Multiscale contour description for pattern recognition

In this paper, we present a new method and its preliminary results within the context of pattern analysis and recognition. This method is based on the multiscale analysis of a curve and deals with the contour of planar objects. Our method uses a low-pass Gaussian kernel to gradually smooth the contour by decreasing the filter bandwidth. Applying gain control to the smoothed contour stretches it to the same scale as the original one so that both contours intersect. By varying the bandwidth and marking all the intersection points between the smoothed contour and the original one we can generate the Intersection Points Map (IPM) function. The initial results obtained by applying this method to various contours appears to indicate that the IPM function has some very interesting properties within the context of pattern recognition. It is translation and rotation insensitive and also scale change resistant for a large range of scaling. The IPM function generated when applied to noisy contours shows that the method is resistant to noise for a range of noise energy. Applying the features extracted by this method to retrieve a pattern from a database confirms the efficiency of the method.