Methods for improving the feature representations of 2-D image patterns
When pattern recognition from 2-D images is implemented using a template matching scheme, several problems are encountered in generating a satisfactory feature representation of each image pattern. This dissertation proposes improved solutions to three of these problems: (1) feature extraction that is invariant to in-plane rotation and uniform radial scaling variations of the image patterns, (2) removal of background shadows from 2-D image patterns, and (3) locating image patterns prior to classifying them. The performance level of the proposed improvements is illustrated using several examples, and the scale and rotation invariant algorithms are tested using Monte Carlo simulations. The experimental results indicate an improvement in the quality of features extracted from shadow removed images, and they show good invariance to scale and rotation variations. In addition, the spectral similarity feature extraction algorithm distinguishes well between out-of-class patterns. The pattern locating algorithm performs well in locating image patterns resulting in a large reduction in the number of features extracted from an image over conventional local window feature extraction techniques.