Scale Invariant Object Detection using a Hybrid Genetic Algorithm

1
Scale Invariant Object Detection using a Hybrid
Genetic Algorithm Fuzzy Logic Approach

• Group 9
• Ayesha Farrukh 04030004
• Junaid Akhtar 04030019

2
Progress

• Matlab Implementation Brute force Template
  Matching
• Initialize Population
• Crossover
• Mutation
• Fitness Function

3
Brute Force Results
Template
Reference Image
4
Cross Correlation Surface
5
Correlation Surface zoomed in
Peak value at (145, 171) Value 0.9666
6
Tic-Toc

• This operation took 152.01 seconds

7
Other Matlab Functions Explored

• DEC2BIN
• BIN2DEC
• RAND
• CORR2

8
GA Pseudocode

• StartGA( img, template )
• // Initialize a random population of
  individuals
• P initPopulation ( populationSize,
  imgDimension, scaleDimension )
• // Evaluate fitness of all initial individuals
  of population
• fitnessP, fuzzyClassifictaionP
  calcFitness ( P, img , template )
• // test for termination criterion
  (Correlation Value gt THRESHOLD)
• WHILE not done do
• // select a sub-population for offspring
  production
• selectedP selectparents( P )
• // Crossover
• newP crossover( selectedP )
• // Mutate x OR y OR scale value randomly
• mutatedP mutate( newP )

9
Chromosome Definition

• 000110001 011111100 1001

x 49
y 252
Index of Scale to Use
10
512 x 512 Reference Image
11
Template Image
12
Scale

• imresize
• Bicubic Interpolation for resizing the template
  image

13
Initial Population

• function initialPopulation, lengthImageDimension,
  lengthMaxScale initialize(populationSize,
  imageDimension, maxScale)
• binaryImageDimension dec2bin(imageDimension)
• binaryMaxScale dec2bin(maxScale)
• lengthImageDimension length(binaryImageDimension
  )
• lengthMaxScale length(binaryMaxScale)
• bitsInitialPopulation (2 lengthImageDimension)
  lengthMaxScale
• initialPopulation zeros(populationSize,
  bitsInitialPopulation)
• for i 1populationSize
• xLocation round(imageDimensionrand)
• xLocation dec2bin(xLocation)
• initialPopulation(i,length(binaryImageDimensio
  n) - length(xLocation) 1 length(binaryImageDime
  nsion)) xLocation - 48
• yLocation round(imageDimensionrand)
• yLocation dec2bin(yLocation)
• initialPopulation(i,(2length(binaryImageDimen
  sion)) - length(yLocation) 1 2length(binaryIma
  geDimension)) yLocation - 48
• scale round(maxScalerand)

14
Select Parents

• Sort the current Population according to their
  respective fitness values
• Select the best

15
CalculateFitness

• Use corr2 function to calculate fitness for each
  individual
• Classify each individual using Fuzzy classifiers
  Negative, Low, Positive correlation

16
Crossover

• function newPopulation crossover(selectedPopul
  ation)
• sizeSelectedPopulation size(selectedPopulation,1
  )
• sizeIndividual size(selectedPopulation,2)
• count 1
• while count lt sizeSelectedPopulation
• crossoverPoint round(sizeIndividualrand)
• newPopulation(count,) selectedPopulation(co
  unt,)
• newPopulation(count1,) selectedPopulation(
  count1,)
• newPopulation(count,crossoverPointsizeIndivid
  ual) selectedPopulation(count1,crossoverPoints
  izeIndividual)
• newPopulation(count1,1crossoverPoint-1)
  selectedPopulation(count,1crossoverPoint-1)
• count count 2
• end

17
Mutate( P )
0011000111111100100100 0011000111111100100100
0011000101111100100100

• For each member of the selected population P
• Randomly decide whether or not to mutate
• Randomly decide index of bit to mutate
• Perform mutation

18
Next Episode

• Integrate the components
• Compile Results on Different Images
• Use information generated during fuzzy
  classification to improve matching time
• Compare Tic-Toc Results for time
• Compare FLOPS for memory