Pattern Recognition Letters

Transcription

1 Pattern Recognition Letters 32 (2011) Contents lists available at SciVerse ScienceDirect Pattern Recognition Letters journal homepage: An improved contour-based thinning method for character images Soumen Bag a,, Gaurav Harit b a Department of Computer Science and Engineering, Indian Institute of Technology Kharagpur, Kharagpur , India b Department of Computer Science and Engineering, Indian Institute of Technology Rajasthan, Jodhpur , India article info abstract Article history: Received 3 September 2010 Available online 11 September 2011 Communicated by A.M. Alimi Keywords: Ambiguous region Contour Medial axis Shape characteristics Skeleton segment Thinning Digital skeleton of character images, generated by thinning method, has a wide range of applications for shape analysis and classification. But thinning of character images is a big challenge. Removal of spurious strokes or deformities in thinning is a difficult problem. In this paper, we propose a contour-based thinning method used for performing skeletonization of printed noisy isolated character images. In this method, we use shape characteristics of text to get skeleton of nearly same as the true character shape. This approach helps to preserve the local features and true shapes of the character images. As a by-product of our thinning approach, the skeleton also gets segmented into strokes in vector form. Hence further stroke segmentation is not required. Experiment is done on printed English, Bengali, Hindi, and Tamil characters and we obtain much better results comparing with other thinning methods without any post-processing. Ó 2011 Elsevier B.V. All rights reserved. 1. Introduction Thinning of shape has a wide range of application in image processing, machine vision, and pattern recognition. But removal of spurious strokes or shape deformation in thinning is a difficult problem. In the past several decades many thinning algorithms have been developed considering all these problems (Lam et al., 1992; Vincze and K}ovári, 2009). They are broadly classified into two groups: raster scan-based and medial axis-based. Raster scan-based methods are classified into two other categories: sequential and parallel. Sequential algorithms consider one pixel at a time and visit the pixel by raster scanning (Arcelli and di Baja, 1989) or contour following (Arcelli, 1981). Parallel thinning algorithms are based on iterative processing and they consider a pixel for removal based on the results of previous iteration only (Datta and Parui, 1994; Huang et al., 2003; Leung et al., 2000; Zhang and Suen, 1984; Zhu and Zhang, 2008). Many of the raster scan-based character thinning methods can not preserve the local properties or features of the character images properly. As a result, they give slightly shape distorted output. Medial axis-based methods generate a central or median line of pattern directly in one pass without examining all the individual pixels (Martinez-Perez et al., 1987). They also give slight distorted result at some local regions. Here we attempt to minimize local distortions by making use of shape characteristics of text. Next we begin a brief description about few prominent raster scanbased thinning algorithms. Corresponding author. Tel.: addresses: bagsoumen@gmail.com (S. Bag), gharit@iitj.ac.in (G. Harit). Datta and Parui (1994) have proposed a parallel thinning algorithm which preserves connectivity and produces skeleton of one pixel thickness. Each iteration of the algorithm is divided into four sub-iterations. These sub-iterations use two 1 3 and two 3 1 templates for removing boundary pixels along east, west, north, and south directions, respectively. The method uses one 3 3 window to avoid the removal of critical point (which alters the connectivity) and end point (which shortens a leg of the skeleton). Leung et al. (2000) have introduced a contour following thinning method having no sub-iteration. They have used a lookup table to avoid the use of multiple templates for removing boundary pixels. The lookup table has 256 entries and each entry contains three fields (neighbor number, weight number, and connection number). The number of entries depends on the different possibilities of 8-connectivity of a pixel. The determination of the possibility of pixel removal depends on the values of the patterns of the spatial contour pixels. To improve the pixel connectivity of thinned character images, Huang et al. (2003) have introduced a new set of templates (three 4 3, one 4 4, and three 3 4). This algorithm considers all possible patterns of 8-connectivity of a pixel (similar to Leung et al. (2000)) and generates template-based elimination rule for deleting boundary pixels. The pixel deletion is done based on the number of black pixels in the 8-neighbor connectivity. Additionally, it compensates information loss by integrating the contour and skeleton of pattern. The information loss is detected based on the ratio of the skeleton and contour pixels. If the value is less than a predefined threshold then the thinned image is replaced by the contour image /$ - see front matter Ó 2011 Elsevier B.V. All rights reserved. doi: /j.patrec

2 S. Bag, G. Harit / Pattern Recognition Letters 32 (2011) Fig. 1. Distortions of different thinning methods: (a) input image; (b) thinned image (Red circle indicates distortion at junction and end points; Red arrow indicates spurious strokes at curvature region). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) The main contribution of Zhu and Zhang (2008) method is that it does not consider all pixels equally for performing character thinning. It is based on the substitution of pixels of strokes or curves which are most valuable parts for character recognition. This method uses a set of thinning templates, similar to Datta and Parui (1994), for boundary pixel removal and handling corners or junction points. This method has tried to preserve the shape of character images such as Chinese characters, English alphabets, and numerals after thinning. All the above discussed algorithms perform boundary pixel deletion using various thinning templates or values from lookup tables. Next we discuss medial axis-based thinning algorithm. For performing thinning of regular shape, Martinez-Perez et al. (1987) have introduced a medial axis-based thinning method. This approach does not use thinning templates or lookup tables for performing thinning. They have generated medial points in between two parallel contour segments and repeat the procedure for all remaining parallel contour segments. But they have not applied this concept for curve. Finally, we give a brief description of a thinning method which is used for skeleton simplification. Telea et al. (2004) have introduced a method to simplify skeleton structure of an image by removing few skeleton branches. The simplification is done by analyzing the quasi-stable points of the Bayesian energy function, parameterized by boundary of contour and internal structure of the skeleton. The experimental results show that it gives multi-scale skeleton at various abstract levels. Apart from these common methods, some new techniques (such as graphical method, wavelet, neural networks, critical kernels, and pulse coupled neural network) have been introduced into image thinning. Melhi et al. (2001) have proposed a method for thinning binary handwritten text images by generating graphical representations of words within the image. Tang and You (2003) have introduced a wavelet-based scheme to extract skeleton of ribbon-like shape. You and Tang (2007) have presented a scheme of extracting the skeleton of English and Chinese characters based on wavelet transform. Krishnapuram and Chen (1993) have applied recurrent neural networks to image thinning. Altuwaijri and Bayoumi (1995) have used a self-organizing neural network to thin digital images. Gu et al. (2004) and Shang et al. (2007) have proposed a binary image thinning algorithm by using the autowaves generated by pulse coupled neural network (PCNN). Bertrand and Couprie (2006) have introduced a new 2D parallel thinning algorithm using critical kernels. All the above mentioned algorithms can not retain the shape of the character images at the junction and end points properly and produce spurious strokes at high curvature region (Fig. 1). From that point of view, we address a contour-based thinning method for generating proper skeleton of isolated character images. The algorithm is non iterative and template free. The main advantage is that it uses shape characteristics of text to determine the areas within the image region to stop thinning partially. This approach helps to preserve the local features and true shape of the character. Additionally, it produces a set of vectorized strokes with the thin skeleton as by-product. These resultant skeletons have stronger ability to oppose shape deformation and more convenience to shape analysis and classification. Fig. 2. System architecture of the proposed method.

