// ========================================================== // Tone mapping operator (Fattal, 2002) // // Design and implementation by // - Hervé Drolon (drolon@infonie.fr) // // This file is part of FreeImage 3 // // COVERED CODE IS PROVIDED UNDER THIS LICENSE ON AN "AS IS" BASIS, WITHOUT WARRANTY // OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, WITHOUT LIMITATION, WARRANTIES // THAT THE COVERED CODE IS FREE OF DEFECTS, MERCHANTABLE, FIT FOR A PARTICULAR PURPOSE // OR NON-INFRINGING. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE COVERED // CODE IS WITH YOU. SHOULD ANY COVERED CODE PROVE DEFECTIVE IN ANY RESPECT, YOU (NOT // THE INITIAL DEVELOPER OR ANY OTHER CONTRIBUTOR) ASSUME THE COST OF ANY NECESSARY // SERVICING, REPAIR OR CORRECTION. THIS DISCLAIMER OF WARRANTY CONSTITUTES AN ESSENTIAL // PART OF THIS LICENSE. NO USE OF ANY COVERED CODE IS AUTHORIZED HEREUNDER EXCEPT UNDER // THIS DISCLAIMER. // // Use at your own risk! // ========================================================== #include "FreeImage.h" #include "Utilities.h" #include "ToneMapping.h" // ---------------------------------------------------------- // Gradient domain HDR compression // Reference: // [1] R. Fattal, D. Lischinski, and M.Werman, // Gradient domain high dynamic range compression, // ACM Transactions on Graphics, special issue on Proc. of ACM SIGGRAPH 2002, // San Antonio, Texas, vol. 21(3), pp. 257-266, 2002. // ---------------------------------------------------------- static const float EPSILON = 1e-4F; /** Performs a 5 by 5 gaussian filtering using two 1D convolutions, followed by a subsampling by 2. @param dib Input image @return Returns a blurred image of size SIZE(dib)/2 @see GaussianPyramid */ static FIBITMAP* GaussianLevel5x5(FIBITMAP *dib) { FIBITMAP *h_dib = NULL, *v_dib = NULL, *dst = NULL; float *src_pixel, *dst_pixel; try { const FREE_IMAGE_TYPE image_type = FreeImage_GetImageType(dib); if(image_type != FIT_FLOAT) throw(1); const unsigned width = FreeImage_GetWidth(dib); const unsigned height = FreeImage_GetHeight(dib); h_dib = FreeImage_AllocateT(image_type, width, height); v_dib = FreeImage_AllocateT(image_type, width, height); if (!h_dib || !v_dib) throw(1); const unsigned pitch = FreeImage_GetPitch(dib) / sizeof(float); // horizontal convolution dib -> h_dib src_pixel = (float*)FreeImage_GetBits(dib); dst_pixel = (float*)FreeImage_GetBits(h_dib); for(unsigned y = 0; y < height; y++) { // work on line y for(unsigned x = 2; x < width - 2; x++) { dst_pixel[x] = src_pixel[x-2] + src_pixel[x+2] + 4 * (src_pixel[x-1] + src_pixel[x+1]) + 6 * src_pixel[x]; dst_pixel[x] /= 16; } // boundary mirroring dst_pixel[0] = (2 * src_pixel[2] + 8 * src_pixel[1] + 6 * src_pixel[0]) / 16; dst_pixel[1] = (src_pixel[3] + 4 * (src_pixel[0] + src_pixel[2]) + 7 * src_pixel[1]) / 16; dst_pixel[width-2] = (src_pixel[width-4] + 5 * src_pixel[width-1] + 4 * src_pixel[width-3] + 6 * src_pixel[width-2]) / 16; dst_pixel[width-1] = (src_pixel[width-3] + 5 * src_pixel[width-2] + 10 * src_pixel[width-1]) / 16; // next line src_pixel += pitch; dst_pixel += pitch; } // vertical convolution h_dib -> v_dib src_pixel = (float*)FreeImage_GetBits(h_dib); dst_pixel = (float*)FreeImage_GetBits(v_dib); for(unsigned x = 0; x < width; x++) { // work on column x for(unsigned y = 2; y < height - 2; y++) { const unsigned index = y*pitch + x; dst_pixel[index] = src_pixel[index-2*pitch] + src_pixel[index+2*pitch] + 4 * (src_pixel[index-pitch] + src_pixel[index+pitch]) + 6 * src_pixel[index]; dst_pixel[index] /= 16; } // boundary mirroring dst_pixel[x] = (2 * src_pixel[x+2*pitch] + 8 * src_pixel[x+pitch] + 6 * src_pixel[x]) / 16; dst_pixel[x+pitch] = (src_pixel[x+3*pitch] + 4 * (src_pixel[x] + src_pixel[x+2*pitch]) + 7 * src_pixel[x+pitch]) / 16; dst_pixel[(height-2)*pitch+x] = (src_pixel[(height-4)*pitch+x] + 5 * src_pixel[(height-1)*pitch+x] + 4 * src_pixel[(height-3)*pitch+x] + 6 * src_pixel[(height-2)*pitch+x]) / 16; dst_pixel[(height-1)*pitch+x] = (src_pixel[(height-3)*pitch+x] + 5 * src_pixel[(height-2)*pitch+x] + 10 * src_pixel[(height-1)*pitch+x]) / 16; } FreeImage_Unload(h_dib); h_dib = NULL; // perform downsampling dst = FreeImage_Rescale(v_dib, width/2, height/2, FILTER_BILINEAR); FreeImage_Unload(v_dib); return dst; } catch(int) { if(h_dib) FreeImage_Unload(h_dib); if(v_dib) FreeImage_Unload(v_dib); if(dst) FreeImage_Unload(dst); return NULL; } } /** Compute a Gaussian pyramid using the specified number of levels. @param H Original bitmap @param pyramid Resulting pyramid array @param nlevels Number of resolution levels @return Returns TRUE if successful, returns FALSE otherwise */ static BOOL GaussianPyramid(FIBITMAP *H, FIBITMAP **pyramid, int nlevels) { try { // first level is the original image pyramid[0] = FreeImage_Clone(H); if(pyramid[0] == NULL) throw(1); // compute next levels for(int k = 1; k < nlevels; k++) { pyramid[k] = GaussianLevel5x5(pyramid[k-1]); if(pyramid[k] == NULL) throw(1); } return TRUE; } catch(int) { for(int k = 0; k < nlevels; k++) { if(pyramid[k] != NULL) { FreeImage_Unload(pyramid[k]); pyramid[k] = NULL; } } return FALSE; } } /** Compute the gradient magnitude of an input image H using central differences, and returns the average gradient. @param H Input image @param avgGrad [out] Average gradient @param k Level number @return Returns the gradient magnitude if successful, returns NULL otherwise @see GradientPyramid */ static FIBITMAP* GradientLevel(FIBITMAP *H, float *avgGrad, int k) { FIBITMAP *G = NULL; try { const FREE_IMAGE_TYPE image_type = FreeImage_GetImageType(H); if(image_type != FIT_FLOAT) throw(1); const unsigned width = FreeImage_GetWidth(H); const unsigned height = FreeImage_GetHeight(H); G = FreeImage_AllocateT(image_type, width, height); if (!G) throw(1); const unsigned pitch = FreeImage_GetPitch(H) / sizeof(float); const float divider = (float)(1 << (k + 1)); float average = 0; float *src_pixel = (float*)FreeImage_GetBits(H); float *dst_pixel = (float*)FreeImage_GetBits(G); for(unsigned y = 0; y < height; y++) { const unsigned n = (y == 0 ? 0 : y-1); const unsigned s = (y+1 == height ? y : y+1); for(unsigned x = 0; x < width; x++) { const unsigned w = (x == 0 ? 