Optical methods for 3d-surface measurement are widely
used in industrial applications as e.g. quality control of
car body parts. Next to increasing accuracy of optical
sensors, research effort has to be put into methods for
processing 3d data.

A modified Associative Memory developed at our re-
search group allows for detection of surface defects close
to optical resolution and noise limits, even for deformable
workpieces. With this approach, dents within the mi-
crometer-range in depth can be detected within a single
measurement of a 1m x 1m surface area.
The Associative Memory is based on an Artificical Neu-
ral Network which is trained with good parts, thus con-
taining the allowed part variance in the weights of the
network. After training, a reference part is obtained
by recalling the network with the currently measured
part. The ANN internally performs a series expansion
with adapted base functions (Karhunen-Lo`ve trans-
e
form). This reference part is then used for compari-
son with the measurement, leading to a difference map
which can be further processed for classification and
projected onto the surface for localisation of defects.
Current research focuses on several issues that allow
higher accuracy and easier setup and training of the As-
sociative Memory for the evaluation of complex parts.
Methods for virtual training with parts derived from
CAD data is researched in a joint project with the
Fraunhofer Institute Magdeburg (IFF) and the INB Vi-
sion AG. (S. v. Enzberg -126, T. Lilienblum +49 391
6117305, E. Lilienblum -126)