ha rm a

Proceedings of the International Conference on Cognition and Recognition

Skew Angle Detection and Correction of Hand Written Gurmukhi Words using Historate Method Dharamveer Sharma1, Sarika Gupta2, Prashant Beri3 1

Ve er S

Lecturer, Department of Computer Science, Punjabi University, Patiala, India [email protected] 2 Student, M. Tech., Department of Computer Science, Punjabi University, Patiala, India [email protected] 3 Student, M. Tech., Department of Computer Science, Punjabi University, Patiala, India [email protected]

Abstract

1. INTRODUCTION

m

Skew angle detection and correction is an indispensable task for handwritten word recognition system. In this paper we describe a robust technique for skew detection and correction of handwritten words in Gurmukhi script, as the existing algorithms can not be effectively applied on the handwritten words in Gurmukhi script. According to this technique, words having minimum variation in their pixel density according to their aspect ratio are skew free. Most characters in Gurmukhi script have horizontal line at the top called headline usually connect to form a single word. Here we are making use of this headline for optimization of skew correction.

ara

There are many proposed approaches as alternatives for skew angle detection of document images. All the approaches require a rich text area to be present in order to work properly. Rich text areas possess a wellknown characteristic structure, one or more (separate) lines of printed or handwritten words, sharing a common direction. Main approaches for skew detection include: 1) Hough transform, 2) Nearest neighbor, 3) Correlation method 4) Projection profile 5) Fourier method.

Dh

Hough transform is a popular method for skew detection [3], [7], [9], [10] and [4]. It is capable of locating fragmented lines in a binary image. Therefore given a group of black pixels, one can easily find a line having maximum number of these black pixels. Given a binary image with a rich text area, the detected lines will most probably go along the whole middle zone of the textual lines. Hence these lines have approximately the same skewness as the reference lines of the text, which define the skewness of the whole page. Whenever Hough transform is used, there is always a tradeoff between accuracy and speed. It

Proceedings of the International Conference on Cognition and Recognition

ha rm a

maps points in Cartesian space (x,y) to sinusoidal curves in (?,?) space via transformation. Relation between them is given as: ?=xcos? + ysin? Each time a sinusoidal curve intersects another at a particular value of ? and ?, the likely hood increases that line corresponding to that ?, ? co-ordinate value is present in the original image, where value of ? corresponding to highest number of counts gives the skew angle, so it can be seen that more the number of pixels in the image, more is the calculations required. Therefore most methods ([7], [10], [4]) make some efforts by thinning up the word.

Ve er S

A bottom-up technique for skew estimation based on nearest-neighbor clustering is described by Hashizume et al [1]. In this work, the l-nearest-neighbors of all connected components are found, the direction vectors for all nearest-neighbor pairs are accumulated in a histogram, and the histogram peak is found to obtain the document skew angle. This method is based on the fact that inter character and inter word spaces between two consecutive characters or words respectively, are usually smaller than spaces between such neighboring elements that belong to different lines. This method has the advantage that it is not limited to any range of skew angle (range free). But its accuracy can be reduced due to noise, subparts of characters, and between-line connections because only one nearest-neighbor connection is made for each component. Yan [l1] introduced a method for determining the skew angle of an image using cross correlation between lines at a fixed distance. It is based on the observation that the correlation between two vertical lines in an image of a skewed document is maximized in general if one line is shifted relatively to the other line such that the character base line levels for the two lines are coincident. Using this approach, attempts are made to find the best correlation between two or more profiles of the image, taken from vertical (horizontal) cuts of the image.

Dh

ara

m

The projection profile tool was used by many to determine the skew angle of a document image [2],[8]. In this method, a series of projection profiles are obtained at a number of angles close to the expected orientation, and the variation is calculated for each of the profiles. Based on the assumption that most of the black pixels appear in the middle zone of each text line, the profile that gives the maximum variation corresponds to the projection with the best alignment to the text lines, gives the skew angle. The other method given by Postl[2] is based on the Fourier transform. In this method, the direction for which the density of the Fourier space is the largest gives the skew angle. For large images, the Fourier method can be computationally expensive. And very often for a document image, the largest density direction of the Fourier space is on a vertical line and the true density direction may not be the largest. This makes the detection/search for the skewness of a document image difficult for the Fourier method. Lehal and Madan[6] have used physical properties of Gurmukhi script to develop a range free skew detection scheme from -180° to 180° for machine printed Gurmukhi script. A limitation of their algorithm is that it may not give correct results if graphic images or tables are present along with the text. The existing methods can not be effectively applied on handwritten Gurmukhi text because the handwritten words of Gurmukhi text may be variedly skewed with in the same text. As this work is part of the Gurmukhi form processing system, isolated words are used for skew detection and correction. In this paper we propose a simple and fast algorithm for finding the skew angle of an isolated word by calculating its aspect ratio, slope and position of headline. The organization of this paper is as follows: In section 2