0 : x-1); const unsigned e = (x+1 == width ? x : x+1); // central difference const float gx = (src_pixel[y*pitch+e] - src_pixel[y*pitch+w]) / divider; // [Hk(x+1, y) - Hk(x-1, y)] / 2**(k+1) const float gy = (src_pixel[s*pitch+x] - src_pixel[n*pitch+x]) / divider; // [Hk(x, y+1) - Hk(x, y-1)] / 2**(k+1) // gradient dst_pixel[x] = sqrt(gx*gx + gy*gy); // average gradient average += dst_pixel[x]; } // next line dst_pixel += pitch; } *avgGrad = average / (width * height); return G; } catch(int) { if(G) FreeImage_Unload(G); return NULL; } } /** Calculate gradient magnitude and its average value on each pyramid level @param pyramid Gaussian pyramid (nlevels levels) @param nlevels Number of levels @param gradients [out] Gradient pyramid (nlevels levels) @param avgGrad [out] Average gradient on each level (array of size nlevels) @return Returns TRUE if successful, returns FALSE otherwise */ static BOOL GradientPyramid(FIBITMAP **pyramid, int nlevels, FIBITMAP **gradients, float *avgGrad) { try { for(int k = 0; k < nlevels; k++) { FIBITMAP *Hk = pyramid[k]; gradients[k] = GradientLevel(Hk, &avgGrad[k], k); if(gradients[k] == NULL) throw(1); } return TRUE; } catch(int) { for(int k = 0; k < nlevels; k++) { if(gradients[k] != NULL) { FreeImage_Unload(gradients[k]); gradients[k] = NULL; } } return FALSE; } } /** Compute the gradient attenuation function PHI(x, y) @param gradients Gradient pyramid (nlevels levels) @param avgGrad Average gradient on each level (array of size nlevels) @param nlevels Number of levels @param alpha Parameter alpha in the paper @param beta Parameter beta in the paper @return Returns the attenuation matrix Phi if successful, returns NULL otherwise */ static FIBITMAP* PhiMatrix(FIBITMAP **gradients, float *avgGrad, int nlevels, float alpha, float beta) { float *src_pixel, *dst_pixel; FIBITMAP **phi = NULL; try { phi = (FIBITMAP**)malloc(nlevels * sizeof(FIBITMAP*)); if (!phi) throw(1); memset(phi, 0, nlevels * sizeof(FIBITMAP*)); for(int k = nlevels-1; k >= 0; k--) { // compute phi(k) FIBITMAP *Gk = gradients[k]; const unsigned width = FreeImage_GetWidth(Gk); const unsigned height = FreeImage_GetHeight(Gk); const unsigned pitch = FreeImage_GetPitch(Gk) / sizeof(float); // parameter alpha is 0.1 times the average gradient magnitude // also, note the factor of 2**k in the denominator; // that is there to correct for the fact that an average gradient avgGrad(H) over 2**k pixels // in the original image will appear as a gradient grad(Hk) = 2**k*avgGrad(H) over a single pixel in Hk. float ALPHA = alpha * avgGrad[k] * (float)((int)1 << k); if(ALPHA == 0) ALPHA = EPSILON; phi[k] = FreeImage_AllocateT(FIT_FLOAT, width, height); if (!phi[k]) throw(1); src_pixel = (float*)FreeImage_GetBits(Gk); dst_pixel = (float*)FreeImage_GetBits(phi[k]); for(unsigned y = 0; y < height; y++) { for(unsigned x = 0; x < width; x++) { // compute (alpha / grad) * (grad / alpha) ** beta const float v = src_pixel[x] / ALPHA; const float value = (float)pow((float)v, (float)(beta-1)); dst_pixel[x] = (value > 1) ? 1 : value; } // next line src_pixel += pitch; dst_pixel += pitch; } if(k < nlevels-1) { // compute PHI(k) = L( PHI(k+1)) * phi(k) FIBITMAP *L = FreeImage_Rescale(phi[k+1], width, height, FILTER_BILINEAR); if (!L) throw(1); src_pixel = (float*)FreeImage_GetBits(L); dst_pixel = (float*)FreeImage_GetBits(phi[k]); for(unsigned y = 0; y < height; y++) { for(unsigned x = 0; x < width; x++) { dst_pixel[x] *= src_pixel[x]; } // next line src_pixel += pitch; dst_pixel += pitch; } FreeImage_Unload(L); // PHI(k+1) is no longer needed FreeImage_Unload(phi[k+1]); phi[k+1] = NULL; } // next level } // get the final result and return FIBITMAP *dst = phi[0]; free(phi); return