2; j++) {
hpfabs = fabs(lab->L[i][j] - lpf[i][j]);
//block average of high pass data
for (i1 = max(0, i - 2), hfnbrave = 0; i1 <= min(i + 2, height - 1); i1++ )
for (j1 = 0; j1 <= j + 2; j1++) {
hfnbrave += fabs(lab->L[i1][j1] - lpf[i1][j1]);
}
impish[i][j] = (hpfabs > ((hfnbrave - hpfabs) * impthrDiv24));
}
#ifdef __SSE2__
for (; j < width - 5; j += 4) {
hfnbravev = ZEROV;
hpfabsv = vabsf(LVFU(lab->L[i][j]) - LVFU(lpf[i][j]));
//block average of high pass data
for (i1 = max(0, i - 2); i1 <= min(i + 2, height - 1); i1++ ) {
for (j1 = j - 2; j1 <= j + 2; j1++) {
hfnbravev += vabsf(LVFU(lab->L[i1][j1]) - LVFU(lpf[i1][j1]));
}
}
int mask = _mm_movemask_ps((hfnbravev - hpfabsv) * impthrDiv24v - hpfabsv);
impish[i][j] = (mask & 1);
impish[i][j + 1] = ((mask & 2) >> 1);
impish[i][j + 2] = ((mask & 4) >> 2);
impish[i][j + 3] = ((mask & 8) >> 3);
}
#endif
for (; j < width - 2; j++) {
hpfabs = fabs(lab->L[i][j] - lpf[i][j]);
//block average of high pass data
for (i1 = max(0, i - 2), hfnbrave = 0; i1 <= min(i + 2, height - 1); i1++ )
for (j1 = j - 2; j1 <= j + 2; j1++) {
hfnbrave += fabs(lab->L[i1][j1] - lpf[i1][j1]);
}
impish[i][j] = (hpfabs > ((hfnbrave - hpfabs) * impthrDiv24));
}
for (; j < width; j++) {
hpfabs = fabs(lab->L[i][j] - lpf[i][j]);
//block average of high pass data
for (i1 = max(0, i - 2), hfnbrave = 0; i1 <= min(i + 2, height - 1); i1++ )
for (j1 = j - 2; j1 < width; j1++) {
hfnbrave += fabs(lab->L[i1][j1] - lpf[i1][j1]);
}
impish[i][j] = (hpfabs > ((hfnbrave - hpfabs) * impthrDiv24));
}
}
}
//now impulsive values have been identified
// Issue 1671:
// often, noise isn't evenly distributed, e.g. only a few noisy pixels in the bright sky, but many in the dark foreground,
// so it's better to schedule dynamic and let every thread only process 16 rows, to avoid running big threads out of work
// Measured it and in fact gives better performance than without schedule(dynamic,16). Of course, there could be a better
// choice for the chunk_size than 16
// race conditions are avoided by the array impish
#ifdef _OPENMP
#pragma omp parallel
#endif
{
int i1, j1, j;
float wtdsum[3], dirwt, norm;
#ifdef _OPENMP
#pragma omp for schedule(dynamic,16)
#endif
for (int i = 0; i < height; i++) {
for (j = 0; j < 2; j++) {
if (!impish[i][j]) {
continue;
}
norm = 0.0;
wtdsum[0] = wtdsum[1] = wtdsum[2] = 0.0;
for (i1 = max(0, i - 2); i1 <= min(i + 2, height - 1); i1++ )
for (j1 = 0; j1 <= j + 2; j1++ ) {
if (impish[i1][j1]) {
continue;
}
dirwt = 1 / (SQR(lab->L[i1][j1] - lab->L[i][j]) + eps); //use more sophisticated rangefn???
