multithresh

Multilevel image thresholds using Otsu’s method

Syntax

thresh = multithresh(A)

thresh = multithresh(A,N)

[thresh,metric] = multithresh(___)

Description

thresh = multithresh(A) returns the single threshold value thresh computed for image A using Otsu’s method. You can use thresh as an input argument to imquantize to convert an image into a two-level image.

thresh = multithresh(A,N) returns thresh a 1-by-N vector containing N threshold values using Otsu’s method. You can use thresh as an input argument to imquantize to convert image A into an image with N+1 discrete levels.

[thresh,metric] = multithresh(___) returns metric, a measure of the effectiveness of the computed thresholds.

Examples

Segment Image Into Two Regions

Read image and display it.

I = imread('coins.png');
imshow(I)

Calculate a single threshold value for the image.

level = multithresh(I);

Segment the image into two regions using imquantize , specifying the threshold level returned by multithresh .

seg_I = imquantize(I,level);
figure
imshow(seg_I,[])

Segment Image into Three Levels Using Two Thresholds

Read an image, convert it to grayscale, and display the result.

I = imread("foggysf2.jpg");
I = rgb2gray(I);
imshow(I)
title("Grayscale Image")

Calculate two threshold levels.

thresh = multithresh(I,2);

Segment the image into three levels using imquantize .

labels = imquantize(I,thresh);

Convert the segmented image into a color image using label2rgb and display it.

labelsRGB = label2rgb(labels);
imshow(labelsRGB)
title("Segmented Image")

Compare Thresholding Entire Image Versus Plane-by-Plane Thresholding

Read truecolor (RGB) image and display it.

I = imread('peppers.png');
imshow(I) 
axis off
title('RGB Image');

Generate thresholds for seven levels from the entire RGB image.

threshRGB = multithresh(I,7);

Generate thresholds for each plane of the RGB image.

threshForPlanes = zeros(3,7);			

for i = 1:3
    threshForPlanes(i,:) = multithresh(I(:,:,i),7);
end

Process the entire image with the set of threshold values computed from entire image.

value = [0 threshRGB(2:end) 255]; 
quantRGB = imquantize(I, threshRGB, value);

Process each RGB plane separately using the threshold vector computed from the given plane. Quantize each RGB plane using threshold vector generated for that plane.

quantPlane = zeros( size(I) );

for i = 1:3
    value = [0 threshForPlanes(i,2:end) 255]; 
    quantPlane(:,:,i) = imquantize(I(:,:,i),threshForPlanes(i,:),value);
end

quantPlane = uint8(quantPlane);

Display both posterized images and note the visual differences in the two thresholding schemes.

imshowpair(quantRGB,quantPlane,'montage') 
axis off
title('Full RGB Image Quantization        Plane-by-Plane Quantization')

To compare the results, calculate the number of unique RGB pixel vectors in each output image. Note that the plane-by-plane thresholding scheme yields about 23% more colors than the full RGB image scheme.

dim = size( quantRGB );
quantRGBmx3   = reshape(quantRGB,   prod(dim(1:2)), 3);
quantPlanemx3 = reshape(quantPlane, prod(dim(1:2)), 3);

colorsRGB   = unique(quantRGBmx3,   'rows' );
colorsPlane = unique(quantPlanemx3, 'rows' );

disp(['Unique colors in RGB image            : ' int2str(length(colorsRGB))]);

Unique colors in RGB image            : 188

disp(['Unique colors in Plane-by-Plane image : ' int2str(length(colorsPlane))]);

Unique colors in Plane-by-Plane image : 231

Check Results Using the Metric Output Argument

Read image.

I = imread('circlesBrightDark.png');

Find all unique grayscale values in image.

uniqLevels = unique(I(:));  

disp(['Number of unique levels = ' int2str( length(uniqLevels) )]);

Number of unique levels = 148

Compute a series of thresholds at monotonically increasing values of N.