3 1838 S. Bag, G. Harit / Pattern Recognition Letters 32 (2011) This paper is organized as follows. Section 2 describes the methodology of the proposed method. Section 3 contains the experimental results of the proposed method and comparison with other thinning methods. This paper concludes with some remarks on the proposed method and future work in Section Proposed contour-based thinning algorithm In this section, we describe the proposed thinning strategy based on contour extraction and medial axis generation for text alphabets. Fig. 2 shows the system architecture of the proposed method. The advantages of our proposed technique are: Shape distortion at junction and end points and many of the spurious thinning branches which inevitably occur when applying other rasterized thinning algorithms do not occur in our proposed algorithm. It provides a vectorized output of the thin skeleton. This vectorized output is a collection of strokes. Hence further stroke segmentation is not required. Our approach is most suitable for thinning noisy printed text alphabets. Particularly it gives correct output even in the presence of changing width of the strokes Contour extraction and ordering Given a scanned document page, we binarize it using the Otsu s algorithm (Otsu, 1979). Currently we are working with documents with all text content, hence we identify the alphabets as connected components in binary images. This works for English (Casey and Lecolinet, 1996) and Tamil alphabets (Shanthi and Duraiswamy, 2005); however, for Bengali or Hindi documents the entire word gets identified as a single connected component because of the matra/shiro-rekha which connects the individual characters. For this case we separate out the individual akshara within a word by using the character segmentation methods reported in (Garain and Chaudhuri, 2002; Pal and Datta, 2003) for Bengali and in (Bansal and Sinha, 2002; Ma and Doermann, 2003) for Hindi. Given an isolated alphabet, its boundary contour is extracted. We detect a boundary pixel on the alphabet using a 3 3 mask centered on every black pixel of the character. If there is even a single background (white) pixel within the mask it would imply that the center black pixel is flagged as a boundary pixel. Starting from this first boundary pixel we traverse the boundary using a connected component aggregation algorithm, i.e., we go on aggregating the black pixels which are adjacent and also on the boundary. This produces a vector representation of the boundary contour. The result of contour extraction for Fig. 4(a) is shown in Fig. 4(b). After that, we perform the ordering of the contour pixels which helps us for traversing the contour in a particular direction. This boundary contour is then segmented into straight line or curved segments which we refer to as contour-strokes. a start point s is denoted C s. The other end i.e., the end point is allowed to extend to the adjacent contour pixel x 1, then to the next one x 2 in sequence and so on. Our objective is to find a suitable contour boundary pixel for the end point of C s. For each possible end point location e i, we get a hypothesized contour segment cse i. The orientation h i of this segment is the slope of the line joining the points s and e i. We analyze the change Dh i = h i h i 1. For a sequence of hypothesized end point positions e i =...,x k 1,x k,x k+1,..., we note the trend of Dh i being positive, negative, or zero. We form groups of successive pixels for which Dh i remains positive, negative, or zero. In other words the pixels forming a group are successive and have a common trend for Dh i. For each such group, say G p, which spans pixels, say x p1 to x p2, we compute the total change in orientation, Dh Gp = h p2 h p1. We also note the total number of pixels N Gp in the group G p. Since our objective is to identify a candidate end point e for the hypothesized contour segment C s which attempt to choose a suitable position for e from the available end points of the pixel groups G p s which have been identified with respect to the start point s of the contour segment C s. If Dh Gp > h threshold (value is set to ±10 ) for any group G p, then take the end point of C s as the end point of G p 1.IfG p is the first pixel group from the start point s, then restrict G p to extend up to only that pixel where Dh Gp = h threshold. If N Gp > N threshold (value is set to 0.5 times of the average height of the characters) for any group G p, then count the total number of pixels N s in C s up to the end pixel of G p 1.IfN s > N threshold then take the end point of G p 1 as the end point of C s. In the above formulation G p 1 denotes the group just preceding G p. Once we get the candidate end point for the hypothesized contour segment C s, we do a refinement of its location on the contour. In other words we search for its most appropriate position in the vicinity of the identified candidate position. This search is based on analyzing local variance of orientations in the forward and backward direction from each contour pixel position in the vicinity of the candidate end point. The position where both the forward and the backward variances are minimum/low is taken as the most appropriate position of the end point of C s. The result of contour segmentation for Fig. 4(b) is shown in Fig. 4(c) where the identified contour-strokes are shown in a different color. Each contour-stroke corresponds to a straight line stroke or a curved stroke within the character. Note that our notion of a stroke is different from the one used in handwriting recognition. Given that the character boundary contour has been now segmented into contour-strokes, as the next step we get the medial axis for the character Medial axis generation We obtain the medial axis by finding the pixel pairs hp i ; P 0 i i on the opposite boundary contour partitioned into contour segments (Fig. 3). To obtain the medial axis we need to identify the parallel 2.2. Contour segmentation In this section, we describe our algorithm to convert the vectorized boundary contour into component segments straight lines or curves. The novelty of this algorithm lies in its search-based identification of contour segmentation points. Given a vectorized contour the candidate segmentation points are identified by analyzing the incremental changes in the orientation as pixels are added at one end of a hypothesized contour segment. The orientation of a contour segment is the slope of the line joining its hypothesized end points. A hypothesized contour segment initiated from Fig. 3. Closest points on opposite contour segments.

