A multi-scale curve smoothing for generalised pattern recognition (MSGPR)

In this paper, we introduce a new method for pattern characterisation in the context of pattern recognition. This method is based on the analysis of the contour of planar objects like the CSS (curvature scale space) method that uses the maxima of the curvature zero-crossing. The input contour is separated into two signals according to its coordinates x and y which are progressively low-pass filtered by decreasing the filter bandwidth. The output signals are then amplified so that the reconstructed contour and the input one have the same scale. By doing so, we detect the intersection points between both contours and then generate the intersection points map that defines features for pattern recognition. Since this method deals only with curve smoothing, it needs only a convolution operation. This way, one can reasonably hope that this method is faster than the CSS one with equivalent performances.