ahd demosaic, reduced processing time and memory usage, #4698

This commit is contained in:
heckflosse 2018-08-01 18:48:08 +02:00
parent 6c8a47ebdf
commit f3ecd14481

View File

@ -2332,209 +2332,201 @@ void RawImageSource::igv_interpolate(int winw, int winh)
/*
Adaptive Homogeneity-Directed interpolation is based on
the work of Keigo Hirakawa, Thomas Parks, and Paul Lee.
Optimized for speed and reduced memory usage 2018 Ingo Weyrich
*/
#define TS 256 /* Tile Size */
#define FORC(cnt) for (c=0; c < cnt; c++)
#define FORC3 FORC(3)
#define TS 144
void RawImageSource::ahd_demosaic()
{
BENCHFUN
int i, j, k, tr, tc, c, d, hm[2];
float val;
float (*pix)[3], (*rix)[3];
static const int dir[4] = { -1, 1, -TS, TS };
float ldiff[2][4], abdiff[2][4], leps, abeps;
float xyz[3], xyz_cam[3][3];
constexpr int dir[4] = { -1, 1, -TS, TS };
float xyz_cam[3][3];
LUTf cbrt(65536);
float (*rgb)[TS][TS][3];
float (*lab)[TS][TS][3];
float (*lix)[3];
char (*homo)[TS][TS], *buffer;
double r;
int width = W, height = H;
unsigned int colors = 3;
const double xyz_rgb[3][3] = { /* XYZ from RGB */
constexpr double xyz_rgb[3][3] = { /* XYZ from RGB */
{ 0.412453, 0.357580, 0.180423 },
{ 0.212671, 0.715160, 0.072169 },
{ 0.019334, 0.119193, 0.950227 }
};
const float d65_white[3] = { 0.950456, 1, 1.088754 };
constexpr float d65_white[3] = { 0.950456, 1, 1.088754 };
volatile double progress = 0.0;
if (plistener) {
plistener->setProgressStr (Glib::ustring::compose(M("TP_RAW_DMETHOD_PROGRESSBAR"), RAWParams::BayerSensor::getMethodString(RAWParams::BayerSensor::Method::AHD)));
plistener->setProgress (0.0);
}
float (*image)[3] = (float (*)[3]) calloc (H * W, sizeof * image);
for (int ii = 0; ii < H; ii++)
for (int jj = 0; jj < W; jj++) {
image[ii * W + jj][fc(ii, jj)] = rawData[ii][jj];
}
for (i = 0; i < 0x10000; i++) {
r = (double)i / 65535.0;
for (int i = 0; i < 0x10000; i++) {
double r = (double)i / 65535.0;
cbrt[i] = r > 0.008856 ? std::cbrt(r) : 7.787 * r + 16 / 116.0;
}
for (i = 0; i < 3; i++)
for (unsigned int j = 0; j < colors; j++)
for (xyz_cam[i][j] = k = 0; k < 3; k++) {
for (int i = 0; i < 3; i++)
for (unsigned int j = 0; j < 3; j++) {
xyz_cam[i][j] = 0;
for (int k = 0; k < 3; k++) {
xyz_cam[i][j] += xyz_rgb[i][k] * imatrices.rgb_cam[k][j] / d65_white[i];
}
}
border_interpolate(5, image);
buffer = (char *) malloc (13 * TS * TS * sizeof(float)); /* 1664 kB */
//merror (buffer, "ahd_interpolate()");
rgb = (float(*)[TS][TS][3]) buffer;
lab = (float(*)[TS][TS][3])(buffer + 6 * TS * TS * sizeof(float));
homo = (char (*)[TS][TS]) (buffer + 12 * TS * TS * sizeof(float));
border_interpolate2(W, H, 5, rawData, red, green, blue);
// helper variables for progress indication
int n_tiles = ((height - 7 + (TS - 7)) / (TS - 6)) * ((width - 7 + (TS - 7)) / (TS - 6));
int tile = 0;
#ifdef _OPENMP
#pragma omp parallel
#endif
{
int progresscounter = 0;
float (*pix), (*rix)[3];
float ldiff[2][4], abdiff[2][4];
float (*lix)[3];
char *buffer = (char *) malloc (12 * TS * TS * sizeof(float) + 2 * TS * TS * sizeof(uint16_t)); /* 1053 kB per core */
float (*rgb)[TS][TS][3] = (float(*)[TS][TS][3]) buffer;
float (*lab)[TS][TS][3] = (float(*)[TS][TS][3])(buffer + 6 * TS * TS * sizeof(float));
uint16_t (*homo)[TS][TS] = (uint16_t(*)[TS][TS])(buffer + 12 * TS * TS * sizeof(float));
for (int top = 2; top < height - 5; top += TS - 6)
#ifdef _OPENMP
#pragma omp for collapse(2) schedule(dynamic) nowait
