MULTISTAGE SEMI-AUTOMATIC TEXT IMAGE SEGMENTATION FOR TRAINING SET ACQUISITION IN HANDWRITING RECOGNITION

摘要

In the paper, a complete method of text image segmentation into the images of individual characters is proposed. The ultimate aim of the segmentation process is to prepare a set of correctly labeled character samples that can be used to train the character classifier applied as the component of the handwritten word recognizer. The method proposed consists of two stages. At the first stage, the text image is first divided into lines and then the lines are segmented into words. In this phase, the known spelling representation of the text on the image is used, so as to obtain as many segments as the number of words in the text. The information about the expected width of known words is also utilized. At the second stage, the obtained images of known words are segmented into individual characters. The multiphase procedure is applied. It first segments individual words independently, using the estimates of character widths obtained by the complete text corpus analysis. Then the global text segmentation is elaborated, which maximizes the similarity measures of samples extracted for all alphabet characters. Genetic algorithm is applied in this phase. Finally, the segmentation variants represented by chromosomes in the terminal population of the genetic algorithm are locally refined and the most dissimilar samples in sets corresponding to the alphabet characters are rejected. The experiments conducted showed that the accuracy of handwriting recognition achieved by recognizers trained with the training set obtained with the proposed method is close to the accuracy achievable with the training set prepared by a human expert.