1 #pragma once 2 /* 3 The EdgePreservingDecomposition files contain standard C++ (standard except the 4 first line) code for creating and, to a limited extent (create your own uses!), 5 messing with multi scale edge preserving decompositions of a 32 bit single 6 channel image. As a byproduct it contains a lot of linear algebra which can be 7 useful for optimization problems that you want to solve in rectangles on 8 rectangular grids. 9 10 Anyway. Basically, this is an implementation of what's presented in the 11 following papers: 12 - Edge-Preserving Decompositions for Multi-Scale Tone and Detail Manipulation 13 - An Iterative Solution Method for Linear Systems of Which the Coefficient 14 Matrix is a Symmetric M-Matrix 15 - Color correction for tone mapping 16 - Wikipedia, the free encyclopedia 17 18 First one is most of what matters, next two are details, last everything else. 19 I did a few things differently, especially: 20 - Reformulated the minimization with finite elements instead of finite 21 differences. This results in better conditioning, slightly better accuracy 22 (less artifacts), the possibility of a better picked edge stopping function, 23 but more memory consumption. 24 - A single rotationally invariant edge stopping function is used instead of two 25 non-invariant ones. 26 - Incomplete Cholseky factorization instead of Szeliski's LAHBF. Slower, but 27 not subject to any patents. 28 - For tone mapping, original images are decomposed instead of their logarithms, 29 and just one decomposition is made; 30 - I find that this way works plenty good (theirs isn't better or worse... just 31 different) and is simpler. 32 33 Written by ben_pcc in Portland, Oregon, USA. Some history: 34 - Late April 2010, I develop interest in this stuff because photos of my 35 ceramics lack local contrast. 36 - Mid 2010, it works but is too slow to be useful. 37 - Fall 2010, various unsuccessful attempts at speeding up are tried. 38 - Early December 2010, I get off the path of least resistance and write a 39 matrix storage class with incomplete Cholesky decomposition. 40 - 31 December 2010, the FEM reformulation works very well. 41 - 1 January 2011, I'm cleaning up this file and readying it for initial release. 42 - 12 - 14 November 2011, further cleanup, improvements, bug fixes, integration 43 into Raw Therapee. 44 45 It's likely that I'll take apart and rerelease contents of this file (in the 46 distant future) as most of it isn't edge preserving decomposition and rather 47 supporting material. SparseConjugateGradient alone is a workhorse I and a few 48 others have been exploiting for a few years. 49 50 EdgePreservingDecomposition.h and EdgePreservingDecomposition.cpp are released 51 under the following licence: 52 - It's free. 53 - You may not incorporate this code as part of proprietary or commercial 54 software, but via freeware you may use its output for profit. 55 - You may modify and redistribute, but keep this big comment block intact and 56 not for profit in any way unless I give specific permission. 57 - If you're unsure about anything else, treat as public domain. 58 - Don't be a dick. 59 60 My email address is my screen name followed by @yahoo.com. I'm also known as 61 ben_s or nonbasketless. Enjoy! 62 */ 63 64 65 66 #include <cmath> 67 #include <cstdio> 68 #include <cstdlib> 69 #include <cstring> 70 71 #include "opthelper.h" 72 #include "noncopyable.h" 73 74 //This is for solving big symmetric positive definite linear problems. 75 float *SparseConjugateGradient(void Ax(float *Product, float *x, void *Pass), float *b, int n, bool OkToModify_b = true, float *x = nullptr, float RMSResidual = 0.0f, void *Pass = nullptr, int MaximumIterates = 0, void Preconditioner(float *Product, float *x, void *Pass) = nullptr); 76 77 //Storage and use class for symmetric matrices, the nonzero contents of which are confined to diagonals. 78 class MultiDiagonalSymmetricMatrix : 79 public rtengine::NonCopyable 80 { 81 public: 82 MultiDiagonalSymmetricMatrix(int Dimension, int NumberOfDiagonalsInLowerTriangle); 83 ~MultiDiagonalSymmetricMatrix(); 84 85 /* Storage of matrix data, and a function to create memory for Diagonals[index]. 86 Here's the storage scheme, designed to be simple both on paper and in C++: 87 88 Let's say you have some diagonal. The StartRows is the row on which, at the left edge of the matrix, the diagonal "starts", 89 and StartRows must strictly increase with its index. The main diagonal for example has start row 0, its subdiagonal has 1, etc. 90 Then, Diagonal[j] is the matrix entry on the diagonal at column j. For efficiency, you're expected to learn this and fill in 91 public Diagonals manually. Symmetric matrices are represented by this class, and all symmetry is handled internally, you 92 only every worry or think about the lower triangular (including main diagonal) part of the matrix. 