Optimized filters for image feature recognition, e.g. for surface inspection

With these customized filter banks, error classifications with significantly higher selectivity and better detection rates can be achieved. The method can be easily integrated as software into existing systems for surface inspection.

The recognition and extraction of image features, i.e. their general classification, plays a major role in digital image processing. For example, for surface inspection, images are taken of surfaces in order to identify different classes of surface defects. However, objects of the same class may occur in various sizes, which poses a challenge to the con­ventional image processing methods available. Solutions that are actually adapted to the respective image feature would considerably improve the reliability and sta­bility of these analysis methods.

The dyadic and M-channel wavelet filter banks that have been widely used cannot be optimally matched to the dif­fer­ent sizes of the image features due to their integer scal­ing. That means their use is far from satisfactory for all kinds of applications.

The filter process according to the invention was developed within the framework of a research project carried out by the University of Applied Sciences Pforzheim in coopera­tion with the Fraunhofer Institute IOSB and the Karlsruhe Institute of Technology KIT. It uses tailored "rational biorthogonal wavelet filter banks". The design is carried out in two steps: first the most suitable rational scaling factor is determined and then the filter coefficients are matched to the image characteristics. Biorthogonal filters are used to create higher degrees of freedom.
The new analysis method was tested in a series using deflectometry for the detection of defects on specular sur­faces. The method achieved a much higher degree of selectivity in defect classification than conventional meth­ods. The corresponding detection rates are also superior to those of known methods. This is illustrated by the figure showing exemplary test results. The invention concerns this method of analysis as well as its implementation into existing systems. The customized filter banks can be inte­grated into existing systems to optimize feature recogni­tion.

• Significantly higher detection rates and selectivity in the classification of image features
• Filter banks and their components are specifically adapted to characteristic features
• New software can be easily integrated into existing systems and processes

• Quality control (defect detection on surfaces)
• Digital image processing

Thomas Greiner, Tan-Toan Le, Mathias Ziebarth, Michael Heizmann, Multiskalige Oberflächeninspektion mit Wavelets und Deflektometrie, tm-Technisches Messen, Band 83, Heft 11, Seiten 617–627, ISSN (Online) 2196-7113, ISSN (Print) 0171-8096, October 2016, De Gruyter Verlag.
DOI: https://doi.org/10.1515/teme-2015-0047

Tan-Toan Le, Matthias Ziebarth, Thomas Greiner, Michael Heizmann, Systematic Design of Object Shape Matched Wavelet Filter Banks for Defect Detection, 39th International Conference on Telecommunications and Signal Processing, June 2016, Vienna, Austria.
DOI: https://doi.org/10.1109/TSP.2016.7760923

Le, T.-T.; Ziebarth, M.; Greiner, T.; Heizmann, M.: Optimized Size-adaptive Feature Extraction Based on Content-matched Rational Wavelet Filters. - ln: Proceedings of the 22th European Signal Processing Conference (EUSIPCO), Lisbon Portugal, September 2014, pp. 1672-1676. (978-0-9928626-1-9).

T. Le, T. Greiner, M. Ziebarth, M. Heizmann, Inspection of Specular Surfaces using Optimized M-Channel Wavelets, IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), Proceedings, Vancouver, Canada, May 2013.