Example radioscopic image processing:
multi-channel filtering for defect detection

Defects are characterised in a radioscopic picture by local changes in x-ray collection density which translates to a two dimensional transition in grey level intensity. Image integration is useful in order to reduce noise whilst the image is digitised. A set of particular convolution kernels have been developed by the MSRR Group which are used to enhance defective areas within an image. Some filters are particularly effective at detecting blob-like defects (eg porosities, inclusions) whilst others are more useful for detecting longitudinal type defects such as cracks.

Once an image has been filtered, one of two post processing operators is applied in order to segment suspect defect pixels from the image background. These are :-

1) a neural network trained to detect areas of 'roughness' in an image
2) calculation of the 2nd moment (variance) of a local area of the image.

The neural network method is more sensitive and hence is likely to locate very subtle areas but has a high false alarm rate. The variance operator is cruder, but is also faster (speed increase of x5) and is also more 'tuneable' so it is applicable to most image types (TIG, SAW).Once suspect defect pixels have been located, some blob analysis is carried out to eliminate any areas that are either too small to warrant further investigation, or are likely to be part of the component geometry (ie. the edges of the weld structure, IQI's etc.).

Defect areas are then grouped (using binary morphology) and geometrically analysed to give an indication of their length, area, and position. It is envisaged that this information will be used in future work to give a classification of each defect which is located. Some example image processing is given in the figure above which shows the defect detection procedure for a radioscopic image of an incomplete TIG weld containing several defects.

Further information :-

For further information on this, and other automated inspection work, undertaken by the MSRR group contact: :-

Shaun Lawson

Back to MSRR Homepage

These pages are administered by Maarten Michielen (m.michielen@surrey.ac.uk)
All legal boffs please read this copyright and disclaimer notice.
Last change: January 2000