liz/rawTherapee
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Dehaze: further speedup, #5456

Browse Source
This commit is contained in:
Ingo Weyrich
2019-09-19 20:56:33 +02:00
parent 991fc94d89
commit 3ab379ad0a
3 changed files with 442 additions and 260 deletions
Download Patch File Download Diff File Expand all files Collapse all files

260
rtengine/ipdehaze.cc
View File

@@ -35,7 +35,10 @@
#include "improcfun.h"
#include "procparams.h"
#include "rt_algo.h"
#include "rt_algo.h"
#include "rt_math.h"
#define BENCHMARK
#include "StopWatch.h"
extern Options options;
@@ -43,24 +46,7 @@ namespace rtengine {
namespace {
#if 0
# define DEBUG_DUMP(arr) \
do { \
Imagefloat im(arr.width(), arr.height()); \
const char *out = "/tmp/" #arr ".tif"; \
for (int y = 0; y < im.getHeight(); ++y) { \
for (int x = 0; x < im.getWidth(); ++x) { \
im.r(y, x) = im.g(y, x) = im.b(y, x) = arr[y][x] * 65535.f; \
} \
} \
im.saveTIFF(out, 16); \
} while (false)
#else
# define DEBUG_DUMP(arr)
#endif
int get_dark_channel(const array2D<float> &R, const array2D<float> &G, const array2D<float> &B, array2D<float> &dst, int patchsize, const float ambient[3], bool clip, bool multithread)
int get_dark_channel(const array2D<float> &R, const array2D<float> &G, const array2D<float> &B, array2D<float> &dst, int patchsize, const float ambient[3], bool clip, bool multithread, float strength)
{
const int W = R.width();
const int H = R.height();
@@ -73,22 +59,12 @@ int get_dark_channel(const array2D<float> &R, const array2D<float> &G, const arr
for (int x = 0; x < W; x += patchsize) {
float val = RT_INFINITY_F;
const int pW = min(x + patchsize, W);
for (int yy = y; yy < pH; ++yy) {
for (int xx = x; xx < pW; ++xx) {
float r = R[yy][xx];
float g = G[yy][xx];
float b = B[yy][xx];
if (ambient) {
r /= ambient[0];
g /= ambient[1];
b /= ambient[2];
}
val = min(val, r, g, b);
for (int xx = x; xx < pW; ++xx) {
for (int yy = y; yy < pH; ++yy) {
val = min(val, R[yy][xx] / ambient[0], G[yy][xx] / ambient[1], B[yy][xx] / ambient[2]);
}
}
if (clip) {
val = LIM01(val);
}
val = 1.f - strength * LIM01(val);
for (int yy = y; yy < pH; ++yy) {
std::fill(dst[yy] + x, dst[yy] + pW, val);
}
@@ -98,41 +74,59 @@ int get_dark_channel(const array2D<float> &R, const array2D<float> &G, const arr
return (W / patchsize + ((W % patchsize) > 0)) * (H / patchsize + ((H % patchsize) > 0));
}
int get_dark_channel_downsized(const array2D<float> &R, const array2D<float> &G, const array2D<float> &B, array2D<float> &dst, int patchsize, bool multithread)
{
const int W = R.width();
const int H = R.height();
#ifdef _OPENMP
#pragma omp parallel for if (multithread)
#endif
for (int y = 0; y < H; y += patchsize) {
int yy = y / patchsize;
const int pH = min(y + patchsize, H);
for (int x = 0, xx = 0; x < W; x += patchsize, ++xx) {
float val = RT_INFINITY_F;
const int pW = min(x + patchsize, W);
for (int xp = x; xp < pW; ++xp) {
for (int yp = y; yp < pH; ++yp) {
val = min(val, R[yp][xp], G[yp][xp], B[yp][xp]);
}
}
dst[yy][xx] = val;
}
}
return (W / patchsize + ((W % patchsize) > 0)) * (H / patchsize + ((H % patchsize) > 0));
}
