regionprops, tables, and struct2table
Today I was talking to Spandan (a regular guest blogger here) about the history of the Image Processing Toolbox function regionprops. This function is a widely-used work-horse of the toolbox. Many of my blog posts here have shown it in practice.
The function regionprops computes properties of regions of pixels. Most often, these regions correspond to individual "objects" (connected components of foreground pixels) of a binary image. Here's an example.
bw = imread('text.png');
imshow(bw)
s = regionprops(bw)
s =
88x1 struct array with fields:
Area
Centroid
BoundingBox
The output is an 88x1 struct array, elements of which contain measurements for objects in the binary image. (We didn't tell regionprops what measurements to compute, so it computed Area, Centroid, and BoundingBox by default.) Here are the measurements for the 10th object.
s(10)
ans =
Area: 61
Centroid: [54 39]
BoundingBox: [49.5000 33.5000 9 11]
This form of output is convenient in some ways but inconvenient or awkward in others. For example, it requires the obscure "comma-separated-list" syntax to get all the area measurements into a single vector to call a function such as mean, min, or max.
a = [s.Area]; mean(a)
ans = 54.2386
min(a)
ans =
6
max(a)
ans = 106
In the comma-separated list syntax for struct arrays, the expression [s.Area] automatically expands into the expression [s(1).Area, s(2).Area, ..., s(end).Area]. But relatively few people recognize or understand this syntax.
Another inconvenient aspect of the struct array aspect is that it's hard to get a tabular view of all the data. As shown above, you can see the data for the 10th object by displaying s(10), but it's hard to see all the data for multiple objects. For example:
s(1:5)
ans =
5x1 struct array with fields:
Area
Centroid
BoundingBox
If I were designing regionprops from scratch today, based on the most recent version of MATLAB, I would have regionprops return a table. Tables were introduced as a new fundamental MATLAB type in R2013b. Here's a simple example to illustrate.
age = [38 43 38 40 49]'; height = [71 69 64 67 64]'; weight = [176 163 131 133 119]'; blood_pressure = [124 93; 109 77; 125 83; 117 75; 122 80]; t = table(age,height,weight,blood_pressure)
t =
age height weight blood_pressure
___ ______ ______ _______________
38 71 176 124 93
43 69 163 109 77
38 64 131 125 83
40 67 133 117 75
49 64 119 122 80
t.height is an ordinary 5-by-1 vector of heights.
t.height
ans =
71
69
64
67
64
t.blood_pressure is an ordinary 5-by-2 matrix. (Table "variables" are often column vectors, but they can have any size as long as the number of rows is the same as the height of the table.)
t.blood_pressure
ans = 124 93 109 77 125 83 117 75 122 80
We are excited about the introduction of tables in MATLAB, and we think they will gradually be used in more and more situations where previously people used struct arrays.
Such as regionprops.
So what to do?
Well, the function struct2table comes in very handy here! As you might have guessed from the name, it automatically converts a struct array into a table.
Try it:
t = struct2table(s)
t =
Area Centroid BoundingBox
____ __________________ ____________
66 11 13.5 [1x4 double]
41 7.6829 38.171 [1x4 double]
63 17.889 38.857 [1x4 double]
80 22.95 15.963 [1x4 double]
53 26.472 36.604 [1x4 double]
63 34.889 16.857 [1x4 double]
63 34.889 38.857 [1x4 double]
41 43.683 38.171 [1x4 double]
48 48.396 15.875 [1x4 double]
61 54 39 [1x4 double]
63 56.889 16.857 [1x4 double]
41 65.683 16.171 [1x4 double]
48 67.396 37.875 [1x4 double]
105 78.79 16.79 [1x4 double]
66 76.5 39 [1x4 double]
68 93.529 39.265 [1x4 double]
92 100 16.696 [1x4 double]
41 106.68 38.171 [1x4 double]
68 113.53 17.265 [1x4 double]
6 114 31.5 [1x4 double]
33 114 39 [1x4 double]
85 123.24 38.259 [1x4 double]
48 121.4 15.875 [1x4 double]
63 129.89 16.857 [1x4 double]
100 134.87 40.86 [1x4 double]
41 138.68 16.171 [1x4 double]
63 145.89 38.857 [1x4 double]
61 149 17 [1x4 double]
80 159.95 15.963 [1x4 double]
85 163.26 185.76 [1x4 double]
6 156.5 195 [1x4 double]
6 156.5 211 [1x4 double]
85 163.26 218.76 [1x4 double]
48 159.4 37.875 [1x4 double]
61 164 107 [1x4 double]
68 164.26 117.47 [1x4 double]
63 163.86 129.11 [1x4 double]
41 163.17 139.32 [1x4 double]
68 164.26 147.47 [1x4 double]
61 164 164 [1x4 double]
63 163.86 175.11 [1x4 double]
33 164 195 [1x4 double]
54 163.44 203 [1x4 double]
33 164 211 [1x4 double]
80 167.95 37.962 [1x4 double]
9 168 233 [1x4 double]
9 168 238 [1x4 double]
9 168 243 [1x4 double]
63 171.89 16.857 [1x4 double]
68 180.53 39.265 [1x4 double]
48 184.88 67.604 [1x4 double]
106 183.6 118.53 [1x4 double]
6 178.5 127 [1x4 double]
85 185.26 134.76 [1x4 double]
85 185.26 164.24 [1x4 double]
85 185.26 179.76 [1x4 double]
6 178.5 212 [1x4 double]
85 185.26 238.76 [1x4 double]
85 184.24 16.259 [1x4 double]
72 185.68 76.653 [1x4 double]
63 185.86 88.111 [1x4 double]
41 185.17 98.317 [1x4 double]
63 185.86 107.11 [1x4 double]
33 186 127 [1x4 double]
74 187.2 152.31 [1x4 double]
63 185.86 192.11 [1x4 double]
72 185.68 203.65 [1x4 double]
33 186 212 [1x4 double]
68 186.26 219.47 [1x4 double]
41 185.17 230.32 [1x4 double]
48 188.4 37.875 [1x4 double]
9 199 43 [1x4 double]
48 206.88 161.6 [1x4 double]
6 200.5 235 [1x4 double]
9 204 43 [1x4 double]
61 208 125 [1x4 double]
105 207.79 139.21 [1x4 double]
63 207.86 153.11 [1x4 double]
61 208 170 [1x4 double]
74 209.2 181.31 [1x4 double]
61 208 192 [1x4 double]
41 207.17 207.32 [1x4 double]
63 207.86 216.11 [1x4 double]
54 207.44 227 [1x4 double]
33 208 235 [1x4 double]
41 207.17 242.32 [1x4 double]
9 209 43 [1x4 double]
9 212 117 [1x4 double]
Immediately, you can see that the data is much more easily visible with a table. You get scrolling and editing if you open t in the Variable Editor.
You can plot directly with the data in a table variable:
imshow(bw) hold on plot(t.Centroid(:,1),t.Centroid(:,2),'r*') axis([.5 100.5 .5 100.5]) hold off
Logical indexing comes in handy for filtering objects based on various criteria. For example, the following code removes objects with area less than 10 from the table.
t(t.Area < 10,:) = [];
And you can easily export tables to various formats. Here is our table exported to a .CSV file and then read into Excel.
writetable(t,'regionprops.csv');
I encourage everyone who has used regionprops to play around with tables (assuming you have MATLAB R2013b) to see how they might be useful in your own work.
Comments
To leave a comment, please click here to sign in to your MathWorks Account or create a new one.