The majority of methods proposed for automatic defect detection in NDT x-ray images generally tend to create problems due to the various other artefacts which typically appear such pictures: notably a high noise content and component structures (edges), which serve to mimic the local properties of a defect, and hence may be detected as false alarms (or false positive, FP, decisions). False positive reactions can not only complicate subsequent processing which must remove them, but may also, and perhaps more significantly, serve to mask actual defect areas.
False negative reactions (ie. the overlooking of a defect area) are also potentially prevalent with many approaches. This is because the sharp intensity changes that characterise a defect are often blurred by the x-ray image process, and the transition between background and defect becomes so slight as to be of similar magnitude to noise spikes.
The problem of finding an ideal defect detection operator is thus : how to produce a very sensitive operator which is not prone to false alarms due to noise or component structure ? This is a very difficult task since the local area properties (eg. the intensity gradient, variance etc.) of, for example, a weld edge very closely resemble those of a true defect. However, inspired by the fact that artificial neural networks (or ANNs) are, theoretically at least, able to simulate any arbitrary mapping, the MSRR group are working on employing such structures in an effort to solve the problem of defect detection without false alarms. A feedforward MLP network (trained using backpropagation) has been optimised on weld inspection data to produce the architecture shown above.
The network is configured with a single output - which registers a maximum when the classification is a defect and a minimum when non-defect. The input stimuli to the network is the set {Fi'j'} of pixel values which comprise the local area population around the pixel to be classified - the number of input elements is therefore (NxN). The optimum value (taking into account both segmentation performance and speed of execution) for N was found to be 9. A classification of each pixel is achieved by moving the sub-window (ie. the input buffer of the ANN)) across the entire image {Fij}, activating the network at each pixel, thus reaching a classification, ie. the output value for the centre pixel P'(i,j), for each image position P(i,j).
The input weightings to the hidden layer are shown in the above diagram as coefficients in spatial filter masks. This is to illustrate the similarity of the behaviour of the network to a applying a sequence of matched filters which are then optimally combined using a decision tree (ie. the remainder of the network). This observation was originally made recently by Nekovei and Sun in their paper "Backpropagation network and its configuration for blood vessel detection in angiograms" in IEEE Trans on Neural Networks, vol.6(1), pp.64-72, 1995.
These pages are administered by Maarten Michielen
(m.michielen@surrey.ac.uk)
All legal boffs please read this copyright and disclaimer notice.
Last change: April 2000