float estimate_ambient_light(const array2D<float> &R, const array2D<float> &G, const array2D<float> &B, const array2D<float> &dark, int patchsize, int npatches, float ambient[3])
{
const int W = R.width();
const int H = R.height();
const auto get_percentile =
[](std::priority_queue<float> &q, float prcnt) -> float
{
size_t n = LIM<size_t>(q.size() * prcnt, 1, q.size());
while (q.size() > n) {
q.pop();
}
return q.top();
};
float darklim = RT_INFINITY_F;
{
std::priority_queue<float> p;
for (int y = 0; y < H; y += patchsize) {
for (int x = 0; x < W; x += patchsize) {
if (!OOG(dark[y][x], 1.f - 1e-5f)) {
p.push(dark[y][x]);
std::vector<float> p;
for (int y = 0, yy = 0; y < H; y += patchsize, ++yy) {
for (int x = 0, xx = 0; x < W; x += patchsize, ++xx) {
if (!OOG(dark[yy][xx], 1.f - 1e-5f)) {
p.push_back(dark[yy][xx]);
}
}
}
darklim = get_percentile(p, 0.95);
const int pos = p.size() * 0.95;
std::nth_element(p.begin(), p.begin() + pos, p.end());
darklim = p[pos];
}
std::vector<std::pair<int, int>> patches;
patches.reserve(npatches);
for (int y = 0; y < H; y += patchsize) {
for (int x = 0; x < W; x += patchsize) {
if (dark[y][x] >= darklim && !OOG(dark[y][x], 1.f)) {
for (int y = 0, yy = 0; y < H; y += patchsize, ++yy) {
for (int x = 0, xx = 0; x < W; x += patchsize, ++xx) {
if (dark[yy][xx] >= darklim && !OOG(dark[yy][xx], 1.f)) {
patches.push_back(std::make_pair(x, y));
}
}
@@ -145,33 +139,38 @@ float estimate_ambient_light(const array2D<float> &R, const array2D<float> &G, c
float bright_lim = RT_INFINITY_F;
{
std::priority_queue<float> l;
std::vector<float> l;
l.reserve(patches.size() * patchsize * patchsize);
for (auto &p : patches) {
const int pW = min(p.first+patchsize, W);
const int pH = min(p.second+patchsize, H);
for (const auto &p : patches) {
const int pW = min(p.first + patchsize, W);
const int pH = min(p.second + patchsize, H);
for (int y = p.second; y < pH; ++y) {
for (int x = p.first; x < pW; ++x) {
l.push(R[y][x] + G[y][x] + B[y][x]);
l.push_back(R[y][x] + G[y][x] + B[y][x]);
}
}
}
bright_lim = get_percentile(l, 0.95);
const int pos = l.size() * 0.95;
std::nth_element(l.begin(), l.begin() + pos, l.end());
bright_lim = l[pos];
}
double rr = 0, gg = 0, bb = 0;
int n = 0;
for (auto &p : patches) {
const int pW = min(p.first+patchsize, W);
const int pH = min(p.second+patchsize, H);
#ifdef _OPENMP
#pragma omp parallel for schedule(dynamic) reduction(+:rr,gg,bb,n)
#endif
for (const auto &p : patches) {
const int pW = min(p.first + patchsize, W);
const int pH = min(p.second + patchsize, H);
for (int y = p.second; y < pH; ++y) {
for (int x = p.first; x < pW; ++x) {
float r = R[y][x];
float g = G[y][x];
float b = B[y][x];
const float r = R[y][x];
const float g = G[y][x];
const float b = B[y][x];
if (r + g + b >= bright_lim) {
rr += r;
gg += g;
@@ -181,6 +180,7 @@ float estimate_ambient_light(const array2D<float> &R, const array2D<float> &G, c
}
}
}
n = std::max(n, 1);
ambient[0] = rr / n;
ambient[1] = gg / n;
@@ -211,12 +211,12 @@ void extract_channels(Imagefloat *img, array2D<float> &r, array2D<float> &g, arr
void ImProcFunctions::dehaze(Imagefloat *img)
{
if (!params->dehaze.enabled) {
