dvg2107-Presentation.ppt

Download Report

Transcript dvg2107-Presentation.ppt

Identifying Tree Species From
Bark
David Gainer
Why Bark?
•
•
•
•
Year-Round
Can be used for dead trees/logs/stumps
Can be observed at ground level
Can be combined with systems for
leafs/needles/twigs/etc.
Inherent Difficulties
• Age:
Old Trees Have Very Different Looking Bark than
Young Trees (So a complete system would have to
identify tree and age range).
Juvenile Bunya Pine
Mature Bunya Pine
Pictures from: http://tree-species.blogspot.com/
Inherent Difficulties
• Shape: Shape of the trunk effects the
appearance of the bark.
Human Performance
• Osterreichische Bundesforste AG Dataset
• 2 Experts: 56.6% and 77.8% (trouble with
the different pines)
Source: Fiel and Sablatnig http://cvww2011.icg.tugraz.at/papers_web/p13.pdf
Prior Results/Approaches
• Wan et al (2004) 77% with GLCM. Higher using
color bands separately. 170 images of 9 classes
• Song et al (2004) 87.5% with GLCM and binary
features. 180 images of 8 classes
• Huang et al (2006) 92.5% with GLCM and fractal
dimensions. 360 images of 24 classes
• Fiel and Sablating (2011) 69.7% vectors of SIFT
pattern matches. 1183 images of 11 classes.
Found that GLCM only yielded 61% for the same
dataset
My Dataset
• Pictures of species in Central Park
• 2 Pictures for each individual tree
• Taken with iPhone, for each tree 1 picture
using HDR setting, 1 with the standard
• Taken between 14” and 16” away from the
tree
• Goal is 10 or more pictures for each
species for 20 or more species
• Currently 4-14 pictures for 14 tree species
My Dataset
• Current Species: American Elm, London
Plane, Red Oak, Eastern White Pine,
Willow, Beech, Cherry (Kwanzan), Norway
Maple, Birch (Himalayan Whitebark),
Linden, American Sycamore, Blue Spruce,
Sycamore Maple, Swamp Oak
• Identified with Help of Ned Barnard and
Ken Chayas “Central Park Entire” map
American Elm
Birch
Birch(Himalayan
(Himalayan Whitebark)
Whitebark)
London Plane
London Plane
American Elm
Beech
Beech
Feature Selection
• From the prior research, I intend to
implement GLCM, look at fractal
dimensions
• I’d also like to come up with some new
statistics especially from various qualities
of edge images
• I’d also like to look at statistics derived
from line detection
Feature Selection
• Gradient/Edge images in X and Y
directions:
Classification
• Compare k-nn, multiclass SVM and
maximum likelihood approaches