Abstract

The thesis first introduces a new adaptive technique of finger upright reorientation by using the Principle of Coordinate System Rotation . The empirical results demonstrate that reorienting the images of fingers of a hand prior to any feature extraction consistently leads to more stable feature values, regardless of the features measured. Hand shape analysis included Central Moments, Fourier Descriptors and Zernike Moments is characterized based on I-D contour transformation. The main contribution of the thesis is the first to use a genetic algorithm to simultaneously achieve dimensionality reduction and object (hand image) clustering. A novel Cooperative Coevolutionary Clustering Algorithm (COCA) with dynamic clustering and feature selection has been developed to search for a proper number (without prior knowledge of it) of clusters of hand images into these clusters based on a smaller set of new features. In addition to the main contribution of the study, an MSE Extended Fitness Function is presented which is particularly suited to an integrated dynamic clustering space. The proposed design and experimental implementation show that the dimensionality of the clustering space can be cut in half, and the GA evolves an average of 4 clusters with a very low standard deviation of 0.4714. Average hand image misplacement number is 5.8 out of 100 hand images. These results open a new way towards other cooperative co-evolutionary applications, in which 3 or more populations are used to co-evolve solutions and designs consisting of 3 or more loosely coupled subsolutions or modules.