.

Spatial filtering

Stefano Ferrari

Università degli Studi di Milano

[email protected]

Methods for Image Processing

academic year 2022–2023

Filtering

I The term filtering refers to a technique developed for the

frequency domain.
I Rapid variations of intensity can be associated to high
frequency components, while slow changes can be associated
to low frequency components.
I The effects of the filtering process is the attenuation or the
enhancement of the components in a target interval of
frequency.

Stefano Ferrari— Methods for Image processing— a.y. 2022/23 1

.

Spatial filtering

I Similar effects can be obtained also in the spatial domain:

techniques of this kind are defined as spatial filtering.
I In general, the spatial filtering techniques operate on an image
taking into consideration the intensity values in an suitable
neighborhood of each pixel.
I For each pixel of the original image, the intensity of the
corresponding pixel in the “filtered” image is computed.
I The transformation rule is often described by a matrix, called
filter, with the same size of the considered neighborhood.
I If the transforming rule is a linear function of the intensities in
the considered neighborhood, the technique is called spatial
linear filtering (non-linear, otherwise).

Linear filtering

Stefano Ferrari— Methods for Image processing— a.y. 2022/23 2

.

Linear filtering (2)

I The pixel (x, y ) in the filtered image, g , is obtained as the

weighted average of the pixels in the original image, f , of a
suitable neighborhood of (x, y ):
a
X b
X
g (x, y ) = w (s, t) f (x + s, y + t)
s=−a t=−b

I The weight matrix, w , is called filter.

I Or mask, template, window.
I For convenience, often a matrix with an odd number of rows
2a + 1, and columns, 2b + 1, is used.

Correlation and convolution

I Correlation
a
X b
X
g (x, y ) = w (x, y )?f (x, y ) = w (s, t) f (x +s, y +t)
s=−a t=−b

I Convolution
a
X b
X
g (x, y ) = w (x, y )∗f (x, y ) = w (s, t) f (x −s, y −t)
s=−a t=−b

I Vectorial representation
mn
X
R = w1 z1 + · · · + wmn zmn = wk zk = w T z
k=1

Stefano Ferrari— Methods for Image processing— a.y. 2022/23 3

.

Correlation

f w
0 0 0 1 0 0 0 0 1 2 3 2 8

↓
f 0 0 0 1 0 0 0 0
w 1 2 3 2 8
↑ starting position alignment

zero padding
f 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0
w 1 2 3 2 8

Correlation (2)

f 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0
w 1 2 3 2 8
w ?f 0

f 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0
w 1 2 3 2 8
w ?f 0 0

f 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0
w 1 2 3 2 8
w ?f 0 0 0

f 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0
w 1 2 3 2 8
w ?f 0 0 0 8

Stefano Ferrari— Methods for Image processing— a.y. 2022/23 4

.

Correlation (3)

f 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0
w 1 2 3 2 8
w ?f 0 0 0 8

f 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0
w 1 2 3 2 8
w ?f 0 0 0 8 2

f 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0
w 1 2 3 2 8
w ?f 0 0 0 8 2 3

f 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0
w 1 2 3 2 8
w ?f 0 0 0 8 2 3 2

Correlation (4)

f 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0
w 1 2 3 2 8
w ?f 0 0 0 8 2 3 2 1

...
f 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0
w 1 2 3 2 8
w ?f 0 0 0 8 2 3 2 1 0 0 0 0

w ?f 0 0 0 8 2 3 2 1 0 0 0 0
Cropping 0 8 2 3 2 1 0 0
f 0 0 0 1 0 0 0 0

Stefano Ferrari— Methods for Image processing— a.y. 2022/23 5

.

Convolution

f w rotated 180◦
0 0 0 1 0 0 0 0 8 2 3 2 1

↓
f 0 0 0 1 0 0 0 0
w 8 2 3 2 1
↑ starting position alignment

...
w ∗f 0 0 0 1 2 3 2 8 0 0 0 0
Cropping 0 1 2 3 2 8 0 0
f 0 0 0 1 0 0 0 0

Correlation and convolution on images (1)

Stefano Ferrari— Methods for Image processing— a.y. 2022/23 6

.

Correlation and convolution on images (2)

Vectorial representation

I When relative position of the coefficients is not important, it

is more convenient to represent the effect of the filter using a
vectorial representation.
I A conventional indexing have to be specified.

mn
X
R = w1 z1 + · · · + wmn zmn = wk zk = w T z
k=1

Stefano Ferrari— Methods for Image processing— a.y. 2022/23 7

.

Filter specification

I The filter coefficients have to be specified.

I Their values depend on the desired effect.
I Direct specification:
9
1X 1
R= zi ⇒ wi = , ∀i
9 9
i=1

I Function based specification:

x 2 +y 2
h(x, y ) = e − 2σ 2 ⇒ w (s, t) = h(s, t)
I Algorithmic specification
I Non linear filters
I e.g., “max” filter

Smoothing filters

I Results in defocused images (blur).

I Small details (smaller that the size of the filter) removal:
I more attention on large objects;
I small gaps bridging;
I noise reduction.
I Depending on the type of noise that affects the image, the
filter can be linear or non-linear.

Stefano Ferrari— Methods for Image processing— a.y. 2022/23 8

.

Linear smoothing filters

I Averaging filters (or low pass)

I Arithmetic average
I Weighted average

Smoothing effects (1)

Original image and image filtered using a 3 × 3 averaging filter.

Stefano Ferrari— Methods for Image processing— a.y. 2022/23 9

.

Smoothing effects (2)

Filtering with averaging filters (5 × 5 and 9 × 9).

Smoothing effects (3)

Filtering with averaging filters (15 × 15 and 35 × 35).

Stefano Ferrari— Methods for Image processing— a.y. 2022/23 10

.

Example: details removal

f g =f ∗h m=g >θ

a b c (a) f , image 353 × 382

d (b) g , image filtered with
13 × 13 smoothing filter
(c) m, thresholded image
(d) k, masked image:
k(x, y ) = f (x, y ) × m(x, y )
k

Non-linear filters

I Statistics-based filters are the most used non-linear filters.

I Ranking (which requires sorting) is more computationally
expensive than linear operations.
I In particular:
I Median filter
I Max and Min filters
I Percentiles based filters
I Example: f = [100, 120, 98, 99, 110, 255, 100, 200, 10]
I mean: 121
I median: 100
I sorted: [10, 98, 99, 100, 100, 110, 120, 200, 255]
I min: 10
I max: 255

Stefano Ferrari— Methods for Image processing— a.y. 2022/23 11

.

Example: median filter

Salt-and-pepper noise removal

a b c

(a) Original image, corrupted by salt-and-pepper noise.

(b) Image filtered with averaging filter 3 × 3.
(c) Image filtered with median filter 3 × 3.

Homeworks and suggested readings

DIP 3e, Sects. 3.4, 3.5 DIP 4e, Sects. 3.4, 3.5
I pp. 144–156 I pp. 177–198

GIMP
I Filters
I Blur
I Enhance
I Despeckle
I Generic
I Convolution matrix

Stefano Ferrari— Methods for Image processing— a.y. 2022/23 12