04 image transformations_ii

Legal Notices and Disclaimers
This presentation is for informational purposes only. INTEL MAKES NO WARRANTIES,
EXPRESS OR IMPLIED, IN THIS SUMMARY.
Intel technologies’ features and benefits depend on system configuration and may require
enabled hardware, software or service activation. Performance varies depending on system
configuration. Check with your system manufacturer or retailer or learn more at intel.com.
This sample source code is released under the Intel Sample Source Code License
Agreement.
Intel and the Intel logo are trademarks of Intel Corporation in the U.S. and/or other countries.
*Other names and brands may be claimed as the property of others.
Copyright © 2018, Intel Corporation. All rights reserved.
2

Canny Edge Detection
This algorithm performs edge detection by completing the following steps:
1. Remove noise with a 5x5 Gaussian
2. Detect horizontal and vertical edges
 The x,y gradients are first computed and then combined into four directional
derivatives
3. Suppress non-maximums (in direction of gradient)
Local maxima define edges

Canny Edge Detection Continued…
4. Apply two thresholds, high and low
Points below low are dropped
Points above high are kept
Points in between are kept if they connect to a kept point
High: Low thresholds should be between 2:1 and 3:1
Note: Direction of gradient is perpendicular to direction of edge

Canny Edge Detection
Low Threshold = 100; High Threshold = 200

Ptolemy and Complex Representation of
Planet Motions
Ptolemy and the ancients had a model of the planets:
• Deferents and epicycles
• A series of smaller circles turning on bigger circles
This modeled the observed motions of the planets fairly well

Planet Motions
But the model was not explanatory.
Why? (CLICK ME)
An infinite sum of circles on top of circles can trace ANY fairly well-behaved (continuous
and not too wiggly) path ...
This is the fundamental idea of Fourier methods.
https://www.youtube.com/watch?v=QVuU2YCwHjw

Planet Motions
The model was descriptive.
• It gave the right positions of planets.
The model was not explanatory.
• The mathematics did not reflect the mechanism of planetary motion (aka gravity).
• Doh!

Fourier Techniques
All Fourier techniques take a function as input and output another function.
They all add up to a very special set of terms that form an orthogonal basis for those fairly
well-behaved functions.
NOTE:
In math, we call this a functional / functional form
In CS, we call this a second-order function
𝑓 𝑔 𝑥 → ℎ 𝑥
𝑓 𝑔 𝑥 → ℎ(𝑦)
𝑋 𝑘 =
1
𝑛
∑
𝑛=0
𝑁−1
𝑋 𝑛
Xi = ith component
of Fourier series,
each a component of
orthogonal basis (see
next slide)
Formula for each Xi
is in later slides.

Fourier Techniques
Orthogonal? Basis?
We have experience with bases from working with data
• We can rewrite any (x,y) point in a 2D plane as (x,y)=x(1,0) + y(0,1)
This is an incredibly powerful mathematical idea that can be applied to many
sets of mathematical objects including functions!
(0,0)
(3,5)
3
0
1
5
0 1
+

How Does Fourier Work?
The bases of Fourier methods are a bit more complicated than (1,0) and (0,1).
They are functions of a special type: They are circles!
• Information needed to coordinate drawing the same circle at the same time (verbal
information)
Image adapted from: https://betterexplained.com/articles/an-interactive-guide-to-the-fourier-transform/
A = Amplitude (radius, size)
P = Phase angle (starting point)
F = Frequency (speed around the circle)A
P
F

Fourier Methods
Fourier methods rewrite a function f in terms of an infinite sum of circles.
The rewritten function is defined at a different set of inputs and produces different
outputs:
• Inputs: Frequencies (speeds of drawing the circles)
• Outputs: The radii and phases (size of circle and angle of starting point)

How Does Fourier Work?
Mathematically, the circles are encoded as either (scary, complex)
exponentials, sins, or sins and cosines.
The three are equivalent because of trigonometric identities and Euler's
formula.
Image adapted from: https://betterexplained.com/articles/an-interactive-guide-to-the-fourier-transform/
cos 𝑥 + 𝑖𝑠𝑖𝑛(𝑥)
𝑒 𝑖𝑥 = cos 𝑥 + 𝑖𝑠𝑖𝑛(𝑥)
𝑒 𝑖𝑥

Fourier Comparison with Taylor Series
Taylor series rewrites a function as a sum of derivative-polynomial terms (a different
basis):
• Function of data to function of data
𝑓 𝑥 =
𝑛=0
∞
𝑓 𝑛 (𝑎)
𝑛!
𝑥 − 𝑎 𝑛

