rawTherapee/rtengine/ipresize.cc

438 lines
15 KiB
C++

/*
* This file is part of RawTherapee.
*
* Copyright (c) 2004-2010 Gabor Horvath <hgabor@rawtherapee.com>
*
* 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/>.
*/
#include <rtengine.h>
#include <improcfun.h>
#include <glibmm.h>
#ifdef _OPENMP
#include <omp.h>
#endif
#include <iostream>
namespace rtengine {
#undef CLIP
#undef CLIPTO
#undef CMAXVAL
#define CMAXVAL 0xffff
#define CLIP(a) ((a)>0?((a)<CMAXVAL?(a):CMAXVAL):0)
#define CLIPTO(a,b,c) ((a)>(b)?((a)<(c)?(a):(c)):(b))
inline double Lanc(double x, double a)
{
if (x * x < 1e-6)
return 1.0;
else if (x * x > a * a)
return 0.0;
else {
x = M_PI * x;
return sin(x) * sin(x / a) / (x * x / a);
}
}
void Lanczos(const Image16* src, Image16* dst, double scale)
{
const double delta = 1.0 / scale;
const double a = 3.0;
const double sc = std::min(scale, 1.0);
const int support = (int)(2.0 * a / sc) + 1;
// storage for precomputed parameters for horisontal interpolation
double * wwh = new double[support * dst->width];
int * jj0 = new int[dst->width];
int * jj1 = new int[dst->width];
// temporal storage for vertically-interpolated row of pixels
double * lr = new double[src->width];
double * lg = new double[src->width];
double * lb = new double[src->width];
// Phase 1: precompute coefficients for horisontal interpolation
for (int j = 0; j < dst->width; j++) {
// x coord of the center of pixel on src image
double x0 = (j + 0.5) * delta - 0.5;
// weights for interpolation in horisontal direction
double * w = wwh + j * support;
// sum of weights used for normalization
double ws = 0.0;
jj0[j] = std::max(0, (int)floor(x0 - a / sc) + 1);
jj1[j] = std::min(src->width, (int)floor(x0 + a / sc) + 1);
// calculate weights
for (int jj = jj0[j]; jj < jj1[j]; jj++) {
int k = jj - jj0[j];
double z = sc * (x0 - jj);
w[k] = Lanc(z, a);
ws += w[k];
}
// normalize weights
for (int k = 0; k < support; k++) {
w[k] /= ws;
}
}
// Phase 2: do actual interpolation
for (int i = 0; i < dst->height; i++) {
// y coord of the center of pixel on src image
double y0 = (i + 0.5) * delta - 0.5;
// weights for interpolation in y direction
double w[support];
// sum of weights used for normalization
double ws= 0.0;
int ii0 = std::max(0, (int)floor(y0 - a / sc) + 1);
int ii1 = std::min(src->height, (int)floor(y0 + a / sc) + 1);
// calculate weights for vertical interpolation
for (int ii = ii0; ii < ii1; ii++) {
int k = ii - ii0;
double z = sc * (y0 - ii);
w[k] = Lanc(z, a);
ws += w[k];
}
// normalize weights
for (int k = 0; k < support; k++) {
w[k] /= ws;
}
// Do vertical interpolation. Store results.
for (int j = 0; j < src->width; j++) {
double r = 0.0, g = 0.0, b = 0.0;
for (int ii = ii0; ii < ii1; ii++) {
int k = ii - ii0;
r += w[k] * src->r[ii][j];
g += w[k] * src->g[ii][j];
b += w[k] * src->b[ii][j];
}
lr[j] = r;
lg[j] = g;
lb[j] = b;
}
// Do horisontal interpolation
for(int j = 0; j < dst->width; j++) {
double * wh = wwh + support * j;
double r = 0.0, g = 0.0, b = 0.0;
for (int jj = jj0[j]; jj < jj1[j]; jj++) {
int k = jj - jj0[j];
r += wh[k] * lr[jj];
g += wh[k] * lg[jj];
b += wh[k] * lb[jj];
}
dst->r[i][j] = CLIP((int)r);
dst->g[i][j] = CLIP((int)g);
dst->b[i][j] = CLIP((int)b);
}
}
delete[] wwh;
delete[] jj0;
delete[] jj1;
delete[] lr;
delete[] lg;
delete[] lb;
}
void ImProcFunctions::resize (Image16* src, Image16* dst) {
//time_t t1 = clock();
if(params->resize.method == "Lanczos") {
Lanczos(src, dst, params->resize.scale);
}
else if(params->resize.method == "Downscale (Better)") {
// small-scale algorithm by Ilia
