Scribd
Upload
38 views

Summary of Lecture #2: Gaussian Filters For

The lecture covered various image filtering techniques including Gaussian filters, scale spaces, Gabor filters, Canny edge detection, and the Hough and Radon transforms. Gaussian filters can be used for smoothing, derivatives, and Laplace operators while non-linear filters enable edge-preserving smoothing. Scale spaces combine smoothing at multiple scales to detect objects and boundaries across scale. The Gabor and Canny filters were discussed for pattern and edge detection applications. The Hough and Radon transforms were also introduced for detecting lines and parameterized shapes.

Uploaded by

Ali Kassab
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
38 views

Summary of Lecture #2: Gaussian Filters For

The lecture covered various image filtering techniques including Gaussian filters, scale spaces, Gabor filters, Canny edge detection, and the Hough and Radon transforms. Gaussian filters can be used for smoothing, derivatives, and Laplace operators while non-linear filters enable edge-preserving smoothing. Scale spaces combine smoothing at multiple scales to detect objects and boundaries across scale. The Gabor and Canny filters were discussed for pattern and edge detection applications. The Hough and Radon transforms were also introduced for detecting lines and parameterized shapes.

Uploaded by

Ali Kassab
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 35

Summary of Lecture #2

● Gaussian filters for:


– Smoothing
– Derivatives
– Laplace operator
● Non-linear filters for edge-preserving smoothing
● Smoothing filters for:
– Noise reduction
– Image simplification
– Shading correction
– Edge sharpening
● First order derivatives used for edge detection
Today's Lecture
● Filtering can be used for analysis & detection
● Special Applications of Filtering:
– Scale spaces
– Gabor filter (for pattern detection)
– Canny's edge detector
– Hough Transform (for line detection)
– Radon Transform
Scale Spaces
● All filters had a scale parameter
● Choosing the right scale is problem-dependent
– Often leads to “magic numbers”
● Smoothing simplifies image for better detecting objects
● Smoothing also shifts object boundaries
● Combine smoothing at many scales to detect objects
and its boundaries
● Add dimension to image: n-D + scale = (n+1)-D
Scale Space
For Gaussian filter,
the “scale” axis
represents time in
the linear diffusion
equation

scale

Non-linear scale
spaces:
● non-linear diffusion

● closings or openings

● wavelet transforms
y
x
Scale Pyramid

scale
Because the filter
reduces the
information content
of the image, it can
be downsampled
with increasing
scale to reduce
memory usage y
x
Gaussian Scale Space

scale space

local maxima & minima

inflection points
Gabor Filter
● We want to show where in the image are grid patterns
with a chosen period and orientation
– Select an area in the Fourier Domain
● We want the resulting spatial filter to be small
– Use a Gaussian region in the FD
● We want the filter to be insensitive to phase
– Use an even and an odd filter
Gabor Filter
spatial domain frequency domain

Gaussian

cosine

sine

Gabor filter
Gabor Filter
spatial domain frequency domain

Gaussian

Gaussian,
shifted in the FD

real
component

imaginary
component
Gabor Filter Example
Gabor Filter Example
Gabor Filter

smaller σ ⇒ less specific

lower ω ⇒ needs larger σ


Edge Detection
●
1st order derivatives measure gradient
● Gradient is strong at edges
● Maximum of gradient magnitude, in direction of
gradient, gives location of edge
– Canny's edge detector
● Zero crossing of second derivatives (Laplace) gives
location of edge
– Difficult to determine which zero-crossings are relevant
● By smoothing the image, we choose the scale of the
edges
Canny's Edge Detector
● Compute gradient magnitude and gradient direction
● Find pixels in the gradient magnitude that are local
maxima in the gradient direction (ridges of gradient
magnitude)
● Many pixels are part of a ridge!
● Prune ridges using hysteresis threshold:
– keep any ridge line that has a large gradient magnitude in
some section
– a normal threshold would break up edges if a small portion
were to have less contrast
Canny's Edge Detector

gradient magnitude gradient direction


Canny's Edge Detector

Non-maxima suppression
Non-Maxima Suppression

a1 & a2 computed by
linear interpolation
∇f a1

f(p)
a2 For each point p check:
f(p) > a1 ∧ f(p) ≥ a2
f(p) ≥ a1 ∧ f(p) > a2
If condition not met, set
pixel to 0.
Canny's Edge Detector
low threshold

hysteresis threshold high threshold


Canny's Edge Detector
Influence of the σ
σ=1 σ=2

σ=4 σ=8
Hough Transform
● Detecting parametrized shapes in an image
● Originally for straight lines in 2-D (2 parameters: p, φ)
● Later generalized to any shape:
– 3-D sphere (4 parameters: x, y, z, r)
– 2-D square (4 parameters: x, y, a, φ)
– 2-D smiley (3 parameters: x, y, r)
– …
● Creates a “parameter space”
● Converts the shape detection into maxima detection
2D Hough Transform for Lines
● Each point in the parameter shape represents one
instance of the shape in the image
p

φ
φ
p

image parameter space


2D Hough Transform for Lines
● Each point in the image gives evidence for all lines that
go through that point
p

image parameter space


2D Hough Transform for Lines
● Each point in the image gives evidence for all lines that
go through that point
p

image parameter space


Hough Transform Example
2D Hough Transform for Lines
● Each local maximum defines a line that goes across the
full image
● We need to detect the line segment in the image
● For each local maximum:
– Look in the input image where there is support for this line
– Fill small gaps in the line
– Is the resulting line long enough?
● In the MATLAB Image Processing Toolbox:
– hough: create the Hough transform
– houghpeaks: detect local maxima
– houghlines: find the line segments
Hough Transform Example
Improving The Hough Transform
● Use additional knowledge about the points in the image,
e.g. local gradient information
p

image parameter space


The Radon Transform
● For each point in the parameter space, examine how
much evidence there is in the image
p

φ
φ
p

image parameter space


The Radon Transform
● For each point in the parameter space, examine how
much evidence there is in the image
● This can be done though one convolution for all values
of the “position” or “distance”
parameters
● This is identical to template
matching!
Radon Transform Example

y
r = 25

r = 30
The Radon Transform
● Because I’m using convolution to create the parameter
space, I can use a grey-value convolution mask!
● Define the circle with a Gaussian profile:
– Yields a band-limited parameter space
– The parameter space can be sampled without aliasing
● It is possible to detect peaks with sub-pixel accuracy
Radon Transform Example
x y r
55.10 51.52 24.94
109.00 88.28 24.90
235.10 181.78 24.92
264.83 107.11 24.86
147.59 35.42 29.06
215.97 73.66 29.21
36.06 111.39 29.93
95.17 152.87 30.06
119.33 218.96 30.42
173.80 125.39 29.62
Hough vs Radon
● The resulting parameter space is identical

● Hough = write paradigm


● Radon = read paradigm

● Hough = efficient for binary input images


● Radon = efficient for grey-value input images

● Radon makes it easier to define the parameter space,


and how to sample it
Summary of Today’s Lecture
● Special Applications of Filtering:

– Scale spaces

– Pattern detection: Gabor filter

– Edge detection: Canny's edge detector

– Hough Transform
– Radon Transform

You might also like