4 S. Bag, G. Harit / Pattern Recognition Letters 32 (2011) contour-strokes on the boundary contour. While doing so we incorporate the most obvious text-specific knowledge: the distance between the parallel contour-strokes should be small (since the pen-width is generally small), and all the pixels in-between the two parallel contour-strokes should be black i.e., belong to the character itself and not the background. Processing starts from any contour-stroke, say C s. We process the pixels {x 1,x 2,...,x Ns } on the contour segment C s. 1. For a pixel x i on C s compute the local orientation. " # h local x i ¼ 1 2 Avg hðx i ; x iþj Þþ Avg hðx i ; x i j Þ j¼1to5 j¼1to5 where h(x i,x i+j ) refers to the orientation of the line segment joining pixel x i to the jth pixel following down the contour (i.e., forward direction), and h(x i,x i j ) refers to the orientation of the line segment joining pixel x i to the jth pixel preceding up the contour (i.e., backward direction). The averaging operation is denoted by Avg. 2. Compute the direction? x i perpendicular to the local orientation h local x i. 3. Starting from pixel x i on C s, traverse pixels along the direction of? x i till a border pixel (not on contour C s ) is reached. Let this border pixel be denoted as x 0 i. 4. Compute the local orientation h x 0 at the pixel x 0 local i i. 5. Compute the distance d xi x 0 between the points x i i and x 0 i. 6. If h local x i is nearly same as h x 0 (differ by ±10 ) and d xi x 0 < n local i i (value is set to 1.4 times of pen width), then mark the mid point of the line joining x i and x 0 i as a pixel on the medial axis. If the two conditions are not satisfied then the medial pixel is not marked since this is likely to be the junction point of two or more pen strokes, hence the ambiguity has to be resolved later. 7. Flag off the pixels x i and x 0 i as been processed and move to the pixel x i+1 and repeat the above steps. The above steps are repeated for all contour strokes. For each contour stroke we consider only the pixels which have not been flagged as already processed. The output of this step is given in Fig. 4(d). We see from the result that the medial axis of the character has come up as skeleton segments. These skeleton segments now need to be extrapolated to join the neighboring segments while ensuring that the extrapolation is within the character region Extrapolating skeleton segments in ambiguous regions We consider the skeleton segments of a character as forming the set S¼ S s 1 ; Se 1 ; S s 2 ; Se 2 ;...; S s m ; Se m, where S s i and S e i denote the start and end points of the ith skeleton segment. The extreme points now need to be extended so that the neighboring skeleton segments can be joined together. The steps to identify the neighboring skeleton segments are as follows: 1. For each extreme point of a skeleton segment find the close extreme points belonging to other skeleton segments. For this purpose we consider that the distance between the two extreme points should be less than e (value is set to 0.45 times of the average height of characters) and that the line joining the two extreme points should be within the character region. 2. If a given extreme point has two or more extreme points of other segments close enough then we make a proximity set, e.g., P¼ S s a ; Ss b ; Se c would indicate proximal extreme point of three skeleton segments a, b, and c. Each proximity set would correspond to extreme points which belong to the junction region belonging to multiple pen strokes. Identifying the medial axis points in the junction region of pen strokes gives skeleton points which tend to distort the true shape of the character. Our approach avoids identifying the medial axis points in such ambiguous regions and instead tries to extrapolate the extreme points of the skeleton segments in a proximity group such that the result would be very close to the true shape of the character. This is the major contribution of this work. The steps for extrapolation are as follows: 1. In a given proximity set P if any two skeleton segments have the orientation differ by ±10 then any