LMA

1
LMA

• Optical Leaf Recognition Through Analysis of Shape

Jason OKane Aaron Brooks Josh Sooy Matt Mastroine
COS 351 Computer Vision Taylor University, Fall
1999
2
Purpose

• To be able to arbitrarily select a leaf sample
  from an unknown tree species, scan the leaf into
  our system, and correctly identify the species by
  the shape of the leaf.

3
Objectives

• Account for characteristic differences between
  leaves of the same species
• Identify leaves with 80 accuracy regardless of
  damage
• Process each leaf in an average of 15 seconds or
  less

4
Assumptions

• All images will be of leaves
• Each leaf will be oriented vertically
• The background of each image will be nearly white
• Exactly one leaf will appear in each image
• All leaves will come from the set of species
  known to the system

5
Data Collection

• Leaves were collected in late September 1999 from
  the Arboretum at Taylor University.
• 134 leaves were selected from 6 species
• Two data sets were formed
• testing and training set
• experimental set

6
Pre-Processing

• The purpose of pre-processing is to eliminate
  defects in the leaf and to prepare it for
  measurement
• Intensity Thresholding
• Stem Clipping
• Removing Strays and Closing Holes
• Euclidean Distance Map
• Contrast Expansion

7
Intensity Thresholding

• Conversion from color to binary
• A constant threshold 250 (of 255) was applied.

8
Stem Clipping

• Clips the stem which would have interfered with
  our analysis
• Erode, dilate, and bit wise and

9
Removing Strays and Closing Holes

• Prepare leaf for EDM
• Step through the image pixel by pixel creating
  regions
• Take a second pass through the image and merge
  the regions that border each other
• Make the largest region (Black/White) and set
  everything else to the inverse color

10
Euclidean Distance map

• The EDM is the easiest way to generate a series
  of summary forms within a given shape
• Each summary form is less sensitive to the micro
  variations along the perimeter
• A summary form will be used for dippie calculation

11
Contrast Expansion

• The EDM will have a varying histogram depending
  on leaf size
• In order to prepare for an isoline to be drawn
  based on intensity to land in the same part of
  the leaf we contrast expand
• Scales histogram to fit min and max values

12
Metrics and Measurements

• Aspect Ratio
• Vertex Detection
• Isoline Drawing
• Counting Jaggies and Dippies

13
Aspect Ratio

• Measures relative ratio of height to width
• Finds maximum and minimum Y values in image that
  are black
• Subtracts maximum Y from minimum Y
• Repeat for width

14
Vertex Detection

• Obtains jaggie count
• A window based on a ratio of the leafs perimeter
  is moved across the leaf
• A point is a vertex if it
• Is the the lines a b and x yield ab/xgtn
• Where a and b are the linear distances from the
  ends window to an intermediate point
• Where x is the linear distance between the ends
  of the window
• Where n is a cutoff ratio defining a vertex
• Is not within a certain perimeter distance from
  the last vertex

15
Isoline Drawing

• Obtains dippie count
• The EDM simplifies this
• Normal isoline generation more difficult
• EDM yields coherent contiguous color bands which
  are used to generate isolines

16
Counting Jaggies and Dippies

• Jaggies are vertices on perimeter
• Dippies are vertices on the internal isoline
• Ratio gives information about general form of
  leaf
• Validity of jaggies and dippies in practical leaf
  recognition

17
Recognition

• Here we combine metrics for each leaf and use
  fuzzy logic to make an assertion about the leaf
  species.
• Training Set
• Assertion of hypothesis
• Combination of hypothesis

18
Training Set

• Results from the test data set were compiled
• For each species, means and standard deviations
  were found for each metric
• These values were used as a pattern to which
  leaves from the experimental set could be
  compared.

19
Forming Hypotheses

• Calculate a z-score for each metric.
• Use the area under the normal curve between the
  mean and this score to make a hypothesis.

20
Combining Hypotheses

• Well have one hypothesis for each species from
  each metric.
• Assertions can be combined using rules developed
  for the MYCIN expert system.

21
Completing the Recognition Process

• Well have hypotheses for each species.
• Choose the hypothesis with the greatest
  certainty.
• Certainty values are not meaningful on their own,
  but only in comparison to other hypotheses from
  the same leaf.

22
Results

• The system successfully identified most leaves.
• Problems existed with red oak leaves.
• All members of the training and experimental sets
  were severely damaged.
• Could not find patterns against which to match.

23
Conclusion

• Possibilities for further study
• Review of objectives

24
Possibilities for further study

• Handling a greater degree of background noise
• Arbitrary orientation
• Multiple leaves in a single image
• Greater diversity of species
• Improved defect elimination
• Adding self-training elements
• Weighting of individual metrics
• Adjustment of magic numbers

25
Review of Objectives

• Accuracy
• Goal 80
• Actual 77.8 (60.3)
• Average processing time
• Goal 15 seconds
• Actual 15.97 seconds