|
A new image segmentation algorithm based on the textural
information of image pixels is presented. Image segmentation is
a fundamental technique for the application of computer vision.
There are two difficulties for the technique of image
segmentation.First, the choice of the characteristic with which
the regions of an image segmentation are homogeneous enough for
the decision.Second,the connection problem for pixels within
the segmented components in an arbitrary shape. In this thesis,
we propose a new property vector related to the concept of
image texture and employ the technique of connected components
labeling to solve the problems. The image structure are defined
to be the linear relationship between pixels with their upper
and left neighbors. Image can be very inhomogenious. However,
the image structure may be uniform enough in some image
components. By this image texture, individual image constraint
equations can be set up for those triple pixel. These equations
are the property vectors chosen by us for segmentation
decision. Discrepancy between equations are measured by
sequential least squared method. A recursive method for
computing the error is developed in this thesis for simplifying
computation. Connected components labeling method was
originally developed for the binary images. By the introduction
of our property vector, the labeling method are extended into
the gray image segmentation. For computing efficiency, the
Ronse and Devijver''s run-length version of labeling method is
modified by us for our application. Our methods of computing
segmentation error for property vectors and labeling
segmentation components are both based upon a top-down and left-
right scanning order. As a summary, our segmentation computing
are very efficient due to the recursive computation structure
and the scanning method.
|