one of the two close extreme points is extended to meet the other one. 2. If there is a skeleton segment whose orientation does not match with any other segment in the same proximity set, then its extreme point is extended till it meets the skeleton segment (extrapolated from other segments) or a junction (formed by extrapolating from other segments in P) on the skeleton segment or to the extreme point of the other skeleton segment. Of the three cases mentioned which one applies depends on the number of extreme points which are part of a proximity set. The result of extrapolation is showed in Fig. 4(e). Note that extrapolation result is almost the true shape of the character and does not produce spurious segments. Our approach gives correct thinning results even for very thick strokes. 3. Experimental results and discussion We collected characters from several heterogeneous printed documents in different scripts, such as English, Bengali, Hindi, and Tamil. The number of characters in the testing set is 2315 (450 for English, 690 for Bengali, 652 for Hindi, and 523 for Tamil). All the characters were collected in a systematic manner from printed pages scanned on a HP scanjet 5590 scanner at 300 dpi. We tested our proposed method on a randomly selected set of character images of different scripts. We find our approach is robust, and good results are achieved for characters of different scripts. Some results are shown in Fig. 5. There are two outputs of the algorithm: the set of vectorized skeleton segments (Fig. 5(d)) and the final skeleton image (Fig. 5(e)) formed by performing the extrapolation of the skeleton segments. All the programs are written in C++ using OpenCV 2.0 in the UNIX platform Comparison of experimental results with other thinning methods Fig. 4. Thinning results of printed character images: (a) Input image; (b) contour image; (c) segmented contour-strokes of different colors; (d) skeleton segments; (e) skeleton image. We applied Datta and Parui (1994), Huang et al. (2003), and Telea et al. (2004) thinning methods on our own dataset, and

5 1840 S. Bag, G. Harit / Pattern Recognition Letters 32 (2011) Fig. 5. Experimental results: (a) input image; (b) contour image; (c) segmented contour-strokes of different colors; (d) skeleton segments; (e) skeleton image. Fig. 6. Comparison of experimental results among different thinning methods: (a) printed character; (b) Datta and Parui; (c) Huang et al.; (d) Telea et al.; (e) proposed method. compared the results with our proposed algorithm. Fig. 6 shows a comparison of experimental results among these four thinning methods. For quantitative comparison of experimental results, we introduce three types of thinning distortions: (1) shape distortion at junction and end points, (2) spurious strokes at high curvature regions, and (3) recession of stroke length. We formulate simple measures for quantifying the amount of different types of distortions present in the thinned characters. The shape distortion at junction points is measured as a fraction of the number of junction points which were distorted to the total number of junction points in all the thinned characters obtained using the method. A junction point is considered as distorted if the branches emanating from the junction point have local deformations just near the junction point. The distorted branches are identified by manually observing the thinned characters obtained by applying the particular thinning method. Shape distortion at junction points # distorted junction points ¼ total # junction points In the same way we quantify the shape distortion at the end points as a fraction of the number of end points which have shape distortion of the stroke to the total number of end points for all thinning images in the experimental data set. We quantify the distortion due to spurious strokes as the fraction of the number of spurious strokes at high curvature regions to the total number of strokes. ð1þ # spurious strokes at high curvature regions Spurious strokes measure ¼ total # strokes ð2þ Some thinning methods yield thinned characters having stroke length shorter