Project Overview

1
Auto-Palynology
AutoStage Presented by Gary Allen
Representing the Pollen Research Group Massey
University
2
Pollen Why bother?

• 81 counting stations in North America mostly
  manned by volunteers for advisements of airborne
  pollen to inhalant allergy and asthma sufferers
• 30 months of microscope work for a PhD student
  using palynology
• Identification and verification of type and
  origin of honey
• Forensic Palynology

SEM

• Answer to take the laborious and time consuming
  work out of palynology.

3
Palynology the study of Pollen

• Pollen multinucleate gametophyte generation of
  flowering plants
• 10 100 microns diameter
• Outer wall is very durable
• Textural features20 microns down to
  nanometresResolution achieved 1 micron

SEM

Birch
Grass
Pine
AutoStage
4
AutoStage Flow

• The study of Pollen
• The system developed
• The Stage
• Lighting
• Microscopes
• Auto-focus
• Segmentation
• Features
• Classification
• Results

5
AutoStage Flow
Extraction of mathematical features
Classification using MLP
User adjustment of counts
Display of grouped pollen
6
The Stage

• Two linear movements in the horizontal plane and
  orthogonal to the optical axis
• Stepper motors/driver
• 1.8steps and
• 1/10th micro-steps
• 2.6 microns per step of linear movement

7
Lighting

• Quartz-halogen
• Transmission lighting
• Band pass filtered
• Yellow/green (550nm)
• Diffusing filter with opaque blocks
• Dark Field illumination
• Non-coherent lighting
• Constant
• location intensity

8
Microscope design

• Design investigation by Craig Holdaway in his
  Masters project
• Two microscopes high magnification low
  magnification
• High magnification 11.2x
• Provides 0.4 microns per pixel, sufficient to
  make use of the one micron resolution
• produces images fit for feature extraction
• Low magnification 1x
• wide field of view for locating objects rapidly
• sufficient depth of field to cover the slide
  without refocus
• Firewire USB serial connections to PC

High Magnification Microscope
9
Auto-Focus

• Low magnification performed once only
• standard deviation function
• High magnification performed for each object
  found
• maximum gradient squared function
• Local peak algorithm to detect peaks in focus
  array

Focus through entire slide
Focus through pollen
10
Segmentation
Edge detection canny
Original image
Dilate fill
Erode
Remove non-pollen

• Bounding rectangle
• Convex hull
• Tests for pollen
• Aspect ratio
• Area ratios

11
Features used for discrimination

• 43 features determined by Dr. Zhang Prof.
  Hodgson within the pollen research group at
  Massey University
• Textural features
• A series of Wavelet transforms that measure
  localised spatial, spatial-frequency content
  using Gabor and Orthogonal Wavelet transforms.
• Orientation sensitivity is reduced by averaging
  the results corresponding to different directions
  
• Grey Level Co-occurrence Matrix
• Grey Gradient Co-occurrence Matrix
• Other features
• Shape, histogram, and second-moments features

12
Classification using a neural network

• Multi Layer Perceptron
• Feed-forward multi-layer
• back-propagation using sum of squares error
• NETLAB algorithm (Ian Nabney)
• No. of Inputs No. of Features (43)
• One hidden Layer with 2n1 hidden nodes
• No. of Outputs No. of Pollen types detritus

13
Results 6 pollen on one slide
14
QUESTIONS