Fusions of Cnn and Svm Classifiers for Recognizing Handwritten Characters

complies with the regulations of the University and meets the accepted standards with respect to originality and quality. Off-line handwritten character recognition plays an important role on a very large scale in handwriting recognition systems. The ultimate goal of this research field is to let the machine read generic materials written by human beings. In order to achieve this, it is necessary to further improve the recognition accuracy and the reliability of current off-line handwritten character recognition systems. The main contribution of this thesis is to present several ways of integrating the synergy of two superior classifiers: Convolutional Neural Network (CNN) and Support Vector Machine (SVM) which have proven results in recognizing different types of patterns. Two models of new fusions have been investigated. In the hybrid model, CNN works as a trainable feature extractor and SVM performs as a recognizer. It automatically extracts the features from the raw images and generates predictions. In the regular combination model, the CNN classifier is trained with raw images but with normalized sizes, and the SVM classifier is trained with handcrafted features. The reliability for both models has been realized through the introduction of a rejection mechanism. The comparisons between the two proposed models were tested on handwritten digits and handwritten letters in the English language, respectively. For the handwritten digit recognition, experiments were conducted on the well-known MNIST database. Experimental results and comparisons with other works on the same database showed that the best results were achieved by the proposed hybrid model: An error rate of 0.19% IV without rejection (compared to the most recent error rate of 0.35%), and a recognition rate of 94.40% under 100% reliability with rejection (compared to a recognition rate of 91.51% under 100% reliability from other studies to our best knowledge). For the experiments on handwritten letters, the NIST database was applied. Three strategies were adopted in classifying handwritten letters into different types of classes. They are a 26-class problem in uppercases and a 26-class problem in lowercases, a 26-metaclass problem, and a 52-class problem. The recognition results without rejection by the hybrid model for the 26-class problem in uppercases and the 26-class problem in lowercases were 96.2289% and 90.2410%, respectively. The recognition rates without rejection were 92.0744% for the 26-metaclass problem and 70.2408% for the 52-class problem. Results showed that the hybrid model outperformed other single classifiers (SVM, CNN) in our experiments. However, the combination …