if (!params->dehaze.enabled || params->dehaze.strength == 0.0) {
return;
}
BENCHFUN
img->normalizeFloatTo1();
const int W = img->getWidth();
const int H = img->getHeight();
const float strength = LIM01(float(params->dehaze.strength) / 100.f * 0.9f);
@@ -229,21 +229,19 @@ void ImProcFunctions::dehaze(Imagefloat *img)
int patchsize = max(int(5 / scale), 2);
float ambient[3];
array2D<float> &t_tilde = dark;
float max_t = 0.f;
{
int npatches = 0;
array2D<float> R(W, H);
array2D<float> G(W, H);
array2D<float> B(W, H);
extract_channels(img, R, G, B, patchsize, 1e-1, multiThread);
patchsize = max(max(W, H) / 600, 2);
npatches = get_dark_channel(R, G, B, dark, patchsize, nullptr, false, multiThread);
DEBUG_DUMP(dark);
max_t = estimate_ambient_light(R, G, B, dark, patchsize, npatches, ambient);
patchsize = max(max(W, H) / 600, 2);
array2D<float> darkDownsized(W / patchsize + 1, H / patchsize + 1);
const int npatches = get_dark_channel_downsized(R, G, B, darkDownsized, patchsize, multiThread);
max_t = estimate_ambient_light(R, G, B, darkDownsized, patchsize, npatches, ambient);
if (options.rtSettings.verbose) {
std::cout << "dehaze: ambient light is "
@@ -251,78 +249,102 @@ void ImProcFunctions::dehaze(Imagefloat *img)
<< std::endl;
}
get_dark_channel(R, G, B, dark, patchsize, ambient, true, multiThread);
}
if (min(ambient[0], ambient[1], ambient[2]) < 0.01f) {
if (options.rtSettings.verbose) {
std::cout << "dehaze: no haze detected" << std::endl;
if (min(ambient[0], ambient[1], ambient[2]) < 0.01f) {
if (options.rtSettings.verbose) {
std::cout << "dehaze: no haze detected" << std::endl;
}
img->normalizeFloatTo65535();
return; // probably no haze at all
}
img->normalizeFloatTo65535();
return; // probably no haze at all
}
DEBUG_DUMP(t_tilde);
#ifdef _OPENMP
#pragma omp parallel for if (multiThread)
#endif
for (int y = 0; y < H; ++y) {
for (int x = 0; x < W; ++x) {
dark[y][x] = 1.f - strength * dark[y][x];
}
get_dark_channel(R, G, B, dark, patchsize, ambient, true, multiThread, strength);
}
const int radius = patchsize * 4;
const float epsilon = 1e-5;
array2D<float> &t = t_tilde;
constexpr float epsilon = 1e-5f;
{
array2D<float> guideB(W, H, img->b.ptrs, ARRAY2D_BYREFERENCE);
guidedFilter(guideB, t_tilde, t, radius, epsilon, multiThread);
guidedFilter(guideB, dark, dark, radius, epsilon, multiThread);
}
DEBUG_DUMP(t);
if (options.rtSettings.verbose) {
std::cout << "dehaze: max distance is " << max_t << std::endl;
}
float depth = -float(params->dehaze.depth) / 100.f;
const float depth = -float(params->dehaze.depth) / 100.f;
const float t0 = max(1e-3f, std::exp(depth * max_t));
const float teps = 1e-3f;
#ifdef _OPENMP
#pragma omp parallel for if (multiThread)
#endif
for (int y = 0; y < H; ++y) {
for (int x = 0; x < W; ++x) {
int x = 0;
#ifdef __SSE2__
const vfloat onev = F2V(1.f);
const vfloat ambient0v = F2V(ambient[0]);
const vfloat ambient1v = F2V(ambient[1]);
const vfloat ambient2v = F2V(ambient[2]);
const vfloat t0v = F2V(t0);
const vfloat tepsv = F2V(teps);
const vfloat c65535v = F2V(65535.f);
for (; x < W - 3; x += 4) {
// ensure that the transmission is such that to avoid clipping...