than that of the same stroke in the given image. We measure the recession of stroke length as follows: Recession of stroke length ¼ X p Y p X p where X p represents total number of object pixels of the thinned image generated by Datta and Parui method and Y p denotes the total number of object pixels in the thinned image as produced by any other method. We observed that Datta and Parui method does not lead to any recession of stroke lengths. So, this type of distortion has been formulated with reference to the thinned image generated by Datta and Parui method. We see that Datta and Parui and Huang et al. thinning results suffer from high shape distortion at the junction and end points and produce large number of unwanted strokes which are not acceptable for stroke segmentation and character recognition. Telea et al. (2004) thinning results suffer from recession of skeleton length and distortion at junction points. Table 1 shows a comparative study of different thinning distortions among these three thinning methods and our proposed method. In this table, we have used four different labels (i.e., High, Medium, Low, and No) to represent the ranges of quantitative values (in %) of different thinning distortions. The ranges are: High: %, Medium: 31 70%, Low: 1 30%, and No: 0%. Finally, we conclude that our method is ð3þ

6 S. Bag, G. Harit / Pattern Recognition Letters 32 (2011) Table 1 Comparative study of different thinning distortions among four thinning methods. Method Shape distortion at Spurious Junction End strokes points points Recession of stroke length Datta and Parui High High High No Huang et al. High Medium High No Telea et al. Medium No No Medium Proposed method No No No No et al., and Telea et al. All skeletonization algorithms are written in C++ and run in a Intel Core 2 Duo 2.20 GHz computer. Table 2 shows a comparison of the average computational times (in ms) taken to thin a character using these four thinning methods. We see that the proposed method is much faster than other three methods. This is because the proposed approach is non-iterative and template free. It provides a direct strategy to find the medial axis, while other methods determine the medial axis by removing the boundary using different templates and lookup tables. 4. Conclusion This paper improves the performance of existent character thinning algorithms. The main challenge of thinning character images is to preserve the shape of characters after thinning. The proposed algorithm avoids shape distortion by detecting the ambiguous regions and avoiding the identification of medial axis points within these regions. The resultant skeleton maintains the pixel connectivity and is very close to the medial axis. Additionally, our algorithm provides strokes of character images as an intermediate result during the thinning process. We have compared our experimental results with three other thinning algorithms and have observed better performance compared to all other algorithms. The proposed thinning approach has a good potential for applications to improve the performance of shape analysis and classification problems. In future, we shall use this method as an important preprocessing step for designing Indian language OCR systems. References Fig. 7. Comparison of experimental results with You and Tang method: (a) input image; (b) You and Tang (2007) method; (c) proposed method. Table 2 Average times in milliseconds taken to thin characters using four different thinning methods. Language Datta and Parui Huang et al. Telea et al. Proposed method English Bengali Hindi Tamil free from all these thinning distortions and gives much better results for thinning character images by maintaining all the basic thinning properties as well. The major improvement is that the deformation of the skeletal structure at the junction and end points of the skeletal branches is not there with our results. We compared our method with a wavelet-based skeletonization method proposed by You and Tang (2007). In their method English and Chinese character sets are used for experimental purpose. As our method is language specific and we