93 */ 94 float **Diagonals; 95 char *buffer; 96 float *DiagBuffer; 97 int *StartRows; 98 bool CreateDiagonal(int index, int StartRow); 99 int n, m; //The matrix is n x n, with m diagonals on the lower triangle. Don't change these. They should be private but aren't for convenience. DiagonalLength(int StartRow)100 inline int DiagonalLength(int StartRow) //Gives number of elements in a diagonal. 101 { 102 return n - StartRow; 103 }; 104 105 //Not efficient, but you can use it if you're lazy, or for early tests. Returns false if the row + column falls on no loaded diagonal, true otherwise. 106 bool LazySetEntry(float value, int row, int column); 107 108 //Calculates the matrix-vector product of the matrix represented by this class onto the vector x. 109 void VectorProduct(float *Product, float *x); 110 111 //Given the start row, attempts to find the corresponding index, or -1 if the StartRow doesn't exist. 112 inline int FindIndex(int StartRow) __attribute__((always_inline)); 113 114 //This is the same as above, but designed to take this class as a pass through variable. By this way you can feed 115 //the meat of this class into an independent function, such as SparseConjugateGradient. PassThroughVectorProduct(float * Product,float * x,void * Pass)116 static void PassThroughVectorProduct(float *Product, float *x, void *Pass) 117 { 118 (static_cast<MultiDiagonalSymmetricMatrix *>(Pass))->VectorProduct(Product, x); 119 }; 120 121 /* CreateIncompleteCholeskyFactorization creates another matrix which is an incomplete (or complete if MaxFillAbove is big enough) 122 LDLt factorization of this matrix. Storage is like this: the first diagonal is the diagonal matrix D and the remaining diagonals 123 describe all of L except its main diagonal, which is a bunch of ones. Read up on the LDLt Cholesky factorization for what all this means. 124 Note that VectorProduct is nonsense. More useful to you is CholeskyBackSolve which fills x, where LDLt x = b. */ 125 bool CreateIncompleteCholeskyFactorization(int MaxFillAbove = 0); 126 void KillIncompleteCholeskyFactorization(void); 127 void CholeskyBackSolve(float *x, float *b); 128 MultiDiagonalSymmetricMatrix *IncompleteCholeskyFactorization; 129 PassThroughCholeskyBackSolve(float * Product,float * x,void * Pass)130 static void PassThroughCholeskyBackSolve(float *Product, float *x, void *Pass) 131 { 132 (static_cast<MultiDiagonalSymmetricMatrix *>(Pass))->CholeskyBackSolve(Product, x); 133 }; 134 135 }; 136 137 class EdgePreservingDecomposition : 138 public rtengine::NonCopyable 139 { 140 public: 141 EdgePreservingDecomposition(int width, int height); 142 ~EdgePreservingDecomposition(); 143 144 //Create an edge preserving blur of Source. Will create and return, or fill into Blur if not NULL. In place not ok. 145 //If UseBlurForEdgeStop is true, supplied not NULL Blur is used to calculate the edge stopping function instead of Source. 146 float *CreateBlur(float *Source, float Scale, float EdgeStopping, int Iterates, float *Blur = nullptr, bool UseBlurForEdgeStop = false); 147 148 //Iterates CreateBlur such that the smoothness term approaches a specific norm via iteratively reweighted least squares. In place not ok. 149 float *CreateIteratedBlur(float *Source, float Scale, float EdgeStopping, int Iterates, int Reweightings, float *Blur = nullptr); 150 151 /*Lowers global contrast while preserving or boosting local contrast. Can fill into Compressed. The smaller Compression 152 the more compression is applied, with Compression = 1 giving no effect and above 1 the opposite effect. You can totally 153 use Compression = 1 and play with DetailBoost for some really sweet unsharp masking. If working on luma/grey, consider giving it a logarithm. 154 In place calculation to save memory (Source == Compressed) is totally ok. Reweightings > 0 invokes CreateIteratedBlur instead of CreateBlur. */ 155 void CompressDynamicRange(float *Source, float Scale = 1.0f, float EdgeStopping = 1.4f, float CompressionExponent = 0.8f, float DetailBoost = 0.1f, int Iterates = 20, int Reweightings = 0); 156 157 private: 158 MultiDiagonalSymmetricMatrix *A; //The equations are simple enough to not mandate a matrix class, but fast solution NEEDS a complicated preconditioner. 159 int w, h, n; 160 161 //Convenient access to the data in A. 162 float * RESTRICT a0, * RESTRICT a_1, * RESTRICT a_w, * RESTRICT a_w_1, * RESTRICT a_w1; 163 }; 164 165