Nvals = [1 2 4 8];
for i = 1:length(Nvals)
    [thresh, metric] = multithresh(I, Nvals(i) );
    disp(['N = ' int2str(Nvals(i)) '  |  metric = ' num2str(metric)]);
end

N = 1  |  metric = 0.54767
N = 2  |  metric = 0.98715
N = 4  |  metric = 0.99648
N = 8  |  metric = 0.99902

Apply the set of 8 threshold values to obtain a 9-level segmentation using imquantize .

seg_Neq8 = imquantize(I,thresh);
uniqLevels = unique( seg_Neq8(:) )

uniqLevels = 9×1

     1
     2
     3
     4
     5
     6
     7
     8
     9

Threshold the image using seg_Neq8 as an input to multithresh. Set N equal to 8, which is 1 less than the number of levels in this segmented image. multithresh returns a metric value of 1.

[thresh, metric] = multithresh(seg_Neq8,8)

thresh = 1×8

    1.8784    2.7882    3.6667    4.5451    5.4549    6.3333    7.2118    8.1216

metric = 1

Threshold the image again, this time increasing the value of N by 1. This value now equals the number of levels in the image. Note how the input is degenerate because the number of levels in the image is too few for the number of requested thresholds. Hence, multithresh returns a metric value of 0.

[thresh, metric] = multithresh(seg_Neq8,9)

Warning: No solution exists because the number of unique levels in the image are too few to find 9 thresholds. Returning an arbitrarily chosen solution.

thresh = 1×9

     1     2     3     4     5     6     7     8     9

metric = 0

Input Arguments

A — Image to be thresholded

numeric array

Image to be thresholded, specified as a numeric array of any dimension. multithresh finds the thresholds based on the aggregate histogram of the entire array. multithresh considers an RGB image as a 3-D numeric array and computes the thresholds for the combined data from all three color planes.

multithresh uses the range of the input image A, [min(A(:)) max(A(:))], as the limits for computing the histogram used in subsequent computations. multithresh ignores any NaNs in computation. Any Infs and -Infs are counted in the first and last bin of the histogram, respectively.

For degenerate inputs where the number of unique values in A is less than or equal to N, there is no viable solution using Otsu's method. For such inputs, the return value thresh contains all the unique values from A and possibly some extra values that are chosen arbitrarily.

Data Types: single | double | int16 | uint8 | uint16

N — Number of threshold values

1 (default) | positive integer

Number of threshold values, specified as a positive integer. For N > 2, multithresh uses search-based optimization of Otsu's criterion to find the thresholds. The search-based optimization guarantees only locally optimal results. Since the chance of converging to local optimum increases with N, it is preferable to use smaller values of N, typically N < 10. The maximum allowed value for N is 20.

Data Types: single | double | int8 | int16 | int32 | int64 | uint8 | uint16 | uint32 | uint64

Output Arguments

thresh — Set of threshold values

1-by-N numeric vector

Set of threshold values used to quantize an image, returned as a 1-by-N numeric vector, whose data type is the same as image A.

These thresholds are in the same range as the input image A, unlike the graythresh function, which returns a normalized threshold in the range [0, 1].

metric — Measure of effectiveness

number in the range [0, 1]

Measure of the effectiveness of the thresholds, returned as a number in the range [0, 1]. Higher values indicates greater effectiveness of the thresholds in separating the input image into N+1 classes based on Otsu's objective criterion. For degenerate inputs where the number of unique values in A is less than or equal to N, metric equals 0.

Data Types: double

References

[1] Otsu, N., "A Threshold Selection Method from Gray-Level Histograms," IEEE Transactions on Systems, Man, and Cybernetics, Vol. 9, No. 1, 1979, pp. 62-66.

Extended Capabilities

C/C++ Code Generation

Generate C and C++ code using MATLAB® Coder™.

Usage notes and limitations:

• multithresh supports the generation of C code (requires MATLAB^® Coder™). Note that if you choose the generic MATLAB Host Computer target platform, multithresh generates code that uses a precompiled, platform-specific shared library. Use of a shared library preserves performance optimizations but limits the target platforms for which code can be generated. For more information, see Types of Code Generation Support in Image Processing Toolbox.

• The input argument N must be a compile-time constant.

GPU Code Generation

Generate CUDA® code for NVIDIA® GPUs using GPU Coder™.

Usage notes and limitations:

• The input argument N must be a compile-time constant.

Version History

Introduced in R2012b