tested our method on English and other three Indian languages, so we took the input images and skeleton images of English alphabets, numerals, and words from their reported dataset for experimental purpose. We observe that the performance of our method is good and comparable with their method. Fig. 7 shows the comparison of experimental results in-between our proposed method and You Tang method Comparison of computational time We compared the computational time of the proposed method with three other thinning methods, i,e., Datta and Parui, Huang Altuwaijri, M., Bayoumi, M., A new thinning algorithm for Arabic characters using self-organizing neural network. In: Proc. IEEE Internat. Symposium on Circuits and Systems, pp Arcelli, C., Pattern thinning by contour tracing. Comput. Graphics Image Process. 17 (2), Arcelli, C., di Baja, G.S., A one-pass two-operation process to detect the skeletal pixels on the 4-distance transform. IEEE Trans. Pattern Anal. Machine Intell. 11 (4), Bansal, V., Sinha, R.M.K., Segmentation of touching and fused Devanagari characters. Pattern Recognition 35, Bertrand, G., Couprie, M., New 2D parallel thinning algorithms based on critical kernels. In: Proc. Internat. Workshop on Combinatorial Image Analysis, pp Casey, R.G., Lecolinet, E., A survey of methods and strategies in character segmentation. IEEE Trans. Pattern Anal. Machine Intell. 18 (7), Datta, A., Parui, S.K., A robust parallel thinning algorithm for binary images. Pattern Recognition 27 (9), Garain, U., Chaudhuri, B.B., Segmentation of touching characters in printed Devnagari and Bangla scripts using fuzzy multifactorial analysis. IEEE Trans. Systems, Man, Cybernet. Part C: Appl. Rev. 32 (4), Gu, X., Yu, D., Zhang, L., Image thinning using pulse coupled neural network. Pattern Recognition Lett. 25, Huang, L., Wan, G., Liu, C., An improved parallel thinning algorithm. In: Proc. Internat. Conf. on Document Analysis and Recognition, vol. 2, pp Krishnapuram, R., Chen, F., Implementation of parallel thinning algorithms using recurrent neural networks. IEEE Trans. Neural Netw. 4, Lam, L., Lee, S.W., Suen, C.Y., Thinning methodologies A comprehensive survey. IEEE Trans. Pattern Anal. Machine Intell. 14 (9), Leung, W., Ng, C.M., Yu, P.C., Contour following parallel thinning for simple binary images. In: Proc. IEEE Internat. Conf. on Systems, Man, and Cybernetics, vol. 3, pp Ma, H., Doermann, D., Adaptive Hindi OCR using generalized hausdorff image comparison. ACM Trans. Asian Language Inform. Process. 2, Martinez-Perez, M.P., Jimenez, J., Navalon, J.L., A thinning algorithm based on contours. Comput. Vision, Graphics. Image Process. 39 (2), Melhi, M., Ipson, S.S., Booth, W., A novel triangulation procedure for thinning hand-written text. Pattern Recognition Lett. 22, Otsu, N., A threshold selection method from gray-level histogram. IEEE Trans. Systems, Man, Cybernet. 9 (1), Pal, U., Datta, S., Segmentation of Bangla unconstrained handwritten text. In: Proc. Internat. Conf. on Document Analysis and Recognition, pp Shang, L., Yi, Z., Ji, L., Binary image thinning using autowaves generated by PCNN. Neural Process. Lett. 25, Shanthi, N., Duraiswamy, K., Preprocessing algorithms for the recognition of Tamil handwritten characters. In: Proc. Internat. CALIBER, pp Tang, Y.Y., You, X.G., Skeletonization of ribbon-like shapes based on a new wavelet function. IEEE Trans. Pattern Anal. Machine Intell. 25,

7 1842 S. Bag, G. Harit / Pattern Recognition Letters 32 (2011) Telea, A., Sminchisescu, C., Dickinson, S., Optimal inference for hierarchical skeleton abstraction. In: Proc. Internat. Conf. on Pattern Recognition, vol. 4, pp Vincze, M., K}ovári, B., Comparative survey of thinning algorithms. In: Proc. Internat. Symposium of Hungarian Researchers on Computational Intelligence and Informatics, pp You, X., Tang, Y.Y., Wavelet-based approach to character skeleton. IEEE Trans. Image Process. 16 (5), Zhang, T.Y., Suen, C.Y., A fast parallel algorithm for thinning digital patterns. Comm. ACM 27 (3), Zhu, X., Zhang, S., A shape-adaptive thinning method for binary images. In: Proc. Internat. Conf. on Cyberworlds,