Proceedings of the International Conference on Cognition and Recognition

2. PROPERTIES OF GURMUKHI SCRIPT

ha rm a

some properties of Gurmukhi script are presented; in section 3 problem formulation, algorithm, solution and implementation details are described. Experimental result, future work is given in section 4. Lastly conclusion is provided in section 5.

Gurmukhi script is used primarily for Punjabi language. It is spoken by 84 million native speakers and is world’s 14th most popular spoken language. Gurmukhi script alphabet consists of 40 consonants (vianjans) and 12 vowels (laga or matra) and 3 half characters. Character set of Gurmukhi script is: Consonants

Vowels in upper zone

Ve er S

Vowels in upper and middle zone Vowel in middle zone Vowels in lower zone

Half characters in lower zone

ara

m

Properties of Gurmukhi Script are as follows: 1) Writing style is from left to right. 2) Most of the characters have a horizontal line at the upper part. This line, called the headline, connects the characters of a words. 3) A word in Gurmukhi script can be partitioned into three horizontal zones as shown in Figure 1:

Figure 1: Representation of Zones

The upper zone denotes the region above the headline containing vowels. The middle zone represents the area below the headline where consonants and some subparts of vowels are present. The major part of the characters is located in the middle zone. iii. The Lower zone represents the area below middle zone where some vowels and certain half characters lie in the foot of full characters. 4) Unlike the Roman script, where the characters appear in successive columns, in Gurmukhi script two characters may appear in same column. Figure 2 below showing two characters in single column.

Dh

i. ii.

ha rm a

Proceedings of the International Conference on Cognition and Recognition

Figure 2: Presence of multi characters in a column 3. SKEW ESTIMATION FOR HANDWRITTEN GURMUKHI SCRIPT 3.1 Problem Formulation

Ve er S

Let A={a(i,j)} denotes an image of size M X N where a(i,j) is a pixel at location (i,j). The skew correction technique is based on optimal estimation of skew angle ? from ?1, ?2, ---, ?m where ?i is the estimation of local skew angle at position i. The policy of optimal estimation is the detection of exact skew angle in the neighborhood of estimated skew angles. The problem starts with calculating aspect ratio, which is followed by identifying the slope of word. Additional problem is presence of uneven headline. Problems in Gurmukhi script largely include: connectivity of character on headline, characters having structural similarity, two or more characters in a word having intersecting minimum bounding rectangles. The words can be classified as follows:

Figure 3: On the basis of slope 3.2 Algorithm

Figure 4: On the basis of Aspect Ratio

Dh

ara

m

Step1: Estimation of Range in which to find the course skew angle: Calculate the Aspect Ratio of the word where Aspect Ratio=width / height If (Aspect Ratio>1) then Rotate word in the original image in +ve direction along x-axis by some angle ? and calculate its height (h1). Then again rotate original image in –ve direction along same axis by angle -? and again calculate the height (h2). If (h1=h2) then No skew in the word and exit. else if (h1>h2) then Slope of the word is –ve w.r.t x, finalangle= -45,incr= -1, else Slope of the word is +ve w.r.t x, finalangle= 45, incr= +1 else Rotate word in the original image in +ve direction along y-axis by some angle say ? and calculate its width (w1). Then again rotate original image in –ve direction along same axis by angle - ? and again calculate the width (w2). If (w1=w2) then No skew in the word and exit. else if (w1>w2) then Angle of the word is –ve w.r.t y, finalangle= -45,incr= -1, else Angle of the word is +ve w.r.t y, final angle= 45, incr= +1 Step2: Estimation of course skew angle