dst; } catch(int) { if(phi) { for(int k = nlevels-1; k >= 0; k--) { if(phi[k]) FreeImage_Unload(phi[k]); } free(phi); } return NULL; } } /** Compute gradients in x and y directions, attenuate them with the attenuation matrix, then compute the divergence div G from the attenuated gradient. @param H Normalized luminance @param PHI Attenuation matrix @return Returns the divergence matrix if successful, returns NULL otherwise */ static FIBITMAP* Divergence(FIBITMAP *H, FIBITMAP *PHI) { FIBITMAP *Gx = NULL, *Gy = NULL, *divG = NULL; float *phi, *h, *gx, *gy, *divg; try { const FREE_IMAGE_TYPE image_type = FreeImage_GetImageType(H); if(image_type != FIT_FLOAT) throw(1); const unsigned width = FreeImage_GetWidth(H); const unsigned height = FreeImage_GetHeight(H); Gx = FreeImage_AllocateT(image_type, width, height); if (!Gx) throw(1); Gy = FreeImage_AllocateT(image_type, width, height); if (!Gy) throw(1); const unsigned pitch = FreeImage_GetPitch(H) / sizeof(float); // perform gradient attenuation phi = (float*)FreeImage_GetBits(PHI); h = (float*)FreeImage_GetBits(H); gx = (float*)FreeImage_GetBits(Gx); gy = (float*)FreeImage_GetBits(Gy); for(unsigned y = 0; y < height; y++) { const unsigned s = (y+1 == height ? y : y+1); for(unsigned x = 0; x < width; x++) { const unsigned e = (x+1 == width ? x : x+1); // forward difference const unsigned index = y*pitch + x; const float phi_xy = phi[index]; const float h_xy = h[index]; gx[x] = (h[y*pitch+e] - h_xy) * phi_xy; // [H(x+1, y) - H(x, y)] * PHI(x, y) gy[x] = (h[s*pitch+x] - h_xy) * phi_xy; // [H(x, y+1) - H(x, y)] * PHI(x, y) } // next line gx += pitch; gy += pitch; } // calculate the divergence divG = FreeImage_AllocateT(image_type, width, height); if (!divG) throw(1); gx = (float*)FreeImage_GetBits(Gx); gy = (float*)FreeImage_GetBits(Gy); divg = (float*)FreeImage_GetBits(divG); for(unsigned y0 = 0; y0 < height; y0++) { for(unsigned x = 0; x < width; x++) { // backward difference approximation // divG = Gx(x, y) - Gx(x-1, y) + Gy(x, y) - Gy(x, y-1) const unsigned index = y0*pitch + x; divg[index] = gx[index] + gy[index]; if(x > 0) divg[index] -= gx[index-1]; if(y0 > 0) divg[index] -= gy[index-pitch]; } } // no longer needed ... FreeImage_Unload(Gx); FreeImage_Unload(Gy); // return the divergence return divG; } catch(int) { if(Gx) FreeImage_Unload(Gx); if(Gy) FreeImage_Unload(Gy); if(divG) FreeImage_Unload(divG); return NULL; } } /** Given the luminance channel, find max & min luminance values, normalize to range 0..100 and take the logarithm. @param Y Image luminance @return Returns the normalized luminance H if successful, returns NULL otherwise */ static FIBITMAP* LogLuminance(FIBITMAP *Y) { FIBITMAP *H = NULL; try { // get the luminance channel FIBITMAP *H = FreeImage_Clone(Y); if (!H) throw(1); const unsigned width = FreeImage_GetWidth(H); const unsigned height = FreeImage_GetHeight(H); const unsigned pitch = FreeImage_GetPitch(H); // find max & min luminance values float maxLum = -1e20F, minLum = 1e20F; BYTE *bits = (BYTE*)FreeImage_GetBits(H); for(unsigned y = 0; y < height; y++) { const float *pixel = (float*)bits; for(unsigned x = 0; x < width; x++) { const float value = pixel[x]; maxLum = (maxLum < value) ? value : maxLum; // max Luminance in the scene minLum = (minLum < value) ? minLum : value; // min Luminance