GROSC: uma proposta de segmentação de caracteres impressos orientada a regiões em níveis de cinza

Abstract

Optical Character Recognition systems (OCR) refer generically to technologies applied to recognize characters from an image file that contains text. It allows, for example, that a scanned sheet of printed text could be converted into an editable text file. These systems have been widespread over the past years, with several commercial versions, including the ones embedded within scanner devices. One of the steps that define the success or failure of this type of system is the correct segmentation of text lines and characters that constitute it. Recently, it has gained importance the correct segmentation of words in the textto assist possible post-processing steps to correct deviations of recognition. This paper presents two different methods for segmentation of texts. The first method, called Segmentation of Lines and Words based Teager energy operator (SLP-TEO), based on the Teager energy operator (TEO), is used in the segmentation of text lines and words. The TEO is applied to the signal abstracted from the linear projection (horizontal or vertical) generated by the binary image of the text. A unique feature of this method is that it can be applied to printed texts or manuscripts, without prior arrangements. Moreover, the same algorithm is used to segmentation of text lines and segmentation of words, no matter if they are printed or handwritten. The adopteddatabase for this project (IAM-Database), widely used in OCR researches, has, for all printed text, handwritten transcripts that were also targeted using SLP-TEO method. The second method, called gray-Region Oriented Segmentation of Characters (gROSC) is applied to grayscale word images for the purpose of character segmentation. This method is based on region oriented methodswhere connected pixels that are visited and labeled. A unique feature of the method is that it should be applied to grayscale images of the segmented words. Moreover, using the Otsu threshold and knowing the gray levels of the image in advance, the method determines the maximum variation between shades of gray that allows or not to visit the neighboring pixel, adding it to the target region of interest. In the adopted database, there are three basic types of characters problems to solve: overlapping characters, connected characters, and font design. The gROSCmethod is applied to all previously segmented words images, and the characters are equally segmented, without previous identification of possible character problems. The experiments and results in all stages of segmentation with both methods are very relevant and demonstrate the efficiency and simplicity of the proposed methods.