Lecture 4: Nonlinear Filters

Image Noise

Physical effects (dust, raindrops)

Electrical noise
Lecture 4: 

Caused by fluctuations in electric circuit.
Thermal noise

Nonlinear Filters

 Impulse noise

Quantization noise

Poor illumination
Math 490
Prof. Todd Wittman
The Citadel
Source: http://photographylife.com/what-is-iso-in-photography

Applications Noise vs. Blur

Astronomy Microscopy
Noise adds static by randomly corrupting individual pixels.

Blur smooths the edges by averaging over multiple pixels.
 =  + 
  =∗

Medical Imaging Computer Animation

Mathematically speaking, noise is generally an additive

process, while blur is generally multiplicative.

Artificial Noise Additive Noise

Gaussian Noise (MRI) Poisson Noise (PET)

We can add artificial noise to images using the Matlab  =  + (0, )  = ()
imnoise command.

We have to specify what type of noise we want and how 50 50

much.
100 100

The most common type of noise
is Gaussian (white noise). 150 150

200 200

A = imread('cameraman.tif');
B = imnoise(A, 'gaussian', 0, 0.3); 250
50 100 150 200 250
250
50 100 150 200 250

A=imread('cameraman.tif'); A=imread('cameraman.tif');
Variance B=imnoise(uint8(A),'gaussian',0,0.02); B=imnoise(uint8(A),'poisson');
Distribution of noise Mean imagesc(B); imagesc(B);

1
 Lecture 4: Nonlinear Filters

Non-additive Noise Filters

Salt & Pepper Noise Speckle Noise (Ultrasound)
(Reset random pixels)  =  ( 1 +  0,  )
A filter is a process that removes or enhances some
feature of an image.

Commonly, the word “filter” describes an operation on
50 50
the neighborhood of an image.
100 100

110

50
150 150 111

100 112

200 200 113

150
114

250 250 200

115
50 100 150 200 250 50 100 150 200 250
250 116
A=imread('cameraman.tif'); A=imread('cameraman.tif'); 50 100 150 200 250 93 94 95 96 97 98 99 100

B=imnoise(uint8(A),'salt & pepper',0.1); B=imnoise(uint8(A),'speckle',0.05);

imagesc(B); imagesc(B);

Linear vs. Nonlinear Median Filter

The median filter computes the median of the

A linear filter can be expressed as a convolution neighborhood centered over the pixel.
(weighted average over a neighborhood):
Compared to the mean filter, the median filter
=∗ preserves edges better.

Any filter that cannot be written as a convolution Mean
255 255 0 
is called a nonlinear filter. =

255 100 0 =152.8

Linear filters can be computed very quickly using
Fourier Transforms by The Convolution Median
255 255 0
Theorem. =0, 0, 0 100, 255, 255, 255, 255, 255

Nonlinear filters will generally be much slower. = 255

Median Filter Visualizing Noise

The median filter and its variants are particularly well-suited
for salt & pepper noise.
A common way to test denoising algorithms is to subtract

The Matlab command nlfilter calculates nonlinear the denoised image from the noisy image.
functions on neighborhoods. It is much slower than imfilter.
This subtracted image is called the residual.
B = nlfilter(A, [3,3], @(x) median(x(:)) );
If the algorithm did a good job denoising, then we should
only see random dots and no structure.
50 50

100 100

150 150

200

250
50 100 150 200 250
200

250
50 100 150 200 250
- =

Can also use the command medfilt2.

2
 Lecture 4: Nonlinear Filters

Thought Exercise
Quantifying Noise

A noisy brain MRI is segmented into 4 classes using

There are two major statistics used to determine the amount K-means clustering.
of noise in an image. K = kmeans(A(:),4);

Both statistics compare the noisy image f to an ideal noise- B = reshape(K, size(A));
free image fideal.
The image B takes on integer values 1-4.