Fourier Comparison with Taylor Series
Same but different:
• Yes it is a rewrite.
• In the Taylor series case, both the original and the rewrite are in terms of
the same inputs and outputs.
Fourier methods are more extreme: They change the domain and range of
the rewritten function!
• f(data)  f(circles)
𝑓 𝑥 =
𝑛=0
∞
𝑓 𝑛 (𝑎)
𝑛!
𝑥 − 𝑎 𝑛

FFT and DFT
We will be using one pair of Fourier methods.
Forward: Data to circle
Backward: Circle to data

FFT and DFT: Forward
Forward: Data to circle
Input fdata([0,N-1]]) 
Output fcircles([0,N-1])
The input function tells us our data values
Input has N data values
Output has N circles
The output function tells us characteristics of the circles in our rewrites

FFT and DFT: Backward
Backward: Circle to data
Input fcircles([0,N-1]]) 
Output fdata([0,N-1])
Here we are going from an input function in terms of circles to an output
function in terms of our data
Input has N circles
Output has N data values

FFT and DFT: p
You might be worried that neither of these are necessarily functions of, or around, a
circle.
What is a function of a circle?
• A function of a circle is defined for each angle of a walk around a circle [0, 2p]
radians

FFT and DFT: p
We can convert to a function of a circle if
1. Our input is discrete and finite:
Map our values from [0, N-1] to points around the circle
Divide [0, 2p] into N parts
This is what we have with images
 Images are discrete and finite
 Edges of the images wrap to make a circle
2. Our input is discrete and repeating:
Do the exact same thing because the circle wraps around itself!

Obligatory Math #1
𝑋𝑘 =
1
𝑁
𝑛=0
𝑁−1
𝑥𝑛 𝑒−𝑖2𝜋𝑘
𝑛
𝑁
Place on Circle
Data
Frequency
Average

Obligatory Math #2
Notice formulas are almost the same
Only fundamental difference is that the sign
of the exponents is opposite.
Different mathematically equivalent equations
will move the 1/N to the circle recipe or split it
as root N on both recipes.
Circle recipe Contributions from a given circle
𝑋𝑘 =
𝑛=0
𝑁−1
𝑥𝑛 𝑒−𝑖2𝜋𝑘
𝑛
𝑁
Data recipe Contributions from a given circle
𝑥𝑛 =
1
𝑁
𝑘=0
𝑁−1
𝑋𝑘 𝑒 𝑖2𝜋𝑛
𝑘
𝑁
Place on Circle
Data
Frequency
Average

Why Fourier?
Filtering
Fast convolutions
Compression
• See JPEG and MP3s!

Why Fourier?
Our rewrite in terms of circles can be hacked.
Different circles move at different speeds
• For example, we even have a term for the stationary circle: the mean
Speeds (frequencies) are related to image regions of slow and fast variation
• High-speed changes are either edges or noise
Low-speed changes are, taken together, a peculiar averaging of the data.

Why Fourier?
These give us:
High-pass filters where we keep high frequencies.
Low-pass filters where we keep low frequencies.

Frequency Based Techniques: High-Pass
Filters
High-pass filtering removes low frequencies.
In images, a high frequency is a contour.

Why Fourier?
We can take advantage of properties of Fourier transforms.
Convolutions, in circle space, are simply multiplication.
This can allow us to drastically speed up convolutions:
1. Discrete Fourier transform (DFT) the data and the kernel
2. Perform multiplication in circle space
3. Transform back to data space

Filters
Simple averaging
Gaussian edge detection
• Laplacian
• Sobel
• Scharr

How to Use Fourier Methods
NumPy and OpenCV (and others):
• It uses the fast Fourier transform to compute the discrete Fourier transform.
Gives data points from an underlying, unknown function
• There are usually some variations for dealing with different types of input.
• There are variations for whether we are going forward (to circle space) or
backwards (to data space).

How to Use Fourier Methods
There are usually some helper functions to deal with the weirdness of
implementation details.
One important detail is that of size:
 DFT algorithms are designed to work best with particular sizes
 Different implementations subdivide the problem into different sizes, so
they have different ideal input sizes
– NumPy implementation ideal size is a power of 2
– OpenCV implementation ideal size is a product of 2s, 3s, 5s
– For example, 2*2*3*5*5 = 300
For OpenCV, we can calculate with cv2.getOptimalDFTSize().

