Machine vision

coatiarfAI and Robotics

Oct 17, 2013 (3 years and 5 months ago)

57 views

1
Copyright (c) Benny Thörnberg
1:39
Machine vision
Introduction

W
V
2
Copyright (c) Benny Thörnberg
3:39
Outline
•Research at Miun
•Pulp Quality Measurement
•Wood chip analyzer
•Paper surface analyzer
•Checkpoint questions
Copyright (c) Benny Thörnberg
4:39
Introduction – A typical machine vision system
Hole Radius OK
1 3.01 Y
2 2.99 Y
3 3.02 Y
4 2.87 N
Reject
3
Copyright (c) Benny Thörnberg
5:39
Introduction – Benefits and
Drawbacks
+ Higher speed compared to manual measurements
+ Every unit can be checked
+ Reduced cost compared to manual checking
+ Contactless measurement
- Initial investments
- Slower switching of production compared to
manual inspection
Copyright (c) Benny Thörnberg
6:39
Introduction – Another machine vision
system?
4
Copyright (c) Benny Thörnberg
7:39
Introduction – What more is possible?
•Measure mean value speed
•Vision systems working together in networks
•Measure traffic flow and visibility
•Report dangerous behavior
•Inquiries for specified vehicles
•Pay road systems
•Control of road illumination
•Dynamic speed limits
• …
Copyright (c) Benny Thörnberg
8:39
Introduction – A basic task for a traffic camera
5
Copyright (c) Benny Thörnberg
9:39
Fundamentals – Steps of a machine vision
system
Image acquisition
Preprocessing
Segmentation
Feature extraction
Classification
Higher data abstraction
Higher data intensity
Labeling
Copyright (c) Benny Thörnberg
10:39
Fundamentals – Image acquisition
The reflected illuminated light is projected in 2D through a lens
onto a pixel sensor. The acquired image is a 2D representation of
the observed 3D surface where each pixel contain information
about reflected intensity (and chromatography).
The most common setup for 2D image acquisition
6
Copyright (c) Benny Thörnberg
11:39
Fundamentals – Preprocessing
It is a common situation that the acquired image is contaminated
with noise and has a slight background shading due to uneven
distributed illumination. A typical preprocessing step can then be
to suppress noise and background variations before further
processing. However, sharpness of the details of interest can also
be reduced by the same preprocessing.
Copyright (c) Benny Thörnberg
12:39
Fundamentals – Segmentation
A segmentation process means that different image components
are separated. In this case, background is separated from letters.
{
}
1,0),( =∧∈ BBcrI12),(0),( −≤≤∧∈
+ N
crIZcrI
Pixel intensity I for a monochromatic image
7
Copyright (c) Benny Thörnberg
13:39
Fundamentals – Segmentation
It can be enough to just use information from the histogram of
pixel intensities I(r,c) but chromatic information, motion or
structure are other powerful features that can be used for
segmentation.
Copyright (c) Benny Thörnberg
14:39
Fundamentals – Labeling
Separate image components need to be identified and coded in a
labeling process. In this case, different colors have been assigned
to different components.
8
Copyright (c) Benny Thörnberg
15:39
Fundamentals – Feature extraction
•Now when different objects are separated into image components,
•How can a nut be distinguished from a screw and how can the
letter A be distinguished from letter B?
•Well selected features of the image components can all together
be used as input vector to a subsequent Classification process.
•Examples of such features are:
•Area
•Color/Intensity
•Moments of statistics of intensity over regions
•Subpixel position
•Compactness
•Anisometry
Copyright (c) Benny Thörnberg
16:39
Fundamentals – Feature extraction
So what features can be used to distinguish between letters?
i
j
5.2≈
b
a
a
a
b
b
0.2≈
b
a
Calculated anisometry
Independent of scaling and rotation.
9
Copyright (c) Benny Thörnberg
17:39
Fundamentals – Feature extraction
•So what features can be used to distinguish between letters?
Perimeter is the length of the region boundary
Compactness
is (Perimeter)
2
/ Area
Copyright (c) Benny Thörnberg
18:39
Fundamentals – Feature extraction
10
Copyright (c) Benny Thörnberg
19:39
Fundamentals – Classification
The graph shows a 2D feature space and its posteriori Probability Density
Function.
Clearly, three distinct classes are visible in this PDF. These classes could
for example in case of OCR correspond to three different characters.
The classification problem
, is how to make the most accurate decision for
a given feature vector to which class it belongs to.
Copyright (c) Benny Thörnberg
20:39
Fundamentals – Classification
Bayes decision rule
)|( xP
i
ω
is the posteriori probability of having class i given a certain feature vector x
This is the probability on which the classification decision can be made.
The right side graph shows hyper lines for which the posteriori probability of
adjacent classes are equal.
11
Copyright (c) Benny Thörnberg
21:39
Fundamentals – Classification
ABC 123
ASCII-codes for
”ABC123”
•In case of traffic surveillance and OCR of registration numbers,
characters are classified and translated into a sequence of ASCII
codes.
•Correlation analysis, partial least squares, support vector machine are
other techniques for classification.
Copyright (c) Benny Thörnberg
22:39
Research at Miun – P
ulp Q
uality M
easurement
Camera
Glass window
12
Copyright (c) Benny Thörnberg
23:39
Research at Miun – PQM