// provides much better quality on small scales
// calculates mean value over source pixels which current destination pixel covers
// works only for scales < 1
// for scales ~1 it is analogous to bilinear
// possibly, for even less scale factors (< 0.2 possibly) boundary pixels are not needed, omitting them can give a speedup
// this algorithm is much slower on small factors than others, because it uses all pixels of the SOURCE image
// Ilia Popov ilia_popov@rambler.ru 2010
double delta = 1.0 / params->resize.scale;
double k = params->resize.scale * params->resize.scale;
#pragma omp parallel for if (multiThread)
for(int i = 0; i < dst->height; i++) {
// top and bottom boundary coordinates
double y0 = i * delta;
double y1 = (i + 1) * delta;
int m0 = y0;
m0 = CLIPTO(m0, 0, src->height-1);
int m1 = y1;
m1 = CLIPTO(m1, 0, src->height-1);
// weights of boundary pixels
double wy0 = 1.0 - (y0 - m0);
double wy1 = y1 - m1;
for(int j = 0; j < dst->width; j++) {
// left and right boundary coordinates
double x0 = j * delta;
double x1 = (j + 1) * delta;
int n0 = x0;
n0 = CLIPTO(n0, 0, src->width-1);
int n1 = x1;
n1 = CLIPTO(n1, 0, src->width-1);
double wx0 = 1.0 - (x0 - n0);
double wx1 = x1 - n1;
double r = 0;
double g = 0;
double b = 0;
// integration
// corners
r += wy0 * wx0 * src->r[m0][n0] + wy0 * wx1 * src->r[m0][n1] + wy1 * wx0 * src->r[m1][n0] + wy1 * wx1 * src->r[m1][n1];
g += wy0 * wx0 * src->g[m0][n0] + wy0 * wx1 * src->g[m0][n1] + wy1 * wx0 * src->g[m1][n0] + wy1 * wx1 * src->g[m1][n1];
b += wy0 * wx0 * src->b[m0][n0] + wy0 * wx1 * src->b[m0][n1] + wy1 * wx0 * src->b[m1][n0] + wy1 * wx1 * src->b[m1][n1];
// top and bottom boundaries
for(int n = n0 + 1; n < n1; n++) {
r += wy0 * src->r[m0][n] + wy1 * src->r[m1][n];
g += wy0 * src->g[m0][n] + wy1 * src->g[m1][n];
b += wy0 * src->b[m0][n] + wy1 * src->b[m1][n];
}
// inner rows
for(int m = m0 + 1; m < m1; m++) {
// left and right boundaries
r += wx0 * src->r[m][n0] + wx1 * src->r[m][n1];
g += wx0 * src->g[m][n0] + wx1 * src->g[m][n1];
b += wx0 * src->b[m][n0] + wx1 * src->b[m][n1];
// inner pixels
for(int n = n0 + 1; n < n1; n++) {
r += src->r[m][n];
g += src->g[m][n];
b += src->b[m][n];
}
}
// overall weight is equal to the DST pixel area in SRC coordinates
r *= k;
g *= k;
b *= k;
dst->r[i][j] = CLIP((int)r);
dst->g[i][j] = CLIP((int)g);
dst->b[i][j] = CLIP((int)b);
}
}
}
else if(params->resize.method == "Downscale (Faster)") {
// faster version of algo above, does not take into account border pixels,
// which are summed with non-unity weights in slow algo. So, no need
// for weights at all
// Ilia Popov ilia_popov@rambler.ru 5.04.2010
double delta = 1.0 / params->resize.scale;
int p = (int) delta;
// if actually we are doing upscaling, behave like Nearest
if(p == 0)
p = 1;
int q = p/2;
// may cause problems on 32-bit systems on extremely small factors.
// In that case change 1024 to smth less
const int divider = 1024;
// scaling factor after summation
int k = divider / (p * p);
#pragma omp parallel for if (multiThread)
for(int i = 0; i < dst->height; i++) {
// y coordinate of center of destination pixel
double y = (i + 0.5) * delta;
int m0 = (int) (y) - q;
m0 = CLIPTO(m0, 0, src->height-1);
int m1 = m0 + p;
if(m1 > src->height) {
m1 = src->height;
m0 = m1 - p;
}
m1 = CLIPTO(m1, 0, src->height);
for(int j = 0; j < dst->width; j++) {
// x coordinate of center of destination pixel
double x = (j + 0.5) * delta;
int n0 = (int) (x) - q;
n0 = CLIPTO(n0, 0, src->width-1);
int n1 = n0 + p;
if(n1 > src->width) {
n1 = src->width;
