Pattern recognition tools for image data

                          Dick de Ridder

                         Delft University

                              Abstract 

The talk will start with a brief overview of my work as a
Ph.D. student, on the application of artificial neural networks (ANNs)
to image processing problems. In this project, the goal was to find
the limits of ANNs, not necessarily to outperform any given standard
approach. The questions we tried to answer were:

- What problems can ANNs be applied to with success? 
- How can we use prior knowledge?
- What can an image processing expert learn from a trained ANN?

I will discuss some of the lessons we learned during this project.

The talk will then continue with a related subject: modelling image
manifolds. In recent years, there has been increasing interest in
fitting low-dimensional, nonlinear models to high-dimensional data,
such as images or image windows represented as grey-value vectors. I
will discuss an approach based on a mixture of local subspaces, and
some applications to texture segmentation and object recognition. This
model still has some problems, and other methods might in some
circumstances be preferable. The talk will finish by giving a broad
overview of interesting methods for learning image manifolds.