in the scene } // next line bits += pitch; } if(maxLum == minLum) throw(1); // normalize to range 0..100 and take the logarithm const float scale = 100.F / (maxLum - minLum); bits = (BYTE*)FreeImage_GetBits(H); for(unsigned y0 = 0; y0 < height; y0++) { float *pixel = (float*)bits; for(unsigned x = 0; x < width; x++) { const float value = (pixel[x] - minLum) * scale; pixel[x] = log(value + EPSILON); } // next line bits += pitch; } return H; } catch(int) { if(H) FreeImage_Unload(H); return NULL; } } /** Given a normalized luminance, perform exponentiation and recover the log compressed image @param Y Input/Output luminance image */ static void ExpLuminance(FIBITMAP *Y) { const unsigned width = FreeImage_GetWidth(Y); const unsigned height = FreeImage_GetHeight(Y); const unsigned pitch = FreeImage_GetPitch(Y); BYTE *bits = (BYTE*)FreeImage_GetBits(Y); for(unsigned y = 0; y < height; y++) { float *pixel = (float*)bits; for(unsigned x = 0; x < width; x++) { pixel[x] = exp(pixel[x]) - EPSILON; } bits += pitch; } } // -------------------------------------------------------------------------- /** Gradient Domain HDR tone mapping operator @param Y Image luminance values @param alpha Parameter alpha of the paper (suggested value is 0.1) @param beta Parameter beta of the paper (suggested value is between 0.8 and 0.9) @return returns the tone mapped luminance */ static FIBITMAP* tmoFattal02(FIBITMAP *Y, float alpha, float beta) { const unsigned MIN_PYRAMID_SIZE = 32; // minimun size (width or height) of the coarsest level of the pyramid FIBITMAP *H = NULL; FIBITMAP **pyramid = NULL; FIBITMAP **gradients = NULL; FIBITMAP *phy = NULL; FIBITMAP *divG = NULL; FIBITMAP *U = NULL; float *avgGrad = NULL; int k; int nlevels = 0; try { // get the normalized luminance FIBITMAP *H = LogLuminance(Y); if (!H) throw(1); // get the number of levels for the pyramid const unsigned width = FreeImage_GetWidth(H); const unsigned height = FreeImage_GetHeight(H); unsigned minsize = MIN(width, height); while(minsize >= MIN_PYRAMID_SIZE) { nlevels++; minsize /= 2; } // create the Gaussian pyramid pyramid = (FIBITMAP**)malloc(nlevels * sizeof(FIBITMAP*)); if (!pyramid) throw(1); memset(pyramid, 0, nlevels * sizeof(FIBITMAP*)); if (!GaussianPyramid(H, pyramid, nlevels)) throw(1); // calculate gradient magnitude and its average value on each pyramid level gradients = (FIBITMAP**)malloc(nlevels * sizeof(FIBITMAP*)); if (!gradients) throw(1); memset(gradients, 0, nlevels * sizeof(FIBITMAP*)); avgGrad = (float*)malloc(nlevels * sizeof(float)); if (!avgGrad) throw(1); if (!GradientPyramid(pyramid, nlevels, gradients, avgGrad)) throw(1); // free the Gaussian pyramid for(k = 0; k < nlevels; k++) { if(pyramid[k]) FreeImage_Unload(pyramid[k]); } free(pyramid); pyramid = NULL; // compute the gradient attenuation function PHI(x, y) phy = PhiMatrix(gradients, avgGrad, nlevels, alpha, beta); if (!phy) throw(1); // free the gradient pyramid for(k = 0; k < nlevels; k++) { if(gradients[k]) FreeImage_Unload(gradients[k]); } free(gradients); gradients = NULL; free(avgGrad); avgGrad = NULL; // compute gradients in x and y directions, attenuate them with the attenuation matrix, // then compute the divergence div G from the attenuated gradient. divG = Divergence(H, phy); if (!divG) throw(1); // H & phy