1
Image Pyramids {#tutorial_pyramids}
7
In this tutorial you will learn how to:
9
- Use the OpenCV functions @ref cv::pyrUp and @ref cv::pyrDown to downsample or upsample a given
15
@note The explanation below belongs to the book **Learning OpenCV** by Bradski and Kaehler.
17
- Usually we need to convert an image to a size different than its original. For this, there are
19
-# *Upsize* the image (zoom in) or
20
-# *Downsize* it (zoom out).
21
- Although there is a *geometric transformation* function in OpenCV that -literally- resize an
22
image (@ref cv::resize , which we will show in a future tutorial), in this section we analyze
23
first the use of **Image Pyramids**, which are widely applied in a huge range of vision
28
- An image pyramid is a collection of images - all arising from a single original image - that are
29
successively downsampled until some desired stopping point is reached.
30
- There are two common kinds of image pyramids:
31
- **Gaussian pyramid:** Used to downsample images
32
- **Laplacian pyramid:** Used to reconstruct an upsampled image from an image lower in the
33
pyramid (with less resolution)
34
- In this tutorial we'll use the *Gaussian pyramid*.
38
- Imagine the pyramid as a set of layers in which the higher the layer, the smaller the size.
40
![](images/Pyramids_Tutorial_Pyramid_Theory.png)
42
- Every layer is numbered from bottom to top, so layer \f$(i+1)\f$ (denoted as \f$G_{i+1}\f$ is smaller
43
than layer \f$i\f$ (\f$G_{i}\f$).
44
- To produce layer \f$(i+1)\f$ in the Gaussian pyramid, we do the following:
45
- Convolve \f$G_{i}\f$ with a Gaussian kernel:
47
\f[\frac{1}{16} \begin{bmatrix} 1 & 4 & 6 & 4 & 1 \\ 4 & 16 & 24 & 16 & 4 \\ 6 & 24 & 36 & 24 & 6 \\ 4 & 16 & 24 & 16 & 4 \\ 1 & 4 & 6 & 4 & 1 \end{bmatrix}\f]
49
- Remove every even-numbered row and column.
51
- You can easily notice that the resulting image will be exactly one-quarter the area of its
52
predecessor. Iterating this process on the input image \f$G_{0}\f$ (original image) produces the
54
- The procedure above was useful to downsample an image. What if we want to make it bigger?:
55
columns filled with zeros (\f$0\f$)
56
- First, upsize the image to twice the original in each dimension, wit the new even rows and
57
- Perform a convolution with the same kernel shown above (multiplied by 4) to approximate the
58
values of the "missing pixels"
59
- These two procedures (downsampling and upsampling as explained above) are implemented by the
60
OpenCV functions @ref cv::pyrUp and @ref cv::pyrDown , as we will see in an example with the
63
@note When we reduce the size of an image, we are actually *losing* information of the image.
68
This tutorial code's is shown lines below. You can also download it from
69
[here](https://github.com/Itseez/opencv/tree/master/samples/cpp/tutorial_code/ImgProc/Pyramids.cpp)
71
@include samples/cpp/tutorial_code/ImgProc/Pyramids.cpp
76
Let's check the general structure of the program:
78
- Load an image (in this case it is defined in the program, the user does not have to enter it
81
/// Test image - Make sure it s divisible by 2^{n}
82
src = imread( "../images/chicky_512.jpg" );
84
{ printf(" No data! -- Exiting the program \n");
88
- Create a Mat object to store the result of the operations (*dst*) and one to save temporal
97
- Create a window to display the result
99
namedWindow( window_name, WINDOW_AUTOSIZE );
100
imshow( window_name, dst );
103
- Perform an infinite loop waiting for user input.
113
{ pyrUp( tmp, dst, Size( tmp.cols*2, tmp.rows*2 ) );
114
printf( "** Zoom In: Image x 2 \n" );
116
else if( (char)c == 'd' )
117
{ pyrDown( tmp, dst, Size( tmp.cols/2, tmp.rows/2 ) );
118
printf( "** Zoom Out: Image / 2 \n" );
121
imshow( window_name, dst );
125
Our program exits if the user presses *ESC*. Besides, it has two options:
127
- **Perform upsampling (after pressing 'u')**
129
pyrUp( tmp, dst, Size( tmp.cols*2, tmp.rows*2 )
131
We use the function @ref cv::pyrUp with 03 arguments:
133
- *tmp*: The current image, it is initialized with the *src* original image.
134
- *dst*: The destination image (to be shown on screen, supposedly the double of the
136
- *Size( tmp.cols*2, tmp.rows\*2 )\* : The destination size. Since we are upsampling,
137
@ref cv::pyrUp expects a size double than the input image (in this case *tmp*).
138
- **Perform downsampling (after pressing 'd')**
140
pyrDown( tmp, dst, Size( tmp.cols/2, tmp.rows/2 )
142
Similarly as with @ref cv::pyrUp , we use the function @ref cv::pyrDown with 03
145
- *tmp*: The current image, it is initialized with the *src* original image.
146
- *dst*: The destination image (to be shown on screen, supposedly half the input
148
- *Size( tmp.cols/2, tmp.rows/2 )* : The destination size. Since we are upsampling,
149
@ref cv::pyrDown expects half the size the input image (in this case *tmp*).
150
- Notice that it is important that the input image can be divided by a factor of two (in
151
both dimensions). Otherwise, an error will be shown.
152
- Finally, we update the input image **tmp** with the current image displayed, so the
153
subsequent operations are performed on it.
161
- After compiling the code above we can test it. The program calls an image **chicky_512.jpg**
162
that comes in the *tutorial_code/image* folder. Notice that this image is \f$512 \times 512\f$,
163
hence a downsample won't generate any error (\f$512 = 2^{9}\f$). The original image is shown below:
165
![](images/Pyramids_Tutorial_Original_Image.jpg)
167
- First we apply two successive @ref cv::pyrDown operations by pressing 'd'. Our output is:
169
![](images/Pyramids_Tutorial_PyrDown_Result.jpg)
171
- Note that we should have lost some resolution due to the fact that we are diminishing the size
172
of the image. This is evident after we apply @ref cv::pyrUp twice (by pressing 'u'). Our output
175
![](images/Pyramids_Tutorial_PyrUp_Result.jpg)