Principles Emerging from the Design of Visual Search Algorithms for Practical Inspection Tasks

This paper is targeted at the design of inspection algorithms involving fast visual search for features, products, and the artefacts that accompany them. The work is based on a number of cereal grain inspection tasks that have presented important problems relating to robustness, sensitivity, accuracy and speed of operation. Techniques covered include multi-level thresholding, morphological operators, mask design, multi-stage processing, and sampling procedures. While in some sense these techniques are simple, they are not trivial to apply, because of the difficulty of specifying algorithms in applications where the image data to be analysed is highly variable, fuzzy, and voluminous. Indeed the difficulty of describing the data sufficiently precisely is both here and in general a highly inhibiting obstacle on the path to specifying and producing optimal algorithms. Nevertheless, it has proved possible to identify a number of generic principles underlying the design of visual search algorithms.