NDTnet 1998 Aug, Vol.3 No.8

Abstract

The reliable detection of flaws is one of the most important tasks in Non Destructive Evaluation (NDE). A special problem in this scope is the crack detection in austenitic welds, which are often evaluated with radiographic methods. Improvements in these methods are necessary, because the human factor has still great influences on the evaluation. For example the differentiation between comparatively harmless undercuts and cracks, which look very similar in the radiograph, is still not solved satisfactorily.

In order to make the job easier for the human inspector and to rise the reliability of correct flaw detection in welds we developed an approach based on a combination of image processing and neuronal networks.
After the digitization of the radiographic film the digitized image is split into several square subimages called ROIs (Regions of Interest). Inside a ROI parameters are determined to describe it as comprehensive as possible. The parameters are estimated with image processing methods. Starting with some classical image processing tools like the morphologic edgefinding operator or the Derivative of Gaussian edge gradient, other parameters are defined by applying a FFT Highpass Filter on the image. Another aspect of describing the ROIs is found with the help of the Discrete Wavelet Transform (DWT). All calculated parameters are combined to a feature vector which is fed into a neuronal network.

There exist various types of networks working well on classification tasks, so the performance of some of them is tested in this investigation. The training and performance testing of the networks was realized with true data provided by a destructive testing of the same welds. The networks have to learn to distinguish three classes. Class number one represents the regions of image background, where no flaw indications can be found. The second class comprises the harmless undercuts, that should not be mixed up with cracks. Cracks are the elements of the third class. It is shown that a correct classification of all classes is possible with a rate of over 90 percent. In order to support the human inspector in his work it is useful to mark the flaws in the digital image of the radiograph. This is done by a final segmentation based on the results of the trained neuronal classificator. This segmentation approach results in a binary image representing the detected flaw. Practical operation can provide support to the human inspectors and lead to improved results in flaw detection and crack identification compared to classical film inspection on the light box.

Abstract Source:
Book of Abstracts, 7th European Conference on Non-Destructive Testing, 26-29 May 1998, ISBN: 87-986898-0-00

Full-Text Source:
Proceedings of the 7th European Conference on Non-Destructive Testing, 26-29 May 1998, ISBN:

Publication Contact:
7th ECNDT, Park Alle 345, DK-2605 Broendby, Denmark, Fax: +45 46 26 70 11, Email: 101373.3414@compuserve.com