Proceedings of the International Conference on Cognition and Recognition

Ve er S

ha rm a

For ? =0 to finalangle step incr if Aspect Ratio <1 then Rotate the skewed word along y-axis and calculate new width (Wn) if Wn < minwidth then minwidth= Wn, final? = ? else break from loop else Rotate the skewed word along x-axis and calculate new height(Hn) if Hn
For words having aspect ratio<1, width is to be minimized otherwise height is to be minimized according to the slope constraint and then new co-ordinates are calculated. But sometimes minimization may give an approximation of skew angle, so it is rotated in its neighborhood to find exact skew angle i.e. F(? 1,?2 ,--- ? i,--- ? m) = fi(? i | ? i-1) Here F is a function to calculate the minimum height/width at angle ?i, so fi(?i |? i-1) gives height/width at angle ?i, given minimum height/width calculated upto ? i-1 angle. 3.4 Implementation

ara

m

1) If we know slope of the word, only then we can rotate image in appropriate direction. If we are rotating our image in +ve direction according to the slope, we will come to know whether we are rotating our image in right direction or not, which will get confirmed by comparison of heights. 2) For words having Aspect Ratio less than one, width is to be minimized, otherwise height is to be minimized to remove skew. 3) Optimization detects and corrects the cases where lower part of the middle zone does not contain much pixel density with large matra in upper zone. 4) Skew corrected image is denoted by B={b(i,j)} and is obtained by rotating image A at exact skew angle. This process is related to displacement of pixels from one location to another to correct skew angle.

Dh

4. EXPERIMENTAL RESULTS AND FUTURE WORK The Historate method (skew detection technique) proposed in paper has been applied to a large number of document images. The constraint of the Historate technique is that it is only applicable for words having slope in the range (+45, -45), because it had been observed after the removal of the skewness of the lines in a scanned document, skew angles of the words greater or lesser then this range are relatively uncommon.

Proceedings of the International Conference on Cognition and Recognition

ha rm a

5. CONCLUSION In this paper we have discussed an efficient Historate method for estimation and correction of skew angles in the handwritten Gurmukhi words. This method uses optimization techniques while calculating height by traversing vertically in the image to find first and last black pixel starting from either end of the image, thus reducing time required to find height or width of the word. Experimental results show that both accuracy of character recognition and segmentation are improved by skew correction. This is the most optimized algorithm w.r.t accuracy and time. This algorithm can be applied to other Indian scripts as well which are structurally similar to Gurmukhi script. 6. REFERENCES

Dh

ara

m

Ve er S

[1] A. Hashizume, P. S. Yeh, and A. Rosenfeld 1986, “A method of detecting the orientation of aligned components”, Pattern Recognition Letters, Volume 4, pp. 125-132. [2] W. Postl October 1986, “Detection of linear oblique structures and skew scan in digitized documents”, In Proceedings Int. Conf. on Pattern Recognition, pp. 687-689. [3] S. N. Srihari and V. Govindaraju 1989, “Analysis of textual images using the Hough transform”, Machine Vision and Applications, Volume 2, pp. 141-153. [4] B. Yu and A. K. Jain 1996, “A robust and fast skew detection algorithm for generic documents”, Pattern Recognition, 29(10), pp. 1599-1630. [5] A. D. Bagdanov and J. Kanai 1996, “Evaluation of document image skew estimation techniques”, In Proceedings of the SPIE - Document Recognition III, pp. 343-353. [6] G. S. Lehal and S. Madan 1998, “A New Approach to Skew Detection and Correction of Machine Printed Gurmukhi Script”, Proceedings Int. Conf : KBCS 98, Mumbai, pp. 215-223. [7] S. C. Hinds, J. L. Fisher, and D. P. D'Amato June 1990, “A document skew detection method using run-length encoding and the Hough transform”, In Proceedings Int. Conf. on Pattern Recognition, pp. 464-468. [8] H. S. Baird May 1987, “The skew angle of printed documents”, In Proc. Conf. of the Society of Photographic Scientists and Engineers, pp. 14-21. [9] Y. Nakano, Y. Shima, H. Fujisawa, J. Higashino, and M. Fujiwara June 1990, “An algorithm for skew normalization of document image”, In Proceedings Int. Conf. on Pattern Recognition, pp. 8-13. [10] D. S. Le, G. R. Thoma, and H. Wechsler 1994, “Automated page orientation and skew angle detection for binary document images”, Pattern Recognition, 27(10), pp.1325-1344. [l1] H. Yan November 1993, “Skew correction of document images using interline cross-correlation”, Computer Vision, Graphics, and Image Processing: Graphical Models and Image Processing, 55(6), pp.538-543.