Image processing and analysis with ImageJ – Exercises - Topic 06 – Noise and Filter

Topic 06 – Noise and filter

Open the image plant-noise.tif. The image contains a high level of noise. Zoom into the image. The
background should be homogeneous, but it contains a random distribution of intensities. The same
random distribution is added to the signal.

Excercise 6.1 – Noise

The noise can pose a problem for the analysis of the image, for example in the separation of the
objects from the background. e might have to pre-process our image to reduce the disturbing effect
of noise.
a) Open the image plant-noise.tif from the folder 06 - noise and filter. Try to use the
Threshold-Adjuster to separate the plant from the background. Is this possible?

Exercise 6.2 – Convolution Filter – Mean Filter

An evident way to smooth our image and to reduce the impact of the noise is to replace each pixel
in the image by an average of the intensities in its neighborhood. This is called a mean filter in
ImageJ. You can find it under Process>Filters>Mean.
a) Apply the mean-filter to the plant-noise image. Compare the histogram before and after
application of the filter. What radius do you need to use, in order to be able to separate the
plant from the background using the threshold-adjuster?
b) Imagine that your image is a white rectangle (intensity 255). How will the result of applying
a mean-filter with radius 1 look like? Control your prediction by applying the filter to the
image rectangle.tif.
c) Use the rectangle-image to compare the results of the command Process>Smooth with the
results from the mean-filter with radius 1.
d) Use the convolver (Process>Filter>Convolve) to define a mean-filter. How does the kernel
look like? On the plant-noise image, is the result of the convolver exactly the same as the
result of the mean-filter command?

Exercise 6.3 - Convolution Filter – Gaussian Blur Filter

A Gaussian blur filter is similar to the mean filter, but instead of using uniform weights in the
kernel, the weights are the values of a normal distribution, also called Gaussian distribution.
a) Open the file Gaussian.txt. Here you see the values of a normal distribution. How does it
look like as an image? Open the file as an image by using File>Import>Text Image. Apply
the ice lut. Now go to Analyze>Surface Plot or use the Interactive 3D surface plot plugin
(Plugins>3D) to get an impression of the 3D form.
b) Apply the Gaussian blur filter with sigma 1.2 from Process>Filters>Gaussian blur to the
rectangle image and compare the result with the result of the mean filter with radius 3.
c) Apply a Gaussian blur filter with a sigma of 6 and a Mean filter of radius 15 to the image
roots.tif and compare the results.

[email protected] 1/5
 Image processing and analysis with ImageJ – Exercises - Topic 06 – Noise and Filter

d) Open the plant-noise image. Open the convolution tool and load the Gaussian.txt into it.
Run the convolution filter on the image.

Exercise 6.4 – Edge enhancing convolution filter - Sobel and Prewitt filter
Other kernels can be used to create edge enhancing filters.
a) The Prewitt filter is defined by the kernels:
−1 0 1 1 1 1
k x =−1 0 1 and k y = 0 0 0 . If Dx and Dy are the results of the convolution
−1 0 1 −1 −1 −1
of kx and ky with the input image then you have to calculate √ D2x +D 2y to apply the Prewitt
filter. Apply the Prewitt filter to the image rectangle.tif. Create a composite image of the
input image and the result (Image>Color>Merge Channels) Hints:
• convert the image to 32-bit first, make a copy and use the convolution-tool with the two
kernels
• you find the square and square-root operations under Process>Math
• Use the image calculator (Process>Image Calculator...) to add two images
b) Apply the Prewitt filter to the image roots.tif. Don't forget to convert it to 32-bits. Create an
overlay of the input image and the filtered image.
c) Another edge enhancing filter is the Sobel filter. It works in the same way but with the
kernels:
−1 0 1 1 2 1
k x =−2 0 2 and k y = 0 0 0 . Apply the Sobel-filter to the image
−1 0 1 −1 −2 −1
rectangle.tif and compare the result with the result of the command Process>Find Edges
d) High-pass or Uni Crisp filter can be used to apply a sharpening effect to an image. Convolve
−1 −1 −1
the image ophrys.tif with the kernel −1 12 −1 and compare the result with the result
−1 −1 −1
of the command Process>Sharpen.
e) A Laplacian of Gaussian (LoG) filter also called Mexican Hat Filter applies a gaussian
smoothing filter and an edge enhancing filter in one step. It can be used to detect spots.
Possible discrete approximations for the laplacian kernel are:
0 −1 0 −1 −1 −1
−1 4 −1 or −1 8 −1 . You can apply a LoG filter by first applying a
0 −1 0 −1 −1 −1
gaussian filter and then convolving with one of the above kernels. However the gaussian and
laplacian can be combined in one single kernel. Install the LoG plugin from
http://bigwww.epfl.ch/sage/soft/LoG3D/. Try to threshold the bright spots on the image.
Run the LoG plugin on the image. While applying it create the kernel as an image and have
a look at it. Try thresholding the spots after application of the filter.