no longer needed FreeImage_Unload(H); H = NULL; FreeImage_Unload(phy); phy = NULL; // solve the PDE (Poisson equation) using a multigrid solver and 3 cycles FIBITMAP *U = FreeImage_MultigridPoissonSolver(divG, 3); if (!U) throw(1); FreeImage_Unload(divG); // perform exponentiation and recover the log compressed image ExpLuminance(U); return U; } catch(int) { if(H) FreeImage_Unload(H); if(pyramid) { for(int i = 0; i < nlevels; i++) { if(pyramid[i]) FreeImage_Unload(pyramid[i]); } free(pyramid); } if(gradients) { for(int i = 0; i < nlevels; i++) { if(gradients[i]) FreeImage_Unload(gradients[i]); } free(gradients); } if(avgGrad) free(avgGrad); if(phy) FreeImage_Unload(phy); if(divG) FreeImage_Unload(divG); if(U) FreeImage_Unload(U); return NULL; } } // ---------------------------------------------------------- // Main algorithm // ---------------------------------------------------------- /** Apply the Gradient Domain High Dynamic Range Compression to a RGBF image and convert to 24-bit RGB @param dib Input RGBF / RGB16 image @param color_saturation Color saturation (s parameter in the paper) in [0.4..0.6] @param attenuation Atenuation factor (beta parameter in the paper) in [0.8..0.9] @return Returns a 24-bit RGB image if successful, returns NULL otherwise */ FIBITMAP* DLL_CALLCONV FreeImage_TmoFattal02(FIBITMAP *dib, double color_saturation, double attenuation) { const float alpha = 0.1F; // parameter alpha = 0.1 const float beta = (float)MAX(0.8, MIN(0.9, attenuation)); // parameter beta = [0.8..0.9] const float s = (float)MAX(0.4, MIN(0.6, color_saturation));// exponent s controls color saturation = [0.4..0.6] FIBITMAP *src = NULL; FIBITMAP *Yin = NULL; FIBITMAP *Yout = NULL; FIBITMAP *dst = NULL; if (!FreeImage_HasPixels(dib)) return NULL; try { // convert to RGBF src = FreeImage_ConvertToRGBF(dib); if (!src) throw(1); // get the luminance channel Yin = ConvertRGBFToY(src); if (!Yin) throw(1); // perform the tone mapping Yout = tmoFattal02(Yin, alpha, beta); if (!Yout) throw(1); // clip low and high values and normalize to [0..1] //NormalizeY(Yout, 0.001F, 0.995F); NormalizeY(Yout, 0, 1); // compress the dynamic range const unsigned width = FreeImage_GetWidth(src); const unsigned height = FreeImage_GetHeight(src); const unsigned rgb_pitch = FreeImage_GetPitch(src); const unsigned y_pitch = FreeImage_GetPitch(Yin); BYTE *bits = (BYTE*)FreeImage_GetBits(src); BYTE *bits_yin = (BYTE*)FreeImage_GetBits(Yin); BYTE *bits_yout = (BYTE*)FreeImage_GetBits(Yout); for(unsigned y = 0; y < height; y++) { float *Lin = (float*)bits_yin; float *Lout = (float*)bits_yout; float *color = (float*)bits; for(unsigned x = 0; x < width; x++) { for(unsigned c = 0; c < 3; c++) { *color = (Lin[x] > 0) ? pow(*color/Lin[x], s) * Lout[x] : 0; color++; } } bits += rgb_pitch; bits_yin += y_pitch; bits_yout += y_pitch; } // not needed anymore FreeImage_Unload(Yin); Yin = NULL; FreeImage_Unload(Yout); Yout = NULL; // clamp image highest values to display white, then convert to 24-bit RGB dst = ClampConvertRGBFTo24(src); // clean-up and return FreeImage_Unload(src); src = NULL; // copy metadata from src to dst FreeImage_CloneMetadata(dst, dib); return dst; } catch(int) { if(src) FreeImage_Unload(src); if(Yin) FreeImage_Unload(Yin); if(Yout) FreeImage_Unload(Yout); return NULL; } }