Camera
Glass window


Pre-filter
Labeling

Post-filter

Parameter
extraction

Segmentation
In

Out

B

C

D

E

BIn
C D
Copyright (c) Benny Thörnberg
24:39
Research at Miun – Wood Chip Analyzer
•Chips of wood are used as raw
material at the manufacturing of
pulp and paper in a paper mill.
•The statistical distribution of wood
chip sizes is an important
parameter for the final quality of
the paper.
13
Copyright (c) Benny Thörnberg
25:39
Research at Miun – Wood Chip Analyzer

Range Imager

Collector
Chip feeder

W

V
Principles of a Wood Chip Analyzer
Copyright (c) Benny Thörnberg
26:39
Research at Miun – Wood Chip Analyzer
Range imaging, means acquisition of a 3D representation of a
surface geometry.

R

X

Y

Sheet of light laser
2-D image sensor
d
Scanning in
X-dimension
14
Copyright (c) Benny Thörnberg
27:39
Research at Miun – Wood Chip Analyzer
A range image of wood chips where the height is represented as
gray levels.
Copyright (c) Benny Thörnberg
28:39
Research at Miun – Wood Chip Analyzer
A real world wood chip analyzer
15
Copyright (c) Benny Thörnberg
29:39
Research at Miun – Paper surface Analyzer
Data for MD
Copyright (c) Benny Thörnberg
30:39
Research at Miun – Paper surface Analyzer
Data for MD
16
Copyright (c) Benny Thörnberg
31:39
Research at Miun – Paper surface Analyzer
Data for MD/CD
Copyright (c) Benny Thörnberg
32:39
Research at Miun – Paper surface Analyzer
Data for MD/CD
17
Copyright (c) Benny Thörnberg
33:39
Research at Miun – Paper surface Analyzer
Data for MD/CD
Copyright (c) Benny Thörnberg
34:39
Research at Miun – Paper surface Analyzer
Range image of MD/CD data
18
Copyright (c) Benny Thörnberg
35:39
Position and orientation in 6DOF
Light spots

Camera


Y
Z
X
kZ
kX
kY
α

β

γ

(a,b,c)
Copyright (c) Benny Thörnberg
36:39
Applications for 6DOF
19
Copyright (c) Benny Thörnberg
37:39
Position and orientation in 6DOF
How to reach a consumer market?•Low power, Low latency, Low price
•Small physical size
Copyright (c) Benny Thörnberg
38:39
Magnetic particle detection in
hydraulic oil
Low sampling frequency
Battery powered ￿￿￿￿Low energy consumption
Simple installation ￿￿￿￿Wireless communication
20
Copyright (c) Benny Thörnberg
39:39
Checkpoint questions
•Explain the fundamental steps of machine vision