Multiscale image decomposition using statistical pattern recognition and eigenanalysis
Eric Graves, J.M. Coggins · 2002
Addresses the problem of segmentation in medical images using multiscale geometric statistical pattern recognition (MGSPR) and applies the method to three images. An artificial visual system (AVS) is proposed which uses multiscale Gaussians and their derivatives to define a feature set that captures the multiscale geometric structure of the image. There are three phases to our method based on MGSPR: training, segmentation and eigenanalysis. The training phase projects manually labeled pixels into a feature space by convolving the training pixels with a set of spatial filters. The distribution of each pixel class is modelled with a Gaussian. The segmentation phase classifies unlabeled pixels based on the models generated by the training phase. In the eigenanalysis phase, optimal filters that are linear combinations of the original filters are calculated. The segmentation procedure is applied to two simulated, but illustrative images, and one medical image of a nerve fiber.>