Neural Network Approach to Scale Space Grouping in Image Analysis (UGN.4496)

Highly developed human and primate visual systems are capable of performing formidable processing tasks in both natural and controlled environments. Such tasks include figure-background segregation, object segmentation, pattern recognition, feature detection and linking. Although a lot of research has been done to understand the basic facts of natural visual processing, no general theory is yet present, and the establishment of a satisfactory system of self-contained fundamental principles still appears to be in its infancy.

Using the amazing abilities of their natural visual system, trained operators can track and highlight connected objects with a satisfactory precision for most medical and industrial applications. The time needed for this "manual" processing, however, makes the procedure slow and costly. Moreover, the biological visual system has intrinsic limitations, as two-dimensionality (partially compensated by stereo vision) and sensitivity to illumination gradients. At the same time, a significant variety of visual tasks in applications goes far beyond the traditional two-dimensional visual perception. Examples are 3D medical imaging devices, 3D geophysical analysis and systems for non-destructive testing of materials. Some features in images, such as oriented structures, textures or structural defects, can appear at different scales and can contain a significant amount of noise of various nature.

Such facts explain the need for systems that can handle a variety of visual problems and have a sufficient degree of universality. While local feature detection can be accomplished by construction of proper local invariants (a non-trivial task on its own), the global groping and segmenting of entire connected structures remains a bottleneck in computer vision.

In this proposal, we focus on a method for detection and segmentation of structures in images, based on multi-scale orientation analysis. Typical application targets are elongated structures like curves and surfaces.

Local orientation filtering, combined with the idea of multi-scale image analysis have proved their significance in image processing. In short, the orientation information in the neighborhood of each image point (pixel) is used to detect and trace structures across finite parts of the image. The size of these neighborhoods is controlled by the scale parameter. In our newly proposed method, we introduce an invertible orientation-dependent transformation that helps restoring the original spatial acuity independently from tlie, scale. Pioneer applications of this technique to 2D scalar images show promising and robust results.

A major problem in the multi-scale orientation approach is its computational complexity. For an image defined in d-dimensional space, we will need another d - 1 angular variables (coordinates) for a full orientation analysis. The local algorithms are simple, but they have to be applied in every image point and at several values of the scale parameter. These facts suggest the possibility of using massively parallel distributed systems for accomplishing the multi-scale orientation-based perceptual grouping and segmentation.

 In living organisms such a distributed system can be associated with the biological neural network. Neurophysiological studies have revealed a lot of data relating visual perception and the activity of the corresponding part of the cortical neural network. Still, it is vastly unknown how to describe the processing power of natural vision in terms of its neural dynamics. In the biological context, our approach will provide an insight into the role of the orientation selectivity, which seems to be crucial for connecting global structures in images.

• Description of the project at ISI
• Powerpoint presentation given at the last user committee meeting (6 MByte!!)

TNO, two companies and three other universities are involved in this project.

Dr.ir. B.M. ter Haar Romeny
Image Sciences Institute
University Hospital Utrecht
Heidelberglaan 100
3584 CX Utrecht
The Netherlands