Thai printed character recognition by combining inductive logic programming with backpropagation neural network

Several approaches to Thai printed character recognition have been proposed such as comparing heading of character, backpropagation neural network (BNN), fuzzy logic and syntactic method, etc. This paper presents a new approach that combines two learning algorithms, i.e. inductive logic programming (ILP) and backpropagation neural network. After features of character images are extracted, they are employed to construct examples for training an ILP algorithm to learn rules that define the characters. The learned rules are then used to classify the unseen data. However, since some character image, especially the noisy image, may not exactly match with any rule, we then employ BNN for approximately matching the image with the rules. Experimental results demonstrate that the accuracy of rules learned by ILP without the help of BNN is comparable to other methods. Moreover, combining BNN with ILP achieves higher accuracy than the other methods tested in our experiment