[email protected] 2/5
 Image processing and analysis with ImageJ – Exercises - Topic 06 – Noise and Filter

Exercise 6.4 – Ranking filter – Median filter

Rank filters are another class of filters. Again we look at a neighborhood of radius r for each pixel.
But this time the intensity values are sorted and the pixel is replaced with the value in a specific
position. If we use the middle position we get a median filter.
a) Open the image plant-sp-noise.tif. Zoom in to the image. You see that in this case the noise
consists of the addition of white and black pixels to the image. This kind of noise is called
salt-and-pepper noise. Apply a median filter of radius 3 from Process>Filters>Median to
the image. Compare the result with the result from the mean filter.
b) Compare the effect of the median filter with radius 1 and the command
Process>Noise>Despeckle.
c) In Process>Noise>Remove Outliers, you find a selective median filter, that replaces a pixel
by the median in the neighborhood if the value of the pixel differs more than a threshold
value from the median. Use it to remove the dark and light components of the salt and
pepper noise separately.
d) In Process>Noise>Remove Outliers, you find a selective median filter, that replaces a pixel
by the median in the neighborhood if the value of the pixel differs more than a threshold
value from the median. Use it to remove the dark and light components of the salt and
pepper noise separately.
e) The signal-noise ratio is a measure of how disturbing the noise is. It can be calculated in the
following way:
IntDen( I² )
2 where I is the ideal image not affected by noise, M the measured,
IntDen[(I −M ) ]
noisy image and IntDen the integrated intensity, i.e. the sum of all intensity values in the
image. With the help of the macro recorder, create a macro that calculates the snr for the
example image boats from the menu File>Open Samples. Create the measured image M by
adding specified noise with sigma = 15 to the input image. Use the macro to calculate the
snr after applying a gaussian-blur filter with sigma 1.2, a mean filter with radius 3 and a
median filter with radius 3. Compare the results. Do it with other example images, as well.

Exercise 6.5 – Ranking filter – Min and Max filter – Grey-morphology

Taking the minimum or maximum value instead of the median gives the min- and max-filter. The
min filter increases dark regions, the max filter bright regions. We can use them to define a
complete grey-morphology, with the only restriction that the structuring element is fixed to a square
of radius r. Some of the operations are:
 min – erode – enlarge dark regions
 max – dilate – enlarge bright regions
 open – dilate(erode(n)) – remove light features smaller than r
 close – erode(dilate(n)) – remove dark features smaller than r
 sequential filter – open(close(n)) – remove bright and dark features smaller than r
 white top hat – I-opening(I) – extract bright features

[email protected] 3/5
 Image processing and analysis with ImageJ – Exercises - Topic 06 – Noise and Filter

 black top hat – closing(I)-I – extract dark features

 …
a) Apply the min and max filter to the image Lena.tif try the values 1, 10 and 20 for the radius.
b) Create a toolset simple grey morphology that has a button for each of the operations erode,
dilate, open, close, sequential filter, white top hat and black top hat. A right click on the first
button should open a dialog in which the user can set the radius that will be used by all the
operations. Run the operations on the image lena.tif and clown.jpg with different values for
the radius. Hints:
• Use the macro editor to write a new toolset (Plugins>New>Macro).
• You can open an existing toolset to have a look at an example by holding down shift
and switching to the toolset with the >> button.
• A macro has the form
macro "Erode Action Tool- C000T4b12e" {
…
}
The string C000T4b12e defines the image on the button. In this case it draws the
letter e in black.
• Declare a variable for the radius of the min and max filter at the beginning
var radius = 1;
To add an option that allows to set the radius to a button you can use for example:
macro 'Erode Action Tool Options' {
radius = getNumber("radius:",radius);
}
• You can declare functions that you can call from a macro:
function erode() {
run("Minimum...", "radius=" + radius);
}
• You can use the macro-recorder if you don't know how to write a specific command
in the macro language.

Exercise 6.6 – Filtering in the Frequency Domain

The Fourier transform changes the coordinate system from the spacial coordinates to the
coordinates angle and frequency. The basis is that any signal can be approximated by a sum of
harmonic functions (sin and cos) of different amplitude and phase. The result is complex numbered,
so we really transform one spacial image into two result images: the phase image (imaginary part)
and the powerspectrum (real part). The part the more interesting is the powerspectrum.
According to the convolution theorem, a point-wise multiplication in the frequency domain
corresponds to a convolution in the spacial-domain. Besides the transform and the inverse transform

[email protected] 4/5
 Image processing and analysis with ImageJ – Exercises - Topic 06 – Noise and Filter

can efficiently be calculated. So we can use it to more efficiently apply convolution filters that do
not have a simple kernel in the spacial-domain.
Possible applications are:
• low-pass filter – smoothing – noise suppression
• high-pass filter – edge detection
• band-pass filter – detect structures with a given frequency
• correlation, template matching, stitching
• deconvolution (inverse filtering) (see 4.5 d) and e))
ImageJ allows to apply a Fourier transform under Process>FFT>FFT. As a result the
powerspectrum is displayed. This is a special image on which you can define a filter by making
regions white or black. Frequencies under the white mask will be passed, those under the black
mask will be suppressed when you calculate the inverse transformation (Process>FFT>Inverse
FFT).
a) Open the image pierre.tif. Calculate the Fourier transform. Do you see the two clusters on in
the upper left half and one in the lower right? Draw a white mask on them and calculate the
inverse transform. Is there an object in the image that corresponds to the frequencies that
passed?
b) Use the Fourier transform to suppress the noise in the image plant-noise.tif. Which
frequencies do you have to suppress?
c) Use the Fourier transform to detect the edges in the image ophrys.tif. Which frequencies do
you have to suppress?
d) The image lines.tif consists of vertical white lines that have a width of one pixel and a
distance of 4 pixel. How will the powerspectrum of the Fourier transform of the image look
like? Verify your answer by creating it.
e) Load the image text.tif and the image template.tif. Use the command correlate from
Process>FFT>FD Math to calculate the correlation between the text and the template.
Transform the resulting (spacial) image to 8-bit, move the minimum threshold value so the
right until only single points remain and apply the threshold. Now use Process>Find
Maxima to create a point selection from the points and transfer the point selection onto the
text image, to see which points gave the maximum response to the template.

[email protected] 5/5