Shape Classification Models

Abstract

The aim of this project is to compare and contrast the performance of a representative set of boundary based shape classification models using a large and common data set. A range of noisy environments are considered to measure their performances in realistic experimental conditions. The effects on performance, such as the sampling algorithm, model order and classifier are also considered. Curvature and angle measurement based sampling methods have been shown to perform poorly in adverse conditions. The low order complex autoregressive (CAR) and complex partial correlation (CPARCOR) models have been shown to be robust in all noise conditions on even the most complex data sets considered in this project. The more complex model of a spatially-varying AR process has been shown to be more sensitive to noise than the more simplistic linear AR models. A high order spectral AR model was also tested and showed inferior performance to the low order linear AR models. The simple Fourier descriptors (FD) showed the most robust performances of all but it is believed that they may perform less well when the data sets contain many similar shapes. Finally a wavelet-based model is presented and improvements in the model are suggested.