PROBABILISTIC ESTIMATION OF BRAILLE DOCUMENT PARAMETERS Majid Yoosefi Babadi 1 , Behrooz Nasihatkon 1 , Zohreh Azimifar 1 , Paul Fieguth 2 1 Computer Science and Engineering Department, Shiraz University, Shiraz, Iran 2 Systems Design Engineering, University of Waterloo, Ontario, Canada ABSTRACT Braille is the most common means of reading and study for visu- ally handicapped people. The need for converting Braille documents into a computer-readable format has motivated research into the im- plementation of Braille recognition systems. The most fundamental steps in such recognition are estimating the scaling, spacing, and skewness properties of a given Braille document. In this paper we propose a statistical method for estimating these parameters, where we model an entire Braille image using a parameterized probabil- ity density function and find the maximum-likelihood (ML) esti- mates using Expectation-Maximization (EM) algorithm. The pro- posed method is robust against noise and other image artifacts, and has a modest computational complexity for straightforward execu- tion on ordinary personal computers. Index Terms— Braille, Expectation-Maximization, Line-spacing, Maximum-Likelihood, Skewness. 1. INTRODUCTION Braille is the most popular convention of reading and writing among visually impaired people. Like other documents, Braille documents need to be converted to a computer-readable format so that they can take advantages of digital documents like maintenance, duplication, translation, text-to-speech conversion, etc. Several researches has been carried out on Braille Document Recognition. Some [1, 2, 3] are based on setting up exclusive equip- ments in which image is illuminated from an special direction, then captured by a camera. Braille characters are then recognized using shadows created by the document protrusions (dots). More recent approaches are based on images obtained by a scan- ner. Ritchings et al. [4] proposed an approach addressing both single and double-sided Braille documents. However, besides some pre- sumptions about scaling properties of the document, no treatment for skewness was mentioned. Antonacopoulos et al. [5] carried out a more sophisticated work, solving the scaling problem using a set of heuristic approaches (like histograms) and solving the skewness problem using Hough transform. It is worth noticing that Braille documents differ from other reg- ular documents in that they are very simple and well-structured. This makes it possible to apply higher-performance methods specialized for Braille recognition. The objective of this research is to find an efficient method to determine the skewness in Braille documents as well as the line spacing. The algorithm corrects high degrees of skewness and efficiently separates lines of the Braille document. In the proposed method first a probability density function (PDF) is defined in a parametric form with parameters representing skew- ness, spacing and scale. Model parameters are then estimated in Fig. 1: Examples of Braille characters Fig. 2: An illustration of Braille document parameters, showing the line spacing β, and the vertical offset b a Maximum-Likelihood framework. Since the PDF has a mixture structure, obtaining a closed form solution is not possible. Hence, we proposed to use the iterative method of Expectation-Maximization to solve the ML problem. 2. METHODOLOGY Braille documents include a number of lines, each comprising of a series of characters. As Fig. 1 shows, each character consists of 6 points arranged in a 3×2 mask. Each point can be either on (raised) or off (flat). 2.1. Modeling Without Skewness Assumption Let us ignore document skewness at this point and focus on esti- mating document scale and line-spacing. Scaling properties of the document are modeled using two parameters b, and β. As shown in Fig. 2, the parameter b is the offset of (center of) the first line of the document from the top of the page. The parameter β is the distance between two consecutive lines. Here, one-sided Braille documents are considered. As the first step, the image is preprocessed in order to remove noise and unde- sirable artifacts and is then scaled to a suitable size. After that, the image is thresholded and locations of the black pixels (shadows) are extracted in the form of xi =[xi yi ] T . Hence, the train data is in the form of X = {xi |1 ≤ i ≤ N }, where N is the total number of 2001 978-1-4244-5654-3/09/$26.00 ©2009 IEEE ICIP 2009