Binary Shape Analysis
Binary Shape Analysis
Binary Shape Analysis
Last week:
{ Mathematical morphology
z Operators
z Applications
{ image cleaning (dilation, erosion, opening, closing)
{ Binary ‘region growing’ (conditional dilation)
{ Binary template matching (hit-and-miss operator)
{ Shape representations: convex hull, skeleton
This week:
z Shape description and representation
{ Contour-based (Sonka 8.2)
{ Region-based (Sonka 8.3)
Note: the following slides are from the lecture notes of Dr. Morse
(Computer Vision 1) http://morse.cs.byu.edu/650/home/index.php 1
Boundary Description and Representation
Shape Description
Shape Description
Shape Description
I Desirable properties:
I Compact representation
I Invariant to as many transformations as possible
(translation, rotation, scale, etc.)
I Useful for matching—
relatively insensitive to small variations
I Approaches:
I Describe the boundary (contour) of the object
I Describe the region occupied by the object (next lecture)
2
Boundary Description and Representation
I Desirable properties:
I Compact representation
I Invariant to as many transformations as possible
(translation, rotation, scale, etc.)
I Useful for matching—
relatively insensitive to small variations
I Approaches:
I Describe the boundary (contour) of the object
I Describe the region occupied by the object (next lecture)
Boundary Description and Representation
Global Properties
3
Boundary Description and Representation
Chain Codes
3 2 1
4 0
5 6 7
Chain 0066002200666446464644222222
Differential 6060202060600602626260600000
Shape No. 0000060602020606006026262606
Data Structures for Image Analysis
Simple Structures
3 2 1
4 0
5 6 7
Chain 0066002200666446464644222222
Differential 6060202060600602626260600000
Shape No. 0000060602020606006026262606
Boundary Description and Representation
Chain Codes
3 2 1
4 0
5 6 7
Chain 0066002200666446464644222222
Differential 6060202060600602626260600000
Shape No. 0000060602020606006026262606
Boundary Description and Representation
Chain Codes
I Problem:
Pixel grid and noise cause change in chain code
(and its length)
I Approach:
Smooth the shape and/or resample to some fixed number
of points (code length)
Boundary Description and Representation
Tangential Representations
!(s)
"
T(s)
!(s)
s
0
s
#"
Boundary Description and Representation
Radial Representations
Curvature Representations
I Encodes the curvature κ(s) as a function of arclength
I Differentiates the tangent vector (not quite the same as
differentiating the ψ-s curve, but close)
!(s)
!(s)
s
0
s
Boundary Description and Representation
Other Representations
Signatures
I In general, a signature is a 1-dimensional function
describing a shape (chain code, differential chain code,
ψ-s curve, r -s curve, etc.)
I Here’s another: orthogonal distance d(s) to opposing side
as a function of arclength s
Boundary Description and Representation
Statistical Representations
Chord Distribution
Convex Hulls
I Build the convex hull of the shape, look at where the shape
touches its convex hull—gives you an idea of the
“extremal” points or sections of the curve.
Boundary Description and Representation
Fourier-Based Representations
Fourier Descriptors
I Can think of the boundary pixels (xk , yk ) as a curve in the
complex plane:
s(k) = x(k) + iy(k )
Boundary Description and Representation
Fourier-Based Representations
I Separates
I low-frequency components of shape (general properties)
I high-frequency ones (detail, small perturbations)
I Can filter shape!
I Low-pass filtering the Fourier descriptor smooths the shape
Boundary Description and Representation
Fourier-Based Representations
I Separates
I low-frequency components of shape (general properties)
I high-frequency ones (detail, small perturbations)
I Can filter shape!
I Low-pass filtering the Fourier descriptor smooths the shape
Boundary Description and Representation
Fourier-Based Representations
Summary
Summary
Reading:
z ELEC 536
{ Jeong and Radke, “Reslicing axially sampled 3D
shapes using elliptic Fourier descriptors”,
Medical Image Analysis 2007
Main idea:
{ Contour-based interpolation
{ Interpolation between parallel slices from a
3D shape is necessary for reslicing and
putting into correspondence organ shapes
acquired from volumetric medical imagery
1
Rationale
2
Interpolation using Elliptic Fourier
descriptors
3
Contour Interpolation via EFD
descriptors
{ Input: 2 or more contour images
5
Disadvantages