Browse Prior Art Database

Edge Class Training for Supervised Image Segmentation

IP.com Disclosure Number: IPCOM000120885D
Original Publication Date: 1991-Jun-01
Included in the Prior Art Database: 2005-Apr-02
Document File: 3 page(s) / 124K

Publishing Venue

IBM

Related People

Dom, BE: AUTHOR

Abstract

Disclosed is an algorithm for automatically forming an edge class for the purpose of training a pixel classifier in supervised multi-or single-band greyscale image segmentation. In image segmentation, it may be desirable to classify pixels as "edges" (thus defining an edge class) for two reasons: 1. For subsequent (post-segmentation) processing, it may be desirable to identify pixels as belonging to the transition region between materials. 2. Due to the difference in appearance of pixels (and their neighborhoods) in these transition regions, it may be desirable to recognize them as edges to reduce overall confusion and the probability of classifying them as something totally erroneous.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 52% of the total text.

Edge Class Training for Supervised Image Segmentation

      Disclosed is an algorithm for automatically forming an
edge class for the purpose of training a pixel classifier in
supervised multi-or single-band greyscale image segmentation.  In
image segmentation, it may be desirable to classify pixels as "edges"
(thus defining an edge class) for two reasons:
      1.   For subsequent (post-segmentation) processing, it may be
desirable to identify pixels as belonging to the transition region
between materials.
      2.   Due to the difference in appearance of pixels (and their
neighborhoods) in these transition regions, it may be desirable to
recognize them as edges to reduce overall confusion and the
probability of classifying them as something totally erroneous.

      The algorithm described is extremely general in the sense that
it utilizes any features, classifier, and classifier training
algorithms and optimizes the edge class definition (implicitly by
forming the training mask) for the given tuple (images, masks,
features, classifier, and training algorithm).  It does this in a
self-consistent fashion by growing the edge class to include pixels
that the classifier cannot identify correctly.

      The following is a description of the solution to this edge
class definition problem.  It is an algorithm that forms and
iteratively modifies an edge class in the training masks until no
improvement in the measured error rate is observed.  The algorithm is
shown in pseudocode form in the figure.

      1.   The initial training masks, with no edge class, are formed
by manually segmenting the training images using a special graphics/
mouse-based program.  These edge less masks are then input to the
rest of the procedure.

      2.   The initial edge class definition (referred to as the
seed) is formed by labeling every pixel adjacent to a class boundary
in the original (edge less) training masks as an "edge".  For long,
straight, horizontal, and vertical edges, this will result in a
two-pixel-wide edge class, for example.  If multiple edge classes are
being used (say, one for every class pair), the edge class label is
determined from the classes of the adjacent classes.

      3.   Next, a limiting band is formed by dilating the seed to a
certain width.  The width is based on prior knowledge about the
images.  This limiting band is intended to define a zone within which
it is reasonable to label pixels as edges.  This is one o...