float rgb[3] = { img->r(y, x), img->g(y, x), img->b(y, x) };
vfloat r = LVFU(img->r(y, x));
vfloat g = LVFU(img->g(y, x));
vfloat b = LVFU(img->b(y, x));
// ... t >= tl to avoid negative values
float tl = 1.f - min(rgb[0]/ambient[0], rgb[1]/ambient[1], rgb[2]/ambient[2]);
const vfloat tlv = onev - vminf(r / ambient0v, vminf(g / ambient1v, b / ambient2v));
// ... t >= tu to avoid values > 1
float tu = t0 - teps;
for (int c = 0; c < 3; ++c) {
if (ambient[c] < 1) {
tu = max(tu, (rgb[c] - ambient[c])/(1.f - ambient[c]));
}
}
float mt = max(t[y][x], t0, tl + teps, tu + teps);
if (params->dehaze.showDepthMap) {
img->r(y, x) = img->g(y, x) = img->b(y, x) = LIM01(1.f - mt);
} else {
float r = (rgb[0] - ambient[0]) / mt + ambient[0];
float g = (rgb[1] - ambient[1]) / mt + ambient[1];
float b = (rgb[2] - ambient[2]) / mt + ambient[2];
r -= ambient0v;
g -= ambient1v;
b -= ambient2v;
img->r(y, x) = r;
img->g(y, x) = g;
img->b(y, x) = b;
vfloat tuv = t0v - tepsv;
tuv = vself(vmaskf_lt(ambient0v, onev), vmaxf(tuv, r / (onev - ambient0v)), tuv);
tuv = vself(vmaskf_lt(ambient1v, onev), vmaxf(tuv, g / (onev - ambient1v)), tuv);
tuv = vself(vmaskf_lt(ambient2v, onev), vmaxf(tuv, b / (onev - ambient2v)), tuv);
const vfloat mtv = vmaxf(LVFU(dark[y][x]), vmaxf(tlv, tuv) + tepsv);
if (params->dehaze.showDepthMap) {
const vfloat valv = vclampf(onev - mtv, ZEROV, onev) * c65535v;
STVFU(img->r(y, x), valv);
STVFU(img->g(y, x), valv);
STVFU(img->b(y, x), valv);
} else {
STVFU(img->r(y, x), (r / mtv + ambient0v) * c65535v);
STVFU(img->g(y, x), (g / mtv + ambient1v) * c65535v);
STVFU(img->b(y, x), (b / mtv + ambient2v) * c65535v);
}
}
#endif
for (; x < W; ++x) {
// ensure that the transmission is such that to avoid clipping...
float r = img->r(y, x);
float g = img->g(y, x);
float b = img->b(y, x);
// ... t >= tl to avoid negative values
const float tl = 1.f - min(r / ambient[0], g / ambient[1], b / ambient[2]);
// ... t >= tu to avoid values > 1
r -= ambient[0];
g -= ambient[1];
b -= ambient[2];
float tu = t0 - teps;
tu = ambient[0] < 1.f ? max(tu, r / (1.f - ambient[0])) : tu;
tu = ambient[1] < 1.f ? max(tu, g / (1.f - ambient[1])) : tu;
tu = ambient[2] < 1.f ? max(tu, b / (1.f - ambient[2])) : tu;
const float mt = max(dark[y][x], tl + teps, tu + teps);
if (params->dehaze.showDepthMap) {
img->r(y, x) = img->g(y, x) = img->b(y, x) = LIM01(1.f - mt) * 65535.f;
} else {
img->r(y, x) = (r / mt + ambient[0]) * 65535.f;
img->g(y, x) = (g / mt + ambient[1]) * 65535.f;
img->b(y, x) = (b / mt + ambient[2]) * 65535.f;
}
}
}
img->normalizeFloatTo65535();
}