Hier finden Sie wissenschaftliche Publikationen aus den Fraunhofer-Instituten.

Neuronale Strukturen zur sichtgestützten Oberflächeninspektion von Objekten in industrieller Umgebung

: König, A.

Darmstadt, 1995
Darmstadt, TH, Diss., 1995
Fraunhofer IITB ( IOSB) ()
anomaly detection; automated visual inspection self-organization; feature selection; multivariate data visualisation; neural network; neurocomputers; novelty filtering; optimization; self-supervision

This thesis contributes to a research effort within the field of automated visual inspection for industrial quality control. In the scope of the project SIOB, that was under grant of the German Federal Ministry of Research and Technology (BMFT), grant number 01 IN 110 B/6 (SIOB), from 01.07.1991 to 31.12.1994, an innovative inspection system architecture was devised and a software prototype was implemented. The objective of the project was to integrate image processing, knowledge processing, neural networks and pattern recognition techniques to form a generic and flexible inspection system that is extremely easy to use and possesses a high degree of autonomy for rapid system configuration towards the goal of Learning by presentation. In this thesis, neural networks are applied for the subtasks classification, feature space reduction, feature generation and visualisation of the inspection process. A qualitative and quantitative comparison of neural networks and conventional methods of p attern recognition was carried out. Special attention was paid to those algorithms that provide excellent performance along with uncomplicated and reliable use. A dynamic radial-basis-function network was developed that fulfilled both requirements. Due to the nonparametric nature of the problem data, neural networks and nonparametric methods of pattern recognition, e.g., k-nearestneighbor-techniques and hypersphere classifiers were best suited for the application. Considerable work in this thesis was put forth into the man/machine-interface, especially the visualisation and manipulation of complete high dimensional sample data sets. Topology and distance preserving mappings were studied for that aim. The investigations showed that conventional mapping algorithms are better suited for the application, than available neural methods, e.g., the Kohonen feature map. A fast geometrical algorithm and a novel topology preserving mapping were developed for visualization purposes. A novel inter a ctive tool for the analysis of data projections was developed for visualization and user interaction. From classical algorithms nonparametric quality measures were derived, that give quantitative assessment of a feature set's quality in terms of overlap and separability. Based on these criteria as cost functions, methods for optimized feature generation and global feature selection were developed and applied. These criteria provide the desired self-supervision capabilities for the inspection system that are prerequisite for system autonomy. The idea of novelty filtering or anomaly detection was applied in this thesis to visual inspection. A novel algorithm for anomaly detection or novelty filtering was developed based on associative memories. This filter is trained by presenting a set of good objects. Blobs are generated for defects on faulty images by novelty filtering. Several architectures for dedicated neurocomputers were developed in this work. Focus on the implementation work w as on associative memory implementation, kNN and hypersphere classifiers, novelty filtering and vector quantization. A very efficient and simple bitserial neuron cell was developed and a chip mit 32 neurons, comprising a very fast minimum search, was implemented. Two prototype systems were completed. The bitserial neuron cell was enhanced for on-chip Kohonen learning and a 128 neuron chip was modeled in VHDI, and synthesized. For high quality Kohonen learning a dedicated signal processor was developed in full-custom design technique and a prototype system was assembled.