A New Class of Learnable Detectors for Categorisation

Matas, Jiri, Zimmermann, Karel
SCIA 2005, Heidelberg, Germany
SCIA '05: Proceedings of the 14th Scandinavian Conference on Image Analysis
Volume 1, Number 3540, Pages 541-550
June, 2005

Abstract

A new class of image-level detectors that can be adapted by machine learning techniques to detect parts of objects from a given category is proposed. A classifier (e.g. neural network or adaboost) within the detector selects a relevant subset of extremal regions, i.e. regions that are connected components of a thresholded image. Properties of extremal regions render the detector very robust to illumination change. Robustness to viewpoint change is achieved by using invariant descriptors and/or by modelling shape variations by the classifier. The approach is brought to bear on three problems: text detection, face segmentation and leopard skin detection. High detection rates were obtained for unconstrained (i.e. brihtness, affine and font invariant) text detection (92%) with reasonable false positive rate. The time-complexity of the detection is approximately linear in the umber of pixel and a non-optimized implementation runs at about 1frame per second for a 640x480 image on a high-end PC.

Keywords

Object recognition, distinguished regions, CSER, extremal regions, MSER, machine learning

Bibtex entry

@InProceedings{Matas-SCIA-2005,
  author =      {Matas, Jiri and Zimmermann, Karel },
  title =       {A New Class of Learnable Detectors for Categorisation},
  year =        {2005},
  pages =       {541--550},
  booktitle =   {SCIA '05: Proceedings of the 14th Scandinavian Conference on
                  Image Analysis},
  editor =      {Kalviainen, Heikki and Parkkinen, Jussi and Kaarna, Arto},
  publisher =   {Springer-Verlag },
  address =     {Heidelberg, Germany },
  isbn =        {0302-9743 },
  volume =      {1},
  series =      {LNCS},
  number =      {3540},
  month =       {June},
  day =         {19-22},
  venue =       {Joensuu, Finland},
  annote = {A new class of image-level detectors that can be adapted
    by machine learning techniques to detect parts of objects from a
    given category is proposed. A classifier (e.g. neural network or
    adaboost) within the detector selects a relevant subset of
    extremal regions, i.e. regions that are connected components of a
    thresholded image. Properties of extremal regions render the
    detector very robust to illumination change.  Robustness to
    viewpoint change is achieved by using invariant descriptors and/or
    by modelling shape variations by the classifier. The approach is
    brought to bear on three problems: text detection, face
    segmentation and leopard skin detection. High detection rates were
    obtained for unconstrained (i.e. brihtness, affine and font
    invariant) text detection (92
    rate. The time-complexity of the detection is approximately linear
    in the umber of pixel and a non-optimized implementation runs at
    about 1frame per second for a 640x480 image on a high-end
    PC.},
  keywords =    {Object recognition, distinguished regions, CSER, 
                 extremal regions, MSER, machine learning},
}