|
| 1 | +function [regions, thresholds, maxSig] = multilevelOtsuThreshold(obj, nClasses, varargin) |
| 2 | +% Multilevel Thresholding using Otsu Method. |
| 3 | +% |
| 4 | +% CLASSES = multilevelOtsuThreshold(IMG, NC) |
| 5 | +% Computes a segmentation of the input grayscale image IMG into NC |
| 6 | +% classes. The number of classes must be comprised between 2 and 5. |
| 7 | +% |
| 8 | +% [CLASSES, THRESHOLDS] = multilevelOtsuThreshold(IMG, NC) |
| 9 | +% Also returns the list of threshold. The number of threshold values is |
| 10 | +% the number of classes minus one. |
| 11 | +% |
| 12 | +% Example |
| 13 | +% % segment cameraman image into three classes |
| 14 | +% img = Image.read('cameraman.tif'); |
| 15 | +% [classes, threshs] = multilevelOtsuThreshold(img, 3); |
| 16 | +% rgb = label2rgb(classes, 'jet', [1 0 0], 'shuffle'); |
| 17 | +% figure; show(rgb); |
| 18 | +% % also displays histogram with threshold levels |
| 19 | +% figure; histogram(img); |
| 20 | +% hold on; |
| 21 | +% for i = 1:2 |
| 22 | +% plot(threshs([i i]), [0 1800], 'color', 'r', 'linewidth', 2); |
| 23 | +% end |
| 24 | +% |
| 25 | +% References |
| 26 | +% * Ping-Sung Liao, Tse-Sheng Chen, Pau-Choo Chung (2001). "A Fast |
| 27 | +% Algorithm for Multilevel Thresholding". Journal of Information Science |
| 28 | +% and Engineering, Vol. 17 No. 5, pp. 713-727. |
| 29 | +% https://jise.iis.sinica.edu.tw/JISESearch/pages/View/PaperView.jsf?keyId=86_1302# |
| 30 | +% * https://imagej.net/Multi_Otsu_Threshold |
| 31 | +% |
| 32 | +% See also |
| 33 | +% otsuThreshold, histogram, maxEntropyThreshold |
| 34 | +% |
| 35 | + |
| 36 | +% ------ |
| 37 | +% Author: David Legland |
| 38 | +% e-mail: david.legland@inrae.fr |
| 39 | +% INRAE - BIA Research Unit - BIBS Platform (Nantes) |
| 40 | +% Created: 2021-02-02, using Matlab 9.8.0.1323502 (R2020a) |
| 41 | +% Copyright 2021 INRAE. |
| 42 | + |
| 43 | + |
| 44 | +%% Input arguments |
| 45 | + |
| 46 | +if ~isGrayscaleImage(obj) |
| 47 | + error('Requires a grayscale uint8 image as first input'); |
| 48 | +end |
| 49 | + |
| 50 | +if nClasses < 2 || nClasses > 5 |
| 51 | + error('The number of classes must be comprised between 2 and 5.'); |
| 52 | +end |
| 53 | + |
| 54 | + |
| 55 | +%% Initialisations |
| 56 | + |
| 57 | +% compute histogram, and convert into probability density |
| 58 | +[h, levels] = histogram(obj, varargin{:}); |
| 59 | +h = h / sum(h); |
| 60 | + |
| 61 | +% number of gray levels |
| 62 | +nLevels = length(levels); |
| 63 | + |
| 64 | +Puv = zeros(nLevels, nLevels); |
| 65 | +Suv = zeros(nLevels, nLevels); |
| 66 | +Huv = zeros(nLevels, nLevels); |
| 67 | + |
| 68 | +% initialize diagonal terms |
| 69 | +for i = 1:nLevels |
| 70 | + Puv(i, i) = h(i); |
| 71 | + Suv(i, i) = h(i) * i; |
| 72 | +end |
| 73 | + |
| 74 | +% initialize first rows |
| 75 | +for i = 2:nLevels |
| 76 | + Puv(1, i) = Puv(1, i-1) + h(i); |
| 77 | + Suv(1, i) = Suv(1, i-1) + i*h(i); |
| 78 | +end |
| 79 | +% initialize the rest of the matrices |
| 80 | +for u = 2:nLevels |
| 81 | + for v = u+1:nLevels |