Wavelets
Wavelets are an alternative basis to represent data.
• There are many, many wavelet basis
• Circles can be exponentials or sin/cos
These are closer to one cycle of sin or cos
Haar wavelet basis

Wavelets
A data representation describes the phenomena point-by-point.
Circle space describes the data in terms of cycles.
Wavelets represent a tradeoff between the two.

Wavelets
We can tell something about both the data and the circle characteristics at the same
time.
But …
• We can’t know both exactly.
• Amazingly, this is the mathematical root of one version of the Heisenberg
Uncertainty Principle.

Wavelet Applications
Wavelets are:
• Used for denoising
• Used for compression
• May outperform DFT and discrete cosine transform (DCT) in peak detection

Wavelet Applications
Low-pass filter
• Approximation coefficient
High-pass filter
• Detail coefficient
Horizontal
Vertical
Diagonal
Approximation Coefficient (low pass) Horizontal details (high pass)
Vertical details (high pass) Diagonal details (high pass)

Discrete Wavelet Transform
Similar to the FFT and DFT, we use the discrete wavelet transform (DWT).
We use the PyWavelets* (pywt) package to do the calculations.
import pywt
cA2, (cH2, cV2, cD2), (cH1, cV1, cD1) =
pywt.wavedec2(messi, 'haar', level=2)

Discrete Wavelet Transform
Level is similar to levels in a Gaussian pyramid:
• One application of DWT results in an approximate image and details.
We can add the two to get back the original image
• Subsequent applications of DWT on the level 1 approximate image give us a level 2
approximate image and level 2 detail images.
• We can continue applying this process recursively.

Line Detection: Successive Approximation
Transforming from a curve-contour to a simpler representation (aka line simplification)
Piecewise-linear polyline
B-spline curve
If the contour is a line (with noise) we can simplify it to that line.
If not, the line simplification is used to approximate the original curve.

Hough Transforms
Propose shape from point
Count points that agree with shape
Keep shapes at points with highest counts
Can do this for any shape we can describe with a small number of parameters (such as
point-slope, angle-distance, center-radius, center-scale-angle).

Hough Transforms
LINES CIRCLES

Hough Transforms: An Example
What potential lines do we see here?

Setup  m,b grid
Count = 3 at
m = -1 and b = 4
m
b
3

Accumulator over (distances, angles)
Fill d, Q table (accumulator) with counts of agreeing pixels
0-(w,h)
# bins
0-p) radians
# bins
0 1 2 3
135
90
45
0
Angle with respect to
line, pixel, and horizontal

L1 L2 L3 L4
D = 2.83 or ~3
Q = 45
D = 2
Q = 90
D = 2
Q = 0
D = 0
Q = 135

Tally of the points that match the d,Q model
0 1 2 3
135
90
45
0
L1
L2
L4
L3
2cts
3cts
2cts
2cts
1 point votes (cts) for 4 lines (L1-L4)
is 1 vote per angle bin

Hough Circle Transform
The circle is similar to the lines
But, the vote is performed by taking the circle center and its radius
Three parameters in this case
Would be too slow because of combinatorics
Would produce a very sparse table because of relatively few matching objects
Instead, we use the gradient and combined edges
Only the connected can vote for the same circle

General Hough Transform
We can generalize this process to other shapes by using a contour instead of the
perimeter of the circle.
Track from a pixel on the contour, rotate the orientation of the contour around a circle
• Count pixels matching at angles.
• Can also have scaling factors.

Line Detection: RANdom SAmple Consensus
(RANSAC)Pair of edge elements are used to form a line hypothesis.
How many others fall on this line?
Set boundaries for minimum matchers
Can also use RANSAC for image alignment.
Note: see OpenCV fitLine:
It may use a RANSAC-like method
Is it RANSAC-like only in use of m-estimator?

Edge Thinning: Detecting Thick Edges
Detect thick edges as one edge (without doubling)
Morphological erosion
Skeletonization
Watershed
Pearling
Convert edges to contours
 findCountours()

Line Detection: Hough Transforms
For occluded lines, we need more advanced techniques like Hough transform.
• NOTE: Hough can include or exclude gradient information.
• With it:
Do not need to orient the line
Need fewer accumulators
Do not need inner loop (work has already been done)

04 image transformations_ii

More Related Content

What's hot

Similar to 04 image transformations_ii

More from ankit_ppt

Recently uploaded

04 image transformations_ii

Editor's Notes