A machine vision system for automated field–level wood identification

coatiarfAI and Robotics

Oct 17, 2013 (3 years and 9 months ago)

62 views

Background
Machine Vision
Conclusion
A machine vision systemfor automated
field–level wood identification
J.Hermanson
1;2
A.Wiedenhoeft
1;2
S.Gardner
3
1
U.S.Forest Service Forest Products Laboratory
2
Stump öch Boles
3
U.S.Forest Service International Programs
GTTN Beijing 2013
Hermanson,Wiedenhoeft,Gardner
FPL WoodID
Background
Machine Vision
Conclusion
Problem
Project
Background
Do not ask me to identify wood without
the “xyloscope”
Experimental and theoretical
mechanics training
Interested in Physical System
Identification (curious)
Love gadgets (nerd)
My “microscopes” that I built =)
The patterns I recognize
r
4
'= 0
g(r;c) =
1
2
2
exp

r
2
+c
2
2
2
Latin pronunciation challenged
Hermanson,Wiedenhoeft,Gardner
FPL WoodID
Background
Machine Vision
Conclusion
Problem
Project
The Problem
Botanical identification of wood and wood products is a
limiting factor for enforcing laws that combat illegal logging
Human capacity is limited
Time to train
Cost of training
Mobility of individuals
Need a technological rather than a human solution
Hermanson,Wiedenhoeft,Gardner
FPL WoodID
Background
Machine Vision
Conclusion
Problem
Project
The Project
Build a machine vision systemthat can be used as a field
deployable tool to identify woody taxa.
1
1
The Fine Print:with an accuracy equal to or better than
average field personnel with one week of training
Hermanson,Wiedenhoeft,Gardner
FPL WoodID
Background
Machine Vision
Conclusion
Approach
Philosophy
Digital Image
Demonstration
The Approach
Pros and cons of a machine vision system for wood ID
Pros
Digital imaging can be
more sensitive than the
human eye
Machine vision systems
collect copious amounts
of data
The collected data can
be"mined"to enhance
the knowledge of wood
anatomical structure
Pros continued
Machines do not forget,
get bored,promoted,
fired,etc
Cons
Machines are dumb
Garbage in garbage
out
Requires robust
programing
Needs Big Data
Hermanson,Wiedenhoeft,Gardner
FPL WoodID
Background
Machine Vision
Conclusion
Approach
Philosophy
Digital Image
Demonstration
The Operational Philosophy
Develop an open system
Utilize low cost off–the–shelf hardware
Utilize open source software
Have a central database
Create an open platform for others to further development
High–throughput
Repeatable
Easy–to–use
Economical
We use this systemto capture digital images of unknown
wood specimens
Hermanson,Wiedenhoeft,Gardner
FPL WoodID
Background
Machine Vision
Conclusion
Approach
Philosophy
Digital Image
Demonstration
Anatomy of a Digital Image
A simple image
A 3D representation
Hermanson,Wiedenhoeft,Gardner
FPL WoodID
Background
Machine Vision
Conclusion
Approach
Philosophy
Digital Image
Demonstration
Anatomy of a Digital Image
A less simple image
A 3D representation
Hermanson,Wiedenhoeft,Gardner
FPL WoodID
Background
Machine Vision
Conclusion
Approach
Philosophy
Digital Image
Demonstration
Anatomy of a Digital Image
Hermanson,Wiedenhoeft,Gardner
FPL WoodID
Background
Machine Vision
Conclusion
Approach
Philosophy
Digital Image
Demonstration
Anatomy of a Digital Image
Hermanson,Wiedenhoeft,Gardner
FPL WoodID
Background
Machine Vision
Conclusion
Approach
Philosophy
Digital Image
Demonstration
Anatomy of a Digital Image
The computational demands
of feature detection exceed
the capacity of affordable,
portable devices,so we need
an alternate method to
analyze the digital images
Hermanson,Wiedenhoeft,Gardner
FPL WoodID
Background
Machine Vision
Conclusion
Approach
Philosophy
Digital Image
Demonstration
Analysis of a digital image
Signal processing methods
Commonly used in digital image processing
Optical Character Recognition
Fingerprint recognition
Computationally,they are fast and efficient
Their output cannot be readily mapped back to specific
wood anatomical features
The signal processing technique
we have chosen is wavelets
Hermanson,Wiedenhoeft,Gardner
FPL WoodID
Background
Machine Vision
Conclusion
Approach
Philosophy
Digital Image
Demonstration
Analysis of a digital image
Wavelets
Characterizes the topology of the image using a function
Extracts signal horizontally and vertically from the image
Extracts signal at iterative reductions in scale,capturing
variability across scales
We calculate the energy in the image at 10 scales
This gives us a 33-dimensional space for analysis and
identification
Haar Wavelet Meyer Wavelet Signal
Hermanson,Wiedenhoeft,Gardner
FPL WoodID
Background
Machine Vision
Conclusion
Approach
Philosophy
Digital Image
Demonstration
Analysis of a digital image
Hermanson,Wiedenhoeft,Gardner
FPL WoodID
Background
Machine Vision
Conclusion
Approach
Philosophy
Digital Image
Demonstration
Analysis of a digital image
Hermanson,Wiedenhoeft,Gardner
FPL WoodID
Background
Machine Vision
Conclusion
Approach
Philosophy
Digital Image
Demonstration
Demonstration
Hermanson,Wiedenhoeft,Gardner
FPL WoodID
Background
Machine Vision
Conclusion
Take-home message
This is a field-deployable
machine vision system that,as of
right now,is as accurate as a
person with one week of wood
identification training
Hermanson,Wiedenhoeft,Gardner
FPL WoodID
Background
Machine Vision
Conclusion
Acknowledgements
Funding from US Department of State
US Forest Service International Programs
Alex Moad and Val Mezainis
US Forest Service Forest Products Laboratory
Andre Lima,Bruna Ferreira,Dave Dostal,Joe Destree,and
Dick Jordan
L
A
T
E
X with the BEAMER class
Hermanson,Wiedenhoeft,Gardner
FPL WoodID
Background
Machine Vision
Conclusion
Thank you – Xie xie
Nothing in Nature is random....A thing appears
random only through the incompleteness of our
knowledge.
Baruch Spinoza
Hermanson,Wiedenhoeft,Gardner
FPL WoodID