🚀 go-pugleaf

Is anybody working in the rather specialised area of automatic defect detection?

I am nearing the end of the first year of my PhD studies, working in collaboration
with a commercial organisation which markets systems to do on-line quality control
for various types of industrial processes. 

The machines generally acquire data by means of a line-scan CCD camera pointing
at the production output which might be, for example, a long continuous sheet
of anodised aluminium which is later cut up and made into printing plates. 
Sometimes the process goes wrong, resulting in faint streaks, chatter marks
(horizontal streaks) or air-bubble type localised defects, and it is these that
the system has to detect on the surface of the product. 

At the moment, all systems on the market which do this job seem to operate
various convolution kernels in hardware, typically one for each class of defect
that is expected. For example, a vertical streak is best enhanced by a long thin
vertical kernel, whereas a smaller, more localised defect might require a small
square one.

This technique has obvious shortcomings! Improvements are the topic of my
research. 

This area seems to fall between the two stools of image enhancement and image
recognition, so it seems to be only lightly covered in image-processing
journals and conferences. 

If you are working on a similar type of problem, perhaps you could e-mail me
at "M.J.Jubb@durham.ac.uk" for an exchange of ideas and techniques. We could
post a summary to the net if there is some interest.