wtdsum[0] += dirwt * lab->L[i1][j1];
wtdsum[1] += dirwt * lab->a[i1][j1];
wtdsum[2] += dirwt * lab->b[i1][j1];
norm += dirwt;
}
if (norm) {
lab->L[i][j] = wtdsum[0] / norm; //low pass filter
lab->a[i][j] = wtdsum[1] / norm; //low pass filter
lab->b[i][j] = wtdsum[2] / norm; //low pass filter
}
}
for (; j < width - 2; j++) {
if (!impish[i][j]) {
continue;
}
norm = 0.0;
wtdsum[0] = wtdsum[1] = wtdsum[2] = 0.0;
for (i1 = max(0, i - 2); i1 <= min(i + 2, height - 1); i1++ )
for (j1 = j - 2; j1 <= j + 2; j1++ ) {
if (impish[i1][j1]) {
continue;
}
dirwt = 1 / (SQR(lab->L[i1][j1] - lab->L[i][j]) + eps); //use more sophisticated rangefn???
wtdsum[0] += dirwt * lab->L[i1][j1];
wtdsum[1] += dirwt * lab->a[i1][j1];
wtdsum[2] += dirwt * lab->b[i1][j1];
norm += dirwt;
}
if (norm) {
lab->L[i][j] = wtdsum[0] / norm; //low pass filter
lab->a[i][j] = wtdsum[1] / norm; //low pass filter
lab->b[i][j] = wtdsum[2] / norm; //low pass filter
}
}
for (; j < width; j++) {
if (!impish[i][j]) {
continue;
}
norm = 0.0;
wtdsum[0] = wtdsum[1] = wtdsum[2] = 0.0;
for (i1 = max(0, i - 2); i1 <= min(i + 2, height - 1); i1++ )
for (j1 = j - 2; j1 < width; j1++ ) {
if (impish[i1][j1]) {
continue;
}
dirwt = 1 / (SQR(lab->L[i1][j1] - lab->L[i][j]) + eps); //use more sophisticated rangefn???
wtdsum[0] += dirwt * lab->L[i1][j1];
wtdsum[1] += dirwt * lab->a[i1][j1];
wtdsum[2] += dirwt * lab->b[i1][j1];
norm += dirwt;
}
if (norm) {
lab->L[i][j] = wtdsum[0] / norm; //low pass filter
lab->a[i][j] = wtdsum[1] / norm; //low pass filter
lab->b[i][j] = wtdsum[2] / norm; //low pass filter
}
}
}
}
//now impulsive values have been corrected
delete [] lpf[0];
delete [] impish[0];
}
void ImProcFunctions::impulse_nrcam (CieImage* ncie, double thresh, float **buffers[3])
{
// %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
// impulse noise removal
// local variables
int width = ncie->W;
int height = ncie->H;
float piid = 3.14159265f / 180.f;
// buffer for the lowpass image
float ** lpf = buffers[0];
// buffer for the highpass image
float ** impish = buffers[1];
//%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
// modified bilateral filter for lowpass image, omitting input pixel; or Gaussian blur
//The cleaning algorithm starts here
//rangeblur<unsigned short, unsigned int> (lab->L, lpf, impish /*used as buffer here*/, width, height, thresh, false);
#ifdef _OPENMP
#pragma omp parallel
#endif
{
gaussianBlur (ncie->sh_p, lpf, width, height, max(2.0, thresh - 1.0));
}
//%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
float impthr = max(1.0f, 5.0f - (float)thresh);
float impthrDiv24 = impthr / 24.0f; //Issue 1671: moved the Division outside the loop, impthr can be optimized out too, but I let in the code at the moment
#ifdef _OPENMP
#pragma omp parallel
#endif
{
int i1, j1, j;
float hpfabs, hfnbrave;
#ifdef __SSE2__
vfloat hfnbravev, hpfabsv;
vfloat impthrDiv24v = F2V( impthrDiv24 );
vfloat onev = F2V( 1.0f );
#endif
#ifdef _OPENMP
#pragma omp for
#endif
for (int i = 0; i < height; i++) {
for (j = 0; j < 2; j++) {
hpfabs = fabs(ncie->sh_p[i][j] - lpf[i][j]);
//block average of high pass data
