The Image Processing Corner

Introduction

Segmenting an image into regions consists in creating a partition of the image with constrained and connex components...
More seriously, the result of a segmentation into regions is a classification of the image into several areas so that "similar" pixels are grouped together.

As talking about image is easier with images, let's give an example: the "pentagon" image below has been segmented into 53 regions, or pieces of image.
The right image is a region image each with random color.

Interest

• Interest area extraction in order to apply a process only on a piece of image:

  • divX encoding compress image area separatly and thus avoid block effects of the MPEG algorithm near contours.
  • All drawing softwares (Gimp, Photoshop, ...) have a "magic wand" which purpose is to select an area of the image.

• Modelling or simplification of the image. An image, array of several thousands of colored pixels, is simplified into a handful of regions of any size and shape. Thus, the following processes will have less information to handle, but with increased complexity.

Limits

The second reason is the difficulty to control the segmentation: segmentation algorithms are seldom stable. Every algorithm gives specific results, and even more, they are often extremely sensitive to parameter variations.
The major problem is that the result is often oversegmented (too much regions) or undersegmented (too few regions) and that it is difficult to automatically obtain a good compromise.

The last reason is the increased algorithmic complexity. Working with regions is much more complex than working with pixels. Their high information content has a price:

• Geometry: pixels are like points, 2 coordinates are enough to describe their geometry. Regions may have any shape: their geometry is described by an outer closed contour and maybe several inner closed contours if some regions are included in it.
• Radiometry: pixels have one color value each. Region have one color value for each included pixel. Taking their average or median value is often too restrictive.
• Connexity: a pixel has a fix neighborhood (up, down left, right in 4-connexity). A region has a variable quantity of neighbors, which may be connected in any place of their contours.
• Memory representation: pixels are stored in an array, so with a direct access. Regions compose a real graph which representation is not obvious.

The example below shows different results of segmentation into regions of the same image, with the same algorithm, but with different parameters.

The first row displays images with special colors to enhance contrasts.
The second row displays images with random colors for each area. The region color is the average color.
The first parameter of the used algorithm is the noise intensity in the image, the second parameter is the minimum size of a region (see next sections for further details).

First remark: for a similar region quantity, segmentations may be very different, even if the same algorithm is used.

It is also clear that "simplifying" an image leads to undesirable pixel associations, especially here between the table and the background. The human eye perceives this like an "error" as it analyzes and separates objects of the scene, but computers only see colored pixels side by side.

The next section deals with some types of current algorithms.