Evaluation of Holistic Features for Word Recognition

This work proposes a testing environment for evaluation of holistic methods for word recognition. The main goal is to determine an optimun set of holistic characteristics to represent the months of the year in Brazilian Portuguese language. The holistic characteristics treat the words as single, indivisible entities and attempts to perform recognition from their overall shape, as opposed to their character contents. No segmentation step is needed in holistic recognition systems. 1 Relevance of holistic methods Holistic approaches circumvent problems of segmentation ambiguity and variability of segment shape, because they make no attempt to segment the word into subunits. Instead, they rely on word-level features and matching to determine the identity of the word. 1.1 Choice of database In recognition of handwritten bankchecks, one of the problems is to recognize the date field, that is formed by words (cities and months) and numerical characters. The task of recognizing the name of the month is adequated for holistic application, since that involves a small lexicon. There is consensus that the utility of the holistic approach is either in the small lexicon cases or in the filtering of large lexicons. 2 Description of testing environment The definition of an optimum feature set is important since there is no standard set of representative characteristics for the task of handwriting recognition. The testing environment proposed is presented in Figure 1. The main function is to analyse the different characteristic sets in order to perform their validation. The fundamental parts of the system are: Image Samples of words obtained from specific forms that been digitilized by a scanner dispositive. Figure 2 ilustrates some samples already obtained. A database with 6000 images has been constructed to be used in this work; Pre-processing Different methods applied to the image to perform noise elimination and slant normalization; Characteristics Extractor Different methods that extract the holistic characteristics of the words, such as, estimated word length, number and position of ascenders and descenders; Neural Classificator Neural network (NN) atributes a confidence measure to the image. For each characteristic set, the NN is trained and tested. The best set is the one that maximizes recognition rate. Characteristics Extractor Neural Classificator Image Pre−processing