for (i1 = max(0, i - 2), hfnbrave = 0; i1 <= min(i + 2, height - 1); i1++ )
for (j1 = 0; j1 <= j + 2; j1++) {
hfnbrave += fabs(ncie->sh_p[i1][j1] - lpf[i1][j1]);
}
impish[i][j] = (hpfabs > ((hfnbrave - hpfabs) * impthrDiv24));
}
#ifdef __SSE2__
for (; j < width - 5; j += 4) {
hpfabsv = vabsf(LVFU(ncie->sh_p[i][j]) - LVFU(lpf[i][j]));
hfnbravev = ZEROV;
//block average of high pass data
for (i1 = max(0, i - 2); i1 <= min(i + 2, height - 1); i1++ ) {
for (j1 = j - 2; j1 <= j + 2; j1++ ) {
hfnbravev += vabsf(LVFU(ncie->sh_p[i1][j1]) - LVFU(lpf[i1][j1]));
}
}
STVFU(impish[i][j], vselfzero(vmaskf_gt(hpfabsv, (hfnbravev - hpfabsv)*impthrDiv24v), onev));
}
#endif
for (; j < width - 2; j++) {
hpfabs = fabs(ncie->sh_p[i][j] - lpf[i][j]);
//block average of high pass data
for (i1 = max(0, i - 2), hfnbrave = 0; i1 <= min(i + 2, height - 1); i1++ )
for (j1 = j - 2; j1 <= j + 2; j1++ ) {
hfnbrave += fabs(ncie->sh_p[i1][j1] - lpf[i1][j1]);
}
impish[i][j] = (hpfabs > ((hfnbrave - hpfabs) * impthrDiv24));
}
for (; j < width; j++) {
hpfabs = fabs(ncie->sh_p[i][j] - lpf[i][j]);
//block average of high pass data
for (i1 = max(0, i - 2), hfnbrave = 0; i1 <= min(i + 2, height - 1); i1++ )
for (j1 = j - 2; j1 < width; j1++ ) {
hfnbrave += fabs(ncie->sh_p[i1][j1] - lpf[i1][j1]);
}
impish[i][j] = (hpfabs > ((hfnbrave - hpfabs) * impthrDiv24));
}
}
}
//now impulsive values have been identified
const float eps = 1.0f;
float** sraa = buffers[0]; // we can reuse buffers[0] because lpf is not needed anymore at this point
float** srbb = buffers[2];
#ifdef _OPENMP
#pragma omp parallel
#endif
{
#ifdef __SSE2__
vfloat2 sincosvalv;
vfloat piidv = F2V( piid );
vfloat tempv;
#endif
#ifdef _OPENMP
#pragma omp for
#endif
for (int i = 0; i < height; i++) {
int j = 0;
#ifdef __SSE2__
for (; j < width - 3; j += 4) {
sincosvalv = xsincosf(piidv * LVFU(ncie->h_p[i][j]));
tempv = LVFU(ncie->C_p[i][j]);
STVFU(sraa[i][j], tempv * sincosvalv.y);
STVFU(srbb[i][j], tempv * sincosvalv.x);
}
#endif
for (; j < width; j++) {
float2 sincosval = xsincosf(piid * ncie->h_p[i][j]);
sraa[i][j] = ncie->C_p[i][j] * sincosval.y;
srbb[i][j] = ncie->C_p[i][j] * sincosval.x;
}
}
}
// Issue 1671:
// often, noise isn't evenly distributed, e.g. only a few noisy pixels in the bright sky, but many in the dark foreground,
// so it's better to schedule dynamic and let every thread only process 16 rows, to avoid running big threads out of work
// Measured it and in fact gives better performance than without schedule(dynamic,16). Of course, there could be a better
// choice for the chunk_size than 16
// race conditions are avoided by the array impish
#ifdef _OPENMP
#pragma omp parallel
#endif
{
int i1, j1, j;
float wtdsum[3], dirwt, norm;
#ifdef _OPENMP
#pragma omp for schedule(dynamic,16)
#endif
for (int i = 0; i < height; i++) {
for (j = 0; j < 2; j++) {
if (!impish[i][j]) {
continue;
}
norm = 0.0f;
wtdsum[0] = wtdsum[1] = wtdsum[2] = 0.0f;
for (i1 = max(0, i - 2); i1 <= min(i + 2, height - 1); i1++ )
for (j1 = 0; j1 <= j + 2; j1++ ) {
if (impish[i1][j1]) {
continue;
}
dirwt = 1.f / (SQR(ncie->sh_p[i1][j1] - ncie->sh_p[i][j]) + eps); //use more sophisticated rangefn???
