Segmentation of Digital Images Using a Priori Information about the Expected Image Contents

Abstract

Various image segmentation algorithms are known from literature where images are segmented according to a general common property denoted “Uniformity predicate P” (Fu et al, 1981). In our concept we establish an additional dependency of the predicate P of the imaged objects which we expect and from which we assume to know some general properties, such as shape, surface proportions and so on.

In this paper, a concept of segmenting a given image is presented, which allows the use of different segmentation algorithms. The basic idea is to use effectively an a priori description of the expected image to guide the segmentation and interpretation process. This starts by looking first for locations of the image where more than two regions meet each other. Small portions around these locations are then segmented by regional algorithms. In a second phase the edges of the image graph must be detected by line following algorithms. This concept is motivated by the fact that computing time of regional segmentation algorithms grows with increasing image size, because they have to inspect in general each pixel at least once even if it is known, for example, that there exists a large connected region of a certain type. Our proposal makes use of such knowledge:

• 1 Regional (complete) segmentation is performed only for small windows where more than two regions are expected to meet each other and

• 2 in the rest of the image -that are the border lines between the regions -the boundaries can be found by line detection algorithms under some special conditions:

  • a) start and endpoint of the line are known as edges in already segmented windows and

  • b) it can be assumed that the two adjacent regions of the line have known properties.

In this manner computation time for segmentation can be reduced by two facts:

• 1 Only pixels of special interest for the expected result are inspected and

• 2 Segmentation of subwindows (nodes of the graph) and

searching for different boundaries can be done in parallel.

First results of a simultaneous region growing algorithm are reported.