#endif
for (int top = 2; top < height - 5; top += TS - 6) {
for (int left = 2; left < width - 5; left += TS - 6) {
/* Interpolate green horizontally and vertically: */
// Interpolate green horizontally and vertically:
for (int row = top; row < top + TS && row < height - 2; row++) {
int col = left + (FC(row, left) & 1);
for (c = FC(row, col); col < left + TS && col < width - 2; col += 2) {
pix = image + (row * width + col);
val = 0.25f * ((pix[-1][1] + pix[0][c] + pix[1][1]) * 2
- pix[-2][c] - pix[2][c]) ;
rgb[0][row - top][col - left][1] = median(val, pix[-1][1], pix[1][1]);
val = 0.25f * ((pix[-width][1] + pix[0][c] + pix[width][1]) * 2
- pix[-2 * width][c] - pix[2 * width][c]) ;
rgb[1][row - top][col - left][1] = median(val, pix[-width][1], pix[width][1]);
for (int col = left + (FC(row, left) & 1); col < std::min(left + TS, width - 2); col += 2) {
pix = &(rawData[row][col]);
float val0 = 0.25f * ((pix[-1] + pix[0] + pix[1]) * 2
- pix[-2] - pix[2]) ;
rgb[0][row - top][col - left][1] = median(val0, pix[-1], pix[1]);
float val1 = 0.25f * ((pix[-width] + pix[0] + pix[width]) * 2
- pix[-2 * width] - pix[2 * width]) ;
rgb[1][row - top][col - left][1] = median(val1, pix[-width], pix[width]);
}
}
/* Interpolate red and blue, and convert to CIELab: */
for (d = 0; d < 2; d++)
for (int row = top + 1; row < top + TS - 1 && row < height - 3; row++)
for (int col = left + 1; col < left + TS - 1 && col < width - 3; col++) {
pix = image + (row * width + col);
// Interpolate red and blue, and convert to CIELab:
for (int d = 0; d < 2; d++)
for (int row = top + 1; row < top + TS - 1 && row < height - 3; row++) {
int cng = FC(row + 1, FC(row + 1, 0) & 1);
for (int col = left + 1; col < std::min(left + TS - 1, width - 3); col++) {
pix = &(rawData[row][col]);
rix = &rgb[d][row - top][col - left];
lix = &lab[d][row - top][col - left];
if ((c = 2 - FC(row, col)) == 1) {
c = FC(row + 1, col);
val = pix[0][1] + (0.5f * ( pix[-1][2 - c] + pix[1][2 - c]
- rix[-1][1] - rix[1][1] ) );
rix[0][2 - c] = CLIP(val);
val = pix[0][1] + (0.5f * ( pix[-width][c] + pix[width][c]
- rix[-TS][1] - rix[TS][1] ) );
} else
val = rix[0][1] + (0.25f * ( pix[-width - 1][c] + pix[-width + 1][c]
+ pix[+width - 1][c] + pix[+width + 1][c]
if (FC(row, col) == 1) {
rix[0][2 - cng] = CLIP(pix[0] + (0.5f * (pix[-1] + pix[1]
- rix[-1][1] - rix[1][1] ) ));
rix[0][cng] = CLIP(pix[0] + (0.5f * (pix[-width] + pix[width]
- rix[-TS][1] - rix[TS][1])));
rix[0][1] = pix[0];
} else {
rix[0][cng] = CLIP(rix[0][1] + (0.25f * (pix[-width - 1] + pix[-width + 1]
+ pix[+width - 1] + pix[+width + 1]
- rix[-TS - 1][1] - rix[-TS + 1][1]
- rix[+TS - 1][1] - rix[+TS + 1][1]) );
rix[0][c] = CLIP(val);
c = FC(row, col);
rix[0][c] = pix[0][c];
xyz[0] = xyz[1] = xyz[2] = 0.f;
FORCC {
xyz[0] += xyz_cam[0][c] * rix[0][c];
xyz[1] += xyz_cam[1][c] * rix[0][c];
xyz[2] += xyz_cam[2][c] * rix[0][c];
- rix[+TS - 1][1] - rix[+TS + 1][1])));
rix[0][2 - cng] = pix[0];
}
float xyz0 = 0.f;
float xyz1 = 0.f;
float xyz2 = 0.f;
for(unsigned int c = 0; c < 3; ++c) {
xyz0 += xyz_cam[0][c] * rix[0][c];
xyz1 += xyz_cam[1][c] * rix[0][c];
xyz2 += xyz_cam[2][c] * rix[0][c];
}
xyz[0] = cbrt[xyz[0]];
xyz[1] = cbrt[xyz[1]];
xyz[2] = cbrt[xyz[2]];
xyz0 = cbrt[xyz0];
xyz1 = cbrt[xyz1];
xyz2 = cbrt[xyz2];
//xyz[0] = xyz[0] > 0.008856 ? pow(xyz[0]/65535,1/3.0) : 7.787*xyz[0] + 16/116.0;
//xyz[1] = xyz[1] > 0.008856 ? pow(xyz[1]/65535,1/3.0) : 7.787*xyz[1] + 16/116.0;