| 82 | + Puv(u, v) = Puv(1, v) - Puv(1, u-1); |
| 83 | + Suv(u, v) = Suv(1, v) - Suv(1, u-1); |
| 84 | + end |
| 85 | +end |
| 86 | + |
| 87 | +% now calculate Huv |
| 88 | +for u = 1:nLevels |
| 89 | + for v = u+1:nLevels |
| 90 | + if Puv(u,v) > 0 |
| 91 | + Huv(u,v) = Suv(u,v) * Suv(u,v) / Puv(u,v); |
| 92 | + end |
| 93 | + end |
| 94 | +end |
| 95 | + |
| 96 | + |
| 97 | +%% Main processing |
| 98 | + |
| 99 | +% create array for threshold values (first one is zeros, last one is inf) |
| 100 | +thresholds = zeros(nClasses+1, 1); |
| 101 | +thresholds(end) = inf; |
| 102 | + |
| 103 | +% iterate over all possible combinations |
| 104 | +maxSig = 0.0; |
| 105 | +switch nClasses |
| 106 | + case 2 |
| 107 | + % two classes -> only need to find second threshold value |
| 108 | + for i = 1:nLevels-1 |
| 109 | + Sq = Huv(1,i) + Huv(i+1, end); |
| 110 | + if Sq >= maxSig |
| 111 | + thresholds(2) = i; |
| 112 | + maxSig = Sq; |
| 113 | + end |
| 114 | + end |
| 115 | + |
| 116 | + case 3 |
| 117 | + % three classes |
| 118 | + for i = 1:nLevels-2 |
| 119 | + for j = i+1:nLevels-1 |
| 120 | + Sq = Huv(1,i) + Huv(i+1, j) + Huv(j+1, end); |
| 121 | + if Sq >= maxSig |
| 122 | + thresholds(2) = i; |
| 123 | + thresholds(3) = j; |
| 124 | + maxSig = Sq; |
| 125 | + end |
| 126 | + end |
| 127 | + end |
| 128 | + |
| 129 | + case 4 |
| 130 | + % four classes |
| 131 | + for i = 1:nLevels-3 |
| 132 | + for j = i+1:nLevels-2 |
| 133 | + for k = j+1:nLevels-1 |
| 134 | + Sq = Huv(1,i) + Huv(i+1, j) + Huv(j+1, k) + Huv(k+1, end); |
| 135 | + if Sq >= maxSig |
| 136 | + thresholds(2) = i; |
| 137 | + thresholds(3) = j; |
| 138 | + thresholds(4) = k; |
| 139 | + maxSig = Sq; |
| 140 | + end |
| 141 | + end |
| 142 | + end |
| 143 | + end |
| 144 | + |
| 145 | + case 5 |
| 146 | + % five classes |
| 147 | + for i = 1:nLevels-3 |
| 148 | + for j = i+1:nLevels-2 |
| 149 | + for k = j+1:nLevels-1 |
| 150 | + for m = k+1:nLevels-1 |
| 151 | + Sq = Huv(1,i) + Huv(i+1, j) + Huv(j+1, k) + Huv(k+1, m) + Huv(m+1, end); |
| 152 | + if Sq >= maxSig |
| 153 | + thresholds(2) = i; |
| 154 | + thresholds(3) = j; |
| 155 | + thresholds(4) = k; |
| 156 | + thresholds(5) = m; |
| 157 | + maxSig = Sq; |
| 158 | + end |
| 159 | + end |
| 160 | + end |
| 161 | + end |
| 162 | + end |
| 163 | + |
| 164 | + otherwise |
| 165 | + error('Can not manage %d number of classes', nClasses); |
| 166 | +end |
| 167 | + |
| 168 | + |
| 169 | +%% Create region image |
| 170 | + |
| 171 | +% compute array of region labels |
| 172 | +regData = zeros(size(obj), 'uint8'); |
| 173 | +for i = 1:nClasses |
| 174 | + inds = obj >= thresholds(i) & obj < thresholds(i+1); |
| 175 | + regData(inds) = i; |
| 176 | +end |
| 177 | + |
| 178 | +% create result image |
| 179 | +newName = createNewName(obj, '%s-OtsuClasses'); |
| 180 | +regions = Image(... |
| 181 | + 'Data', regData, ... |
| 182 | + 'Parent', obj, ... |
| 183 | + 'Type', 'Label', ... |
| 184 | + 'Name', newName); |
| 185 | + |
| 186 | +% format threshold result |
| 187 | +thresholds([1 end]) = []; |
0 commit comments