week03-patternRec

  • Upload
    ayaz

  • View
    216

  • Download
    0

Embed Size (px)

Citation preview

  • 7/30/2019 week03-patternRec

    1/18

    CSE803 Fall 2012 1

    Pattern Recognition Concepts Chapter 4: Shapiro and Stockman

    How should objects be represented?

    Algorithms for recognition/matching

    * nearest neighbors

    * decision tree

    * decision functions* artificial neural networks

    How should learning/training be done?

  • 7/30/2019 week03-patternRec

    2/18

    CSE803 Fall 2012 2

    Feature Vector Representation X=[x1, x2, , xn], each xj

    a real number

    Xj may be objectmeasurement

    Xj may be count of objectparts

    Example: object rep.[#holes, Area, moments, ]

  • 7/30/2019 week03-patternRec

    3/18

    CSE803 Fall 2012 3

    Possible features for char rec.

  • 7/30/2019 week03-patternRec

    4/18

    CSE803 Fall 2012 4

    Some Terminology Classes: set of m known classes of objects

    (a) might have known description for each

    (b) might have set of samples for each

    Reject Class:

    a generic class for objects not in any of

    the designated known classes Classifier:

    Assigns object to a class based on features

  • 7/30/2019 week03-patternRec

    5/18

    CSE803 Fall 2012 5

    Classification paradigms

  • 7/30/2019 week03-patternRec

    6/18

    CSE803 Fall 2012 6

    Discriminant functions Functions f(x, K)

    perform some

    computation onfeature vector x

    Knowledge K fromtraining orprogramming isused

    Final stagedetermines class

  • 7/30/2019 week03-patternRec

    7/18

    CSE803 Fall 2012 7

    Decision-Tree Classifier Uses subsets of

    features in seq.

    Feature extractionmay be interleavedwith classificationdecisions

    Can be easy todesign and efficientin execution

  • 7/30/2019 week03-patternRec

    8/18

    CSE803 Fall 2012 8

    Decision Trees

    #holes

    moment of

    inertia#strokes #strokes

    best axis

    direction #strokes

    - / 1 x w 0 A 8 B

    0

    12

    < t t

    2 4

    0 1

    060

    90

    0 1

  • 7/30/2019 week03-patternRec

    9/18

    CSE803 Fall 2012 9

    Classification using nearest

    class mean Compute the Euclidean

    distance between

    feature vector X and themean of each class.

    Choose closest class, ifclose enough (reject

    otherwise) Low error rate at left

  • 7/30/2019 week03-patternRec

    10/18

    CSE803 Fall 2012 10

    Nearest mean might yield poor

    results with complex structure Class 2 has two

    modes

    If modes aredetected, twosubclass meanvectors can be

    used

  • 7/30/2019 week03-patternRec

    11/18

    CSE803 Fall 2012 11

    Scaling coordinates by std dev

  • 7/30/2019 week03-patternRec

    12/18

    CSE803 Fall 2012 12

    Another problem for nearest

    mean classification If unscaled, object X

    is equidistant fromeach class mean

    With scaling X closerto left distribution

    Coordinate axes notnatural for this data

    1D discriminationpossible with PCA

  • 7/30/2019 week03-patternRec

    13/18

    CSE803 Fall 2012 13

    Receiver Operating Curve ROC Plots correct

    detection rateversus false

    alarm rate Generally, false

    alarms go upwith attempts todetect higherpercentages ofknown objects

  • 7/30/2019 week03-patternRec

    14/18

    CSE803 Fall 2012 14

    Confusion matrix shows

    empirical performance

  • 7/30/2019 week03-patternRec

    15/18

    CSE803 Fall 2012 15

    Bayesian decision-making

  • 7/30/2019 week03-patternRec

    16/18

    CSE803 Fall 2012 16

    Normal distribution

    0 mean and unitstd deviation

    Table enables us tofit histograms andrepresent themsimply

    New observation ofvariable x can thenbe translated intoprobability

  • 7/30/2019 week03-patternRec

    17/18

    CSE803 Fall 2012 17

    Cherry with bruise

    Intensities at about 750 nanometers wavelength

    Some overlap caused by cherry surface turning away

  • 7/30/2019 week03-patternRec

    18/18

    CSE803 Fall 2012 18

    Parametric Models can beused