wtdsum[0] += dirwt * ncie->sh_p[i1][j1];
wtdsum[1] += dirwt * sraa[i1][j1];
wtdsum[2] += dirwt * srbb[i1][j1];
norm += dirwt;
}
if (norm) {
ncie->sh_p[i][j] = wtdsum[0] / norm; //low pass filter
sraa[i][j] = wtdsum[1] / norm; //low pass filter
srbb[i][j] = wtdsum[2] / norm; //low pass filter
}
}
for (; j < width - 2; j++) {
if (!impish[i][j]) {
continue;
}
norm = 0.0f;
wtdsum[0] = wtdsum[1] = wtdsum[2] = 0.0f;
for (i1 = max(0, i - 2); i1 <= min(i + 2, height - 1); i1++ )
for (j1 = j - 2; j1 <= j + 2; j1++ ) {
if (impish[i1][j1]) {
continue;
}
dirwt = 1.f / (SQR(ncie->sh_p[i1][j1] - ncie->sh_p[i][j]) + eps); //use more sophisticated rangefn???
wtdsum[0] += dirwt * ncie->sh_p[i1][j1];
wtdsum[1] += dirwt * sraa[i1][j1];
wtdsum[2] += dirwt * srbb[i1][j1];
norm += dirwt;
}
if (norm) {
ncie->sh_p[i][j] = wtdsum[0] / norm; //low pass filter
sraa[i][j] = wtdsum[1] / norm; //low pass filter
srbb[i][j] = wtdsum[2] / norm; //low pass filter
}
}
for (; j < width; j++) {
if (!impish[i][j]) {
continue;
}
norm = 0.0f;
wtdsum[0] = wtdsum[1] = wtdsum[2] = 0.0f;
for (i1 = max(0, i - 2); i1 <= min(i + 2, height - 1); i1++ )
for (j1 = j - 2; j1 < width; j1++ ) {
if (impish[i1][j1]) {
continue;
}
dirwt = 1.f / (SQR(ncie->sh_p[i1][j1] - ncie->sh_p[i][j]) + eps); //use more sophisticated rangefn???
wtdsum[0] += dirwt * ncie->sh_p[i1][j1];
wtdsum[1] += dirwt * sraa[i1][j1];
wtdsum[2] += dirwt * srbb[i1][j1];
norm += dirwt;
}
if (norm) {
ncie->sh_p[i][j] = wtdsum[0] / norm; //low pass filter
sraa[i][j] = wtdsum[1] / norm; //low pass filter
srbb[i][j] = wtdsum[2] / norm; //low pass filter
}
}
}
}
//now impulsive values have been corrected
#ifdef _OPENMP
#pragma omp parallel
#endif
{
#ifdef __SSE2__
vfloat interav, interbv;
vfloat piidv = F2V(piid);
#endif // __SSE2__
#ifdef _OPENMP
#pragma omp for
#endif
for(int i = 0; i < height; i++ ) {
int j = 0;
#ifdef __SSE2__
for(; j < width - 3; j += 4) {
interav = LVFU(sraa[i][j]);
interbv = LVFU(srbb[i][j]);
STVFU(ncie->h_p[i][j], (xatan2f(interbv, interav)) / piidv);
STVFU(ncie->C_p[i][j], vsqrtf(SQRV(interbv) + SQRV(interav)));
}
#endif
for(; j < width; j++) {
float intera = sraa[i][j];
float interb = srbb[i][j];
ncie->h_p[i][j] = (xatan2f(interb, intera)) / piid;
ncie->C_p[i][j] = sqrt(SQR(interb) + SQR(intera));
}
}
}
}
}