D.B. Megherbi and Y. Iyassu (USA)
non-linear filtering, character segmentation, character recognition
Optical Character Recognition (OCR) is the process of recognizing printed or hand-written characters from a document image and converting these images into an electronically recognized format such as ASCII. The OCR process involves the following steps. (1) The document is pre-processed to eliminate unnecessary information. (2) Individual characters are extracted from a cluster of other objects. (3) Extracted characters or objects are further processed for recognition based on feature and shape. In this paper, we are mainly concerned with the process of eliminating unnecessary information from the document image and segmentation of individual characters from words. In particular, documents are often distorted when migrated electronically due to the equipment used. Eliminating background noises in images has been the subject of extensive research. Special emphasis ought to be given to the noise that is being introduced when scanning a binary document image into a bitmap image. We propose a morphological operator non linear-filter based method to eliminate noise from document images. In particular, we show in this paper that a series of four proposed masks are used for this purpose; two are diagonal (one leaning right and other left), one is vertical, and one horizontal. We show that first the document is eroded with each mask sequentially which results in four images. The purpose of the erosion stage is to remove noise that resides outside the targeted object. The resulting eroded images contain different component of the targeted object. We show that the union sum of all the images is necessary to assemble the object without the noisy background. As gaps and openings are introduced with the erosion process, an extra step is required to close the existing gaps. We show that further dilation with a relatively small mask is performed to close the object and erosion with the said mask is performed to keep the shape of the image unchanged. Once we have a relatively noise free document image, we can proceed with segmentation process. For this approach, we introduce information that is useful to us referred as Vertical Width (VW). The VW, a function of a column in a given image document, is the number of pixels between the top most pixel and the bottom most pixel of a given column. This can be used as information to identify segmentation points. Once the VW is calculated for a given document image (word), further processing is performed to enhance the accuracy of segmentation points. The steps taken are (1) re-orientation of the image so that it is orthogonal to the vertical axis, (2) a median filter to smoothen the VW function, (3) calculate the first derivative of VW function to identify the change of rate of the VW function. Finally experimental results are shown to illustrate the potential and efficient noise removal and segmentation of the proposed technique as applied on hand written and scanned objects.
Important Links:
Go Back
