133-murray (1)

Embed Size (px)

Citation preview

  • 7/28/2019 133-murray (1)

    1/15

    SELECTIVELY FILTERING IMAGEFEATURES USING A

    PERCENTAGE OCCUPANCY

    HIT-OR-MISS TRANSFORM

    Paul Murrayand

    Stephen Marshall

  • 7/28/2019 133-murray (1)

    2/15

    Overview

    The Hit-or-Miss Transform (HMT) The Percentage Occupancy Hit-or-Miss Transform

    (POHMT)

    The Percentage Occupancy Plot (PO Plot) Properties of the PO Plot

    A Method for Selectively Filtering Image Features

    Examples

    Conclusions

  • 7/28/2019 133-murray (1)

    3/15

    The Hit-or-Miss TransformMorphological transform that can be designed for featuredetection/pattern recognition

    Uses two SEs which probe the image in search of a match

    One SE searches for places where it fits inside features in the foreground

    A second searches for places where it fits around features in the background

    A positive result is returned if and only i f both SEs find a match at somepoint in the image

    ForegroundProbe

    BackgroundProbe

  • 7/28/2019 133-murray (1)

    4/15

    The Hit-or-Miss Transform

    BBG

    BFG

    BBG

    BFG

    In some cases, noise prevents the probes from fitting a feature whichshould otherwise be matched

    This can be overcome by allowing a percentage P < 100% of the probes tofit the image

  • 7/28/2019 133-murray (1)

    5/15

    The Percentage Occupancy

    Hit-or-Miss Transform

    The POHMT may be computed by measuring the extentto which image features occupy the SEs

    A feature is detected by the POHMT ifP% of the SEs areoccupied by some feature

    0 50 100 150 200 2500

    20

    40

    60

    80

    100

    Extent to which BFG

    and BBG

    are occupied

    Intensity

    POo

    fSEs

    POFG

    POBG

    T=50

    T=150

    0 50 100 150 200 2500

    20

    40

    60

    80

    100

    Extent to which BFG

    and BBG

    are occupied

    Intensity

    POo

    fSEs

    POFG

    POBG

    100P 96P

  • 7/28/2019 133-murray (1)

    6/15

    The PO Plot

    A PO plot is generated by plotting the percentageoccupancy of both probes against each other

    The parameter P of the POHMT may be estimated byfinding the value of the critical point in the plot

  • 7/28/2019 133-murray (1)

    7/15

    The PO Plot and the POHMT

    A PO plot may be used to estimate P from training data

    before applying the POHMT

    0 20 40 60 80 1000

    20

    40

    60

    80

    100

    PO P lot

    POFG

    POBG

    Top Left CellTop Right Cell

    Bottom Left Cell

    Bottom Right CellP=70

    One pair of SEs were designed to match all 4 cells in the

    image

    The PO plot was used to estimate the only parameter, P

    The POHMT accurately detects all 4 cells despite thenoise in the image when P = 70%

  • 7/28/2019 133-murray (1)

    8/15

    Selective Filtering

    The PO plot can be used to estimate values ofP such that some featuresare discarded by the POHMT

    Increasing the value ofP makes the transform more strict leading to fewercells being detected

    0 20 40 60 80 1000

    20

    40

    60

    80

    100

    PO Plot

    POFG

    PO

    BG

    Top Left Cell

    Top Right Cell

    Bottom Left Cell

    Bottom Right Cell

    P=75%

    3 cells

    P = 96%

    1 cell

    P = 90%

    2 cells

  • 7/28/2019 133-murray (1)

    9/15

    Selective Filtering

    The previous example shows a powerful but simpletechnique for selectively filtering image features

    It is obvious that increasing the strictness of thetransform by increasing the value ofP leads to fewerfeatures being detected

    However, it is NOT easy to isolate any of the fourcells in the image

    The following novel method allows any combination

    of the image features to be segmented using justone composite SE

    Lets try to isolate the cell in the bottom right of theimage which occupies around 80% of the SEs

  • 7/28/2019 133-murray (1)

    10/15

    Selective Filtering

    Set P low enough to detect the feature of interest

    This leads to features other than the desired one to be detected

    Overlay the bounding boxes computed for the previous image Set P high enough to exclude the feature of interest Compute a bounding box for each connected component in the

    image

    Locate any empty bounding boxes these are the locations of the

    sought feature Mark the position of any empty bounding boxes in the ouput image

  • 7/28/2019 133-murray (1)

    11/15

    Selective Filtering - Example

    0 20 40 60 80 1000

    20

    40

    60

    80

    100

    PO Plot

    POFG

    POBGP = 95 P = 91 P = 88

    P = 84 P = 80 P = 76

    A complementary pair of SEs were designed to detect dice in an image

    A PO plot was generated for each face of the die using these SEs

    The critical points of each SE were computed to allow accurate setting ofparameters for the selective filtering method using bounding boxes

  • 7/28/2019 133-murray (1)

    12/15

    Selective Filtering - Example

    Isolating dice displaying 1 Isolating dice displaying 6

    Isolating dice displaying 3 Isolating dice displaying 2 or 5

  • 7/28/2019 133-murray (1)

    13/15

    Selective Filtering - Example

    0 20 40 60 80 1000

    20

    40

    60

    80

    100

    PO Plot

    POFG

    POBG

    Noise was added to the imageThe PO plot was regenerated

    The selective filtering algorithm was usedto locate dice displaying a value of 2

  • 7/28/2019 133-murray (1)

    14/15

    Summary

    The HMT is a powerful method which can be used for fortemplate matching

    The HMT is unreliable for processing noisy data

    The POHMT is an extension of the HMT which offersimproved performance for processing noisy data

    The PO plot is a novel design tool which can be used toestimate parameters for the POHMT and similar transforms

    The PO plot can be exploited to make the POHMT capable ofselectively filtering image features

    A novel approach which uses bounding boxes to locatefeatures of interest has been demonstrated

    This approach allows features to be isolated in an imagebased on the extent to which they occupy a complementary

    pair of SEs

  • 7/28/2019 133-murray (1)

    15/15