Project description

Previous work By Neville & Holland [SH1-3] the use of symmetry features as a means of extracting or pre-processing information presented to neural networks. This work has led to a substantial block of research work that needs to be undertaken. The crux of the work would address the following issues. The past research has investigated the normalisation of pattern position and orientation using symmetry features. This has enabled a neural network to perform pattern recognition, invariant to differences in position and orientation; this leads to the following unanswered research questions [#RQ1 to #Q7]:[#RQ1] Could symmetry also be used to normalise the pattern's size? A good starting point for such an investigation is the research of multi-resolution schemes for determining the optimal value of the distance weight function's variance. Previous research has focused on the investigation of reflectional symmetries.[#RQ2] How can rotational and transformation symmetries be used to aid pattern recognition? [#RQ3] Can we utilise symmetry in multi-dimensional patterns? [#RQ4] Can the three symmetry functions of the Reflectional Symmetry Transform be modified to produce a rotational symmetry transform based on the same concepts?Symmetry has been used as a pre-classification step. Traditional programming has been used to present the end product of the transform to a network. An identical value of the distance weight function's Gaussian width parameter, was used for detecting symmetry at every point in the image.