Fraunhofer-Gesellschaft

Publica

Hier finden Sie wissenschaftliche Publikationen aus den Fraunhofer-Instituten.

Ultrafast surface inspection system based on image processing with cellular neural networks (CNN)

 
: Blug, Andreas; Jetter, Volker; Strohm, Peter; Höfler, Heinrich

:

AMA Fachverband für Sensorik e.V., Wunstorf:
Sensor + Test Conference 2011. Proceedings. CD-ROM : 06.07.-09.07.2011, Nürnberg, 15th International Conference on Infrared Sensors & Systems (SENSOR 2011); 12th International Conference on Infrared Sensors and Systems (IRS(2) 2011); 10th INternational Confernce on Optical Technologies for Sensing Measurement (OPTO 2011)
Wunstorf: AMA Service, 2011
ISBN: 978-3-9810993-9-3
pp.49-53
Sensor + Test Conference <2011, Nürnberg>
International Conference on Infrared Sensors & Systems (IRS2) <12, 2011, Nürnberg>
International Conference on Optical Technologies for Sensing and Measurement (OPTO) <10, 2011, Nürnberg>
International Conference on Sensors and Measurement Technology (SENSOR) <15, 2011, Nürnberg>
English
Conference Paper
Fraunhofer IPM ()
cellular neural network; CNN; machine vision; image processing; quality assurance; defect detection

Abstract
An important issue in quality assurance is the inspection of rapidly moving surfaces for small defects such as scratches, dents, grooves, or chatter marks. This paper describes an image processing system based on a novel camera technique - the so called Cellular Neural Networks (CNN) - for an industrial application which was not feasible for conventional machine vision systems so far. The application is a test site for a 100% surface control in the production of aluminum wires using chip less shaping processes at feeding rates of 10 m/s. Within such process, rather small surface defects such as grooves with lateral extend of 100 µm in forward direction can cause cracks during further shaping steps and therefore a stop of production. With conventional image processing equipment like line cameras, this system would have to process more than 200,000 lines per second for an optical resolution of 50 µm, which is not feasible with current line cameras. To solve this problem, a camera based on CNN was used. CNN can be considered as a technology to integrate a Single Instruction Multiple Data (SIMD) processor architecture in the electronic circuitry of CMOS cameras. The result is a machine vision system based on a camera where every pixel has its own processor and which enables real time processing of up to 5800 images per second with a resolution of 176 x 144 pixels. The exposure time which limits the motion blur - is 10 µs for a dark field illumination. In comparison to line cameras this corresponds to more than one million lines per second if no overlap is taken into account. With a more realistic overlap of 50 % between two succeeding images, the equivalent line rate is still 500,000 lines per second with a resolution of 144 pixels - which is about five times above the acquisition rates achieved with conventional imaging systems.

: http://publica.fraunhofer.de/documents/N-193116.html