//xyz[2] = xyz[2] > 0.008856 ? pow(xyz[2]/65535,1/3.0) : 7.787*xyz[2] + 16/116.0;
lix[0][0] = (116 * xyz[1] - 16);
lix[0][1] = 500 * (xyz[0] - xyz[1]);
lix[0][2] = 200 * (xyz[1] - xyz[2]);
lix[0][0] = 116.f * xyz1 - 16.f;
lix[0][1] = 500.f * (xyz0 - xyz1);
lix[0][2] = 200.f * (xyz1 - xyz2);
}
}
/* Build homogeneity maps from the CIELab images: */
memset (homo, 0, 2 * TS * TS);
// Build homogeneity maps from the CIELab images:
for (int row = top + 2; row < top + TS - 2 && row < height - 4; row++) {
tr = row - top;
int tr = row - top;
for (int col = left + 2; col < left + TS - 2 && col < width - 4; col++) {
tc = col - left;
for (d = 0; d < 2; d++) {
for (int col = left + 2, tc = 2; col < left + TS - 2 && col < width - 4; col++, tc++) {
for (int d = 0; d < 2; d++) {
lix = &lab[d][tr][tc];
for (i = 0; i < 4; i++) {
for (int i = 0; i < 4; i++) {
ldiff[d][i] = std::fabs(lix[0][0] - lix[dir[i]][0]);
abdiff[d][i] = SQR(lix[0][1] - lix[dir[i]][1])
+ SQR(lix[0][2] - lix[dir[i]][2]);
}
}
leps = min(max(ldiff[0][0], ldiff[0][1]),
max(ldiff[1][2], ldiff[1][3]));
abeps = min(max(abdiff[0][0], abdiff[0][1]),
max(abdiff[1][2], abdiff[1][3]));
float leps = min(max(ldiff[0][0], ldiff[0][1]),
max(ldiff[1][2], ldiff[1][3]));
float abeps = min(max(abdiff[0][0], abdiff[0][1]),
max(abdiff[1][2], abdiff[1][3]));
for (d = 0; d < 2; d++)
for (i = 0; i < 4; i++)
for (int d = 0; d < 2; d++) {
homo[d][tr][tc] = 0;
for (int i = 0; i < 4; i++) {
if (ldiff[d][i] <= leps && abdiff[d][i] <= abeps) {
homo[d][tr][tc]++;
}
}
}
}
}
/* Combine the most homogenous pixels for the final result: */
// Combine the most homogenous pixels for the final result:
for (int row = top + 3; row < top + TS - 3 && row < height - 5; row++) {
tr = row - top;
int tr = row - top;
for (int col = left + 3; col < left + TS - 3 && col < width - 5; col++) {
tc = col - left;
for (int col = left + 3, tc = 3; col < std::min(left + TS - 3, width - 5); col++, tc++) {
uint16_t hm0 = 0, hm1 = 0;
for (int i = tr - 1; i <= tr + 1; i++)
for (int j = tc - 1; j <= tc + 1; j++) {
hm0 += homo[0][i][j];
hm1 += homo[1][i][j];
}
for (d = 0; d < 2; d++)
for (hm[d] = 0, i = tr - 1; i <= tr + 1; i++)
for (j = tc - 1; j <= tc + 1; j++) {
hm[d] += homo[d][i][j];
}
if (hm[0] != hm[1]) {
FORC3 image[row * width + col][c] = rgb[hm[1] > hm[0]][tr][tc][c];
} else
FORC3 image[row * width + col][c] =
0.5f * (rgb[0][tr][tc][c] + rgb[1][tr][tc][c]) ;
if (hm0 != hm1) {
int dir = hm1 > hm0;
red[row][col] = rgb[dir][tr][tc][0];
green[row][col] = rgb[dir][tr][tc][1];
blue[row][col] = rgb[dir][tr][tc][2];
} else {
red[row][col] = 0.5f * (rgb[0][tr][tc][0] + rgb[1][tr][tc][0]);
green[row][col] = 0.5f * (rgb[0][tr][tc][1] + rgb[1][tr][tc][1]);
blue[row][col] = 0.5f * (rgb[0][tr][tc][2] + rgb[1][tr][tc][2]);
}
}
}
tile++;
if(plistener) {
plistener->setProgress((double)tile / n_tiles);
}
}
progresscounter++;
if(progresscounter % 32 == 0) {
#ifdef _OPENMP
#pragma omp critical (ahdprogress)
#endif
{
progress += (double)32 * ((TS - 32) * (TS - 32)) / (height * width);
progress = progress > 1.0 ? 1.0 : progress;
plistener->setProgress(progress);
}
}
}
}
}
free (buffer);
}
if(plistener) {
plistener->setProgress (1.0);
}
free (buffer);
for (int i = 0; i < H; i++) {
for (int j = 0; j < W; j++) {
red[i][j] = image[i * W + j][0];
green[i][j] = image[i * W + j][1];
blue[i][j] = image[i * W + j][2];
}
}
free (image);
}
#undef TS