Both statistics are easier to write using the Frobenius norm: 50

 = ! !  , " 
100

150

$ #
200

250

1. Root Mean Square Error (RMSE) 300

350


%& =  − $)*+,
400

where N=#pixels in image

'
450

500

2. Signal-to-Noise Ratio (SNR)

50 100 150 200 250 300 350 400 450 500

123456

Mean and median filters are not appropriate for this
% = 20 log Larger SNR = Less Noise type of data. Can you explain why?
17123456

Thought Exercise
Beyond Local

The data is qualitative. The values 1,2,3,4 are just labels

Suppose Lena’s left eye was obscured by noise.
and the order is irrelevant.

A neighborhood around the eye will not tell us the eye’s true

We could run a ___________ filter on the data. color.
C = nlfilter(B,[3,3], @(x)________(x(:)) );
50 50

100 100

150 150

200 200

250 250

300 300

350 350

400 400

450 450

500
50 100 150 200 250 300 350 400 450 500
500
50 100 150 200 250 300 350 400 450 500

Solution: Look at the other eye!

Image Similarity Nonlocal-Means

(Buades-Coll-Morel, 2005)

Natural images tend to have similar neighborhoods. proposed a denoising method that
averages over all pixels in the
image that have a similar
neighborhood.

They call their method Nonlocal

(NL) Means.
∑< 9, ; (;)
8 9 =
∑< (9, ;)
where w is a weight that judges the
similarity between the
neighborhoods.
Similar ⟶ w=1, Different ⟶ w=0

3
 Lecture 4: Nonlinear Filters

Nonlocal-Means Weights from Clean Images

Let N(x) denote the neighborhood around pixel x.

We want w(x,q)=1 if the neighborhoods N(x) and N(q)
are similar.

And we want w(x,q)=0 if the neighborhoods N(x) and
N(q) are completely different.


 9 − ;
9, ; = >9? −
ℎ

The parameter h controls the similarity.

Generally choose square neighborhood, e.g. 5x5.

Can add Gaussian weights to norm to emphasize center.

Weights from Noisy Images NL-Means

Smaller neighborhoods remove noise
better, but may over-smooth the image.

NL-Means NL-Means

Many consider NL-Means and its variants to be
NL-Means performs best on images with
the best denoising method for natural images. repeating patterns.

Noisy Gauss Filtered NL-Means

Noisy NL-Means

4
 Lecture 4: Nonlinear Filters

Implementation Color Images

NL-Means is painfully slow. There are 2 tricks to speed
We can extend NL-Means to color images by comparing
it up: 3D neighborhood cubes.

NL-Means actually achieves better results on color
1. Search Somewhat Local -- Restrict the neighborhood images than on grayscale. We can shrink the
comparisons to a window around the current pixel, neighborhood size because of the extra information.
rather than the whole image.

2. Save Pixel Comparisons -- Store all pixel-to-pixel

comparisons in an efficient data structure to avoid
making unnecessary passes through the image
(Darbon-Cunha-Chan-Osher-Jensen, 2008).

Noisy NL-Means

3D NL-Means Ultrasound

Nonlocal methods are good at removing the speckle

NL-Means extends to 3D for surface denoising (Dong- noise in Ultrasound images (Sawatzky, 2011).
Ye-Osher-Dinov, 2007).

NL-Means recovers corners and flattens smooth regions
better than many denoising methods.

NL-Means is particularly good for microscopy

(Darbon-Cunha-Chan-Osher-Jensen, 2008).

Noisy Mean Curvature Smoothing NL-Means

More Neighborhoods Use with Caution

Standard NL-Means only compares neighborhoods
under translation.

Nonlocal methods are powerful denoisers, but
they have to be used carefully.

We could also consider neighborhoods under scaling
and rotation transformations (Lou-Favaro-Soatto-
Bertozzi, 2008).
Exploiting image similarity could erase small-

But this will slow down the computation. scale features that only occur in one part of the
image.

But these small features are often the most

important part in some applications. For
example, in medical imaging we look for tumors,
Noisy NL-Means Extended NL-Means hemorrhages, muscle tears, bone fractures, etc.