Abstract

Over the past decade, there has been considerable progress in the design of statistical machine learning strategies, including Semi-Supervised Learning (SSL) approaches. However, researchers still have difficulties in applying most of these learning strategies when two or more classes overlap, and/or when each class has a bimodal/multimodal distribution. In this thesis, an efficient, robust, and reliable recognition system with a novel SSL scheme has been developed to overcome overlapping problems between two classes and bimodal distribution within each class. This system was based on the nature of category learning and recognition to enhance the system's performance in relevant applications. In the training procedure, besides the supervised learning strategy, the unsupervised learning approach was applied to retrieve the "extra information" that could not be obtained from the images themselves. This approach was very helpful for the classification between two confusing classes. In this SSL scheme, both the training data and the test data were utilized in the final classification. In this thesis, the design of a promising supervised learning model with advanced state-of-the-art technologies is firstly presented, and a novel rejection measurement for verification of rejected samples, namely Linear Discriminant Analysis Measurement (LDAM), is defined. Experiments on CENPARMI's Hindu-Arabic Handwritten Numeral Database, CENPARMI's Numerals Database, and NIST's Numerals Database were conducted in order to evaluate the efficiency of LDAM. Moreover, multiple verification modules, including a Writing Style Verification (WSV) module, have been developed according to four newly defined error categories. The error categorization was based on the different costs of misclassification. The WSV module has been developed by the unsupervised learning approach to automatically retrieve the person's writing styles so that the rejected samples can be classified and verified accordingly. As a result, errors on CENPARMI's Hindu-Arabic Handwritten Numeral Database (24,784 training samples, 6,199 testing samples) were reduced drastically from 397 to 59, and the final recognition rate of this HAHNR reached 99.05%, a significantly higher rate compared to other experiments on the same database. When the rejection option was applied on this database, the recognition rate, error rate, and reliability were 97.89%, 0.63%, and 99.28%, respectively