n0 = n1 - p;
}
n1 = CLIPTO(n1, 0, src->width);
int r = 0;
int g = 0;
int b = 0;
// integration
for(int m = m0; m < m1; m++) {
for(int n = n0; n < n1; n++) {
r += src->r[m][n];
g += src->g[m][n];
b += src->b[m][n];
}
}
dst->r[i][j] = CLIP( r * k / divider);
dst->g[i][j] = CLIP( g * k / divider);
dst->b[i][j] = CLIP( b * k / divider);
}
}
}
else if (params->resize.method.substr(0,7)=="Bicubic") {
double Av = -0.5;
if (params->resize.method=="Bicubic (Sharper)")
Av = -0.75;
else if (params->resize.method=="Bicubic (Softer)")
Av = -0.25;
#pragma omp parallel for if (multiThread)
for (int i=0; i<dst->height; i++) {
double wx[4], wy[4];
double Dy = i / params->resize.scale;
int yc = (int) Dy; Dy -= (double)yc;
int ys = yc - 1; // smallest y-index used for interpolation
// compute vertical weights
double t1y = -Av*(Dy-1.0)*Dy;
double t2y = (3.0-2.0*Dy)*Dy*Dy;
wy[3] = t1y*Dy;
wy[2] = t1y*(Dy-1.0) + t2y;
wy[1] = -t1y*Dy + 1.0 - t2y;
wy[0] = -t1y*(Dy-1.0);
for (int j=0; j<dst->width; j++) {
double Dx = j / params->resize.scale;
int xc = (int) Dx; Dx -= (double)xc;
int xs = xc - 1; // smallest x-index used for interpolation
if (ys >= 0 && ys <src->height-3 && xs >= 0 && xs <= src->width-3) {
// compute horizontal weights
double t1 = -Av*(Dx-1.0)*Dx;
double t2 = (3.0-2.0*Dx)*Dx*Dx;
wx[3] = t1*Dx;
wx[2] = t1*(Dx-1.0) + t2;
wx[1] = -t1*Dx + 1.0 - t2;
wx[0] = -t1*(Dx-1.0);
// compute weighted sum
int r = 0;
int g = 0;
int b = 0;
for (int x=0; x<4; x++)
for (int y=0; y<4; y++) {
double w = wx[x]*wy[y];
r += w*src->r[ys+y][xs+x];
g += w*src->g[ys+y][xs+x];
b += w*src->b[ys+y][xs+x];
}
dst->r[i][j] = CLIP(r);
dst->g[i][j] = CLIP(g);
dst->b[i][j] = CLIP(b);
}
else {
xc = CLIPTO(xc, 0, src->width-1);
yc = CLIPTO(yc, 0, src->height-1);
int nx = xc + 1;
if (nx>=src->width)
nx = xc;
int ny = yc + 1;
if (ny>=src->height)
ny = yc;
dst->r[i][j] = (1-Dx)*(1-Dy)*src->r[yc][xc] + (1-Dx)*Dy*src->r[ny][xc] + Dx*(1-Dy)*src->r[yc][nx] + Dx*Dy*src->r[ny][nx];
dst->g[i][j] = (1-Dx)*(1-Dy)*src->g[yc][xc] + (1-Dx)*Dy*src->g[ny][xc] + Dx*(1-Dy)*src->g[yc][nx] + Dx*Dy*src->g[ny][nx];
dst->b[i][j] = (1-Dx)*(1-Dy)*src->b[yc][xc] + (1-Dx)*Dy*src->b[ny][xc] + Dx*(1-Dy)*src->b[yc][nx] + Dx*Dy*src->b[ny][nx];
}
}
}
}
else if (params->resize.method=="Bilinear") {
#pragma omp parallel for if (multiThread)
for (int i=0; i<dst->height; i++) {
int sy = i/params->resize.scale;
sy = CLIPTO(sy, 0, src->height-1);
double dy = i/params->resize.scale - sy;
int ny = sy+1;
if (ny>=src->height)
ny = sy;
for (int j=0; j<dst->width; j++) {
int sx = j/params->resize.scale;
sx = CLIPTO(sx, 0, src->width-1);
double dx = j/params->resize.scale - sx;
int nx = sx+1;
if (nx>=src->width)
nx = sx;
dst->r[i][j] = (1-dx)*(1-dy)*src->r[sy][sx] + (1-dx)*dy*src->r[ny][sx] + dx*(1-dy)*src->r[sy][nx] + dx*dy*src->r[ny][nx];
dst->g[i][j] = (1-dx)*(1-dy)*src->g[sy][sx] + (1-dx)*dy*src->g[ny][sx] + dx*(1-dy)*src->g[sy][nx] + dx*dy*src->g[ny][nx];
dst->b[i][j] = (1-dx)*(1-dy)*src->b[sy][sx] + (1-dx)*dy*src->b[ny][sx] + dx*(1-dy)*src->b[sy][nx] + dx*dy*src->b[ny][nx];
}
}
}
else {
#pragma omp parallel for if (multiThread)
for (int i=0; i<dst->height; i++) {
int sy = i/params->resize.scale;
sy = CLIPTO(sy, 0, src->height-1);
for (int j=0; j<dst->width; j++) {
int sx = j/params->resize.scale;
sx = CLIPTO(sx, 0, src->width-1);
dst->r[i][j] = src->r[sy][sx];
dst->g[i][j] = src->g[sy][sx];
dst->b[i][j] = src->b[sy][sx];
}
}
}
//time_t t2 = clock();
//std::cout << "Resize: " << params->resize.method << ": "
// << (double)(t2 - t1) / CLOCKS_PER_SEC << std::endl;
}
}