Optimum shape representation based on Fisher’s discriminant analysis

Shape recognition is an important part of machine intelligence in both decision making and data processing. A good shape representation in shape recognition should describe the shape in the way that makes it distinguishable from other shapes and be invariant to transform of position, size, angle and skew. More importantly, developing and finding appropriate shape representation are still a challenging problem. In our previous works, the 1-D area representation at various triangle side lengths has been proposed as an affine invariant shape representation. Finding the optimum triangle side length, the best 1-D area representation, needs to conduct through recognition systems experimentally. In this work, Fisherpsilas discriminant analysis is applied to predict the optimum triangle side length instead of obtaining it from the experiment. This method has been evaluated over a number of affine distorted shapes. The predicted optimum triangle side length is compared with the optimum results obtained from simulation through the recognition systems (neural networks and normalized cross-correlation). The results demonstrate that the performance of recognition depends on the discriminant power of shape representation. The higher Fisherpsilas ratio gives the better recognition performance. The best performance is achieved by using shape representation that has the maximum Fisherpsilas ratio.