SIGMEDIA - Example Based Image Segmentation

by Claire Gallagher

Keywords: Image Segmentation, Image modelling, Wavelet Based Analysis, Bayesian framework.

The segmentation of an observed image into an unknown number of distinct and in some way homogeneous regions remains a fundamental issue in low-level image analysis. Image segmentation is the research area that addresses this problem. The goal of an image segmentation process is to assign to each pixel in an observed image a label indicating to which region or class that pixel belongs.

Fully automated or unsupervised segmentation is a difficult problem to solve and often the solution is largely subjective. Often and as a means to constrain the nature of the problem some information regarding the nature of the observed image is included in the segmentation process. In example based segmentation , this information is provided by means of an example set of images which are similar in content to the image to be segmented. These example images are composed of texture components which are similar in content to those found in the observed image and each of these texture images has associated with it a label indicating to which class that texture belongs. This new example set of textures will be used as a guide in the segmentation of the observed image.

This type of semi-automated segmentation can be viewed as the interleaving of segmentation and object recognition. The example based segmentation technique follows on from the success of the texture synthesis algorithm and uses the same inplicit modelling process over the observed image. In order to regularise the solution, the implicit modelling of the observed image is combined with explicit modelling of the label field. The Bayesian framework provides a natural expression for such parallel modelling techniques and the new algorithm is presented under this framework.

Below are some preliminary results which demonstrate how the example based technique can be used for object recognition. An outlier class condition has been added to the segmentation process to account for texture regions in the observed image which are similar to the example texture images.