Comparison of Crisp and Fuzzy Character Networks in Handwritten Word Recognition

Experiments involving handwritten word recognition on words taken from images of handwritten address blocks from the United States Postal Service mailstream are described. The word recognition algorithm relies on the use of neural networks at the character level. The neural networks were trained using crisp and fuzzy desired outputs. The fuzzy outputs were defined using a fuzzy k-nearest neighbor algorithm. The crisp networks slightly outperformed the fuzzy networks at the character level but the fuzzy networks outperformed the crisp networks at the word level. INTRODUCTION Handwritten word recognition by computer is a very difficult task. Although considerable research has been performed in character recognition, not much has been done in word recognition. Interest has picked up lately, as can be seen by viewing the contents of the proceedings of recent conferences in these areas [1,2,3,4]. Even in the machine-printed case, word recognition consists of more than just reading the individual characters in the word [5,6,7]. People are able to read words with illegible and ambiguous characters. Many alphabetic characters are ambiguous when read out of context In fact, the same pixel pattern can represent difference characters in different words. Furthermore, multiple characters can look like characters. For example, the "tl" in the image of the word "Portland" in Figure 8 could be an "H". The implication of this is that high recognition rates may not be the ultimate goal of an alphabetic character classifier that is to be used in word reading. Accurate representation of ambiguity is more important. Thus if a certain character in the training set is called a "u" but could be either a "u" or "v", then the desired output of a classifier for that sample should reflect the ambiguity. That is, the notion of fuzzy set membership of characters is very natural and important in the development of character classifiers to be used in word recognition. In this paper, we discuss a handwritten word recognition algorithm that uses neural network classifiers on the character level to attempt to read a word. The algorithm is designed to read words that are amenable to segmentation-based approaches; handprinted and well-formed cursive words. We discuss experiments involving the using of assigning desired outputs in the training of the neural networks using a fuzzy k-nearest neighbor algorithm. We compare the use of such networks with crisply trained networks at the character level and at the word level. Our experimental results indicate that the fuzzy output networks do not perform as well on the character level but perform better at the word level.