Date of Award
2015
Document Type
Dissertation
Degree Name
Doctor of Philosophy (PhD)
Department
Computer Engineering
Subject(s)
Supervised learning (Machine learning), Support vector machines., Pattern recognition systems.
Abstract
Margin setting is a novel supervised learning algorithm for pattern recognition. In this research, we studied margin setting in terms of its algorithm, performance, and application. The algorithm and performance were studied from a new perspective: margin. Margin measures the associated distance from the data to the classification boundaries. Margin is an important design parameter that has a direct impact on the performances of margin setting. The performance impact of margin was comprehensively analyzed in two aspects. First, given the fact that margin setting generates a spherical decision boundary in pattern recognition, we proposed a novel approach that combines margin concept in the support vector machine to spherical classification in margin setting. The margin impact analysis in spherical classification was presented using probabilities of miss classification (MC) and over classification (OC). Experiments were carried out through the Monte Carlo method. The result showed that margin setting is a margin classifier whose performance tends to improve with an increased margin within a certain range. Besides, the multi-sphere strategy employed by the margin setting algorithm allowed it to achieve lower probabilities of MC, OC and non-classification than classifiers using a single sphere as its decision boundary. On the other hand, to explore margin impact on the performances of margin setting algorithm, we analyzed and compared with the support vector machine. The margin impact on training performance and generalization were discussed theoretically. Next, experiments were conducted using artificial data sets and benchmark data sets to analyze margin impact as well. The experimental results showed that training performance got worse with an increased margin, and generalization tended to improve with an increased margin within a certain range. As a novel application of margin setting in pattern recognition, we proposed a new learning-based switching filter for suppression of impulse noise in highly corrupted digital images. Margin setting detected the corrupted pixels by classification. A new filter scheme called the Noise-Free Two-Stage (NFTS) filter was employed to restore the image. The superior performance of margin setting indicates it is a powerful supervised learning algorithm that outperforms the support vector machine when applied to salt and pepper noise detection.
Recommended Citation
Wang, Yi, "A study of margin setting in pattern recognition :algorithm, performance and application" (2015). Dissertations. 67.
https://louis.uah.edu/uah-dissertations/67