- Email: [email protected]
Modeling of palm leaf character recognition system using transform based techniques
Pattern Recognition Letters 84 (2016) 29–34
Contents lists available at ScienceDirect
Pattern Recognition Letters journal homepage: www.elsevier.com/locate/patrec
Modeling of palm leaf character recognition system using transform based techniques✩ Narahari Sastry Panyam a,∗, Vijaya Lakshmi T.R. b, RamaKrishnan Krishnan c, Koteswara Rao N.V. a a b c
CBIT, Hyderabad 500075, India MGIT, Hyderabad 500075, India Department of Space, ISRO, Coimbatore 641046, India
a r t i c l e
i n f o
Article history: Received 2 November 2015 Available online 2 August 2016 Keywords: Palm leaf character recognition (PLCR) 3D feature Discrete wavelet transform (DWT) Discrete cosine transform (DCT) Fast Fourier transform (FFT)
a b s t r a c t Optical character recognition (OCR) has been a well-known area of research for last five decades. This is an important application of pattern recognition in image processing. Automatic mail sorting generated interest in the handwritten character recognition (HCR) over a period of time. Palm leaf manuscripts which are very fragile and susceptible to damage caused by insects, contain huge amount of information relating to music, astrology, astronomy etc. Hence it becomes necessary for these manuscripts digitized and stored. These palm leaf manuscripts created interest for the young generation researchers since the last decade. This work exploits a special 3D feature (depth of indentation) which is proportional to the pressure applied by the scriber at that point. This 3D feature is obtained at each of the pixel point of a Telugu palm leaf character. In this work two dimensional Discrete wavelet transform (2-D DWT), two dimensional fast Fourier transform (2-D FFT) and two dimensional discrete cosine transform (2-D DCT) are used for feature extraction. The 3D feature along with the proposed two level transform based technique helps to obtain better recognition accuracy. The best recognition accuracy obtained in this model is 96.4%. © 2016 Elsevier B.V. All rights reserved.
1. Introduction Handwritten character recognition is a computer enabled process that facilitates interpreting and reading legible handwritten inputs provided on old manuscripts like palm leaves, stone inscriptions, etc. [1]. This involves the classification of handwritten characters into appropriate and recognized class [2]. It involves the minute reading of features which define each character. Optical scanning can read the image of a written text from a piece of paper. The most important principle involved in the recognition of handwriting is optical character recognition (OCR). This may be augmented with the help of formatting which will entail correct segmentation into characters and recognize to read the plausible words. Since the advent of digital computers it has become difficult to conduct research on machine simulation for human functions. It is unfortunate that the digital word has yet to devise a powerful
✩ ∗
This paper has been recommended for acceptance by Prof. L. Heutte. Corresponding author. E-mail address: [email protected] (N.S. Panyam).
http://dx.doi.org/10.1016/j.patrec.2016.07.020 0167-8655/© 2016 Elsevier B.V. All rights reserved.
computer that can substitute for the functions of optical senses of the humans. In recent times research focused on machine simulation that will mimic reading as the human being does. However, the recognition accuracy is very low for handwritten characters [3]. The process of character recognition can be structured into two broad categories Offline and Online character recognitions [4]. The Offline character recognition process involves the scanning of the document, digitalization of the data and storing it in a computer for the process of character recognition. Contrary to this, characters are first extracted simultaneously while they are created for online character recognition systems. Obviously, Offline character recognition is not inhibited by external factors like speed and stroke in the movement of writing. The Online system of recognition will be impacted by these features. Pattern recognition techniques till recent times resorted to template or feature based approach [4]. In the template based recognition, the desired pattern for recognition is superposed upon a template pattern that is considered ideal. The extent of correlation between the pattern for recognition and ideal pattern help to decide the degree of match. This was the approach in the earlier OCR systems. In recent times, the feature based approach is clubbed with the template approach for better results [4]. For example, the
30
N.S. Panyam et al. / Pattern Recognition Letters 84 (2016) 29–34
Bangla OCR system incorporated feature based approach for basic characters and template based approach for compound characters [4]. The more refined feature based approach extracts special and specific properties from the test pattern and implies them in a more sophisticated model. This can be two types, [4,5] spatial domain and transform domain. The spatial domain approach depends on deriving features directly from the pixel image representation individual to a pattern. The transform domain is more elaborate. It involves the transformation of image into another space using any one of the transform techniques. Today a modern set of techniques are available to derive any feature from any pattern to a degree of clarity. Nevertheless a prominent graph based work for popular Indian language script is not reported in the literature [4]. Though Sinha and Mahabala [4,5] employed embedded picture language for Devanagari (a script similar to Hindi) OCR studies. India is a diverse nation and rich in literature. As of today there are 33 languages and 20 0 0 dialects of which 22 are recognized under the constitution. Thus, it is a multilingual nation – Assamese, Hindi, Kannada, Bengali, Konkani, Malayalam, Manipuri, Tamil, Telugu, Punjabi, Sanskrit and Urdu [6] are the languages officially recognized. Most of these use 12 scripts in various forms. Hindi, Konkani and Sanskrit languages use Devanagari script, Bangla, Assamese and Manipuri use Bengali script and Punjabi language use the Gurumukhi script. A common trait of these languages is that there is no upper case and lower case letters and they are all descendants of the Brahmi script [7,8]. All of them are phonetic i.e., the symbols used in the script relate to a phoneme or the sound of the language. Another prominent feature of these languages is the vowel is not explicitly written when it follows a consonant in a word, which leads to writing a composite character. Consonants sometimes combine with the vowels or another consonant to form a complex character [7,8]. The most common features of these languages is that they are scripted from left to right except for Urdu which is written from right to left [8]. Chaudhuri and co-workers reported that there is no standard database for Indian languages and the research contributed in this area is low [3]. The ancient literature of Southeast Asia including India is preserved as written on palm leaves [9]. This was the main writing material for many centuries. Popular literature, scientific treaties and history are found since fifth century B.C. [9,10] on palm leaves. One of the best preserved oldest existing documents is recognized to be recorded in the second century A.D. Palm leaves were used as writing material to record art, medicine, astronomy, etc., and were preserved and passed through generations [9,11–15]. Telugu script which is an offshoot of Brahmi script has complex structural characteristics, which are difficult for character recognition [16]. It has 16 vowels and 36 consonants [14]. The challenges for Telugu character recognition are as follows. •
•
•
Compared to English, Indian languages have more number of basic and composite characters. With smaller number of users, languages like Telugu have not attracted equivalent efforts for character recognition. Grammatical operation of sandhi, which literally means “junction” or “union”.
2. Related work Sastry et al. [11,12] have reported that Principal Component Analysis (PCA) using nearest neighborhood classifier (NNC) classification method got very low recognition accuracy in all the three planes “XY”, “YZ” and “XZ” ranging between 37% and 40%. They further published that using 2-D correlation features and NNC classification method the recognition accuracy obtained was 90% in
“YZ” plane of projection. The confusing characters like “Ya”, “Ma”, “Na”, and “Va” in “XY” plane, could give higher recognition accuracy in “YZ” plane of projection using their algorithm. Pal and Chaudhuri [4] devised the completely capable system for the first time to perform OCR of the printed Bangla scripts. Pre-processing in this system starts with skew correction, noise reduction is next and gradation of the images (lines, zones and characters) follow suit. The blend of salient features and pattern similarity is engaged for identification using the combination of feature based and template based approaches. For recognizing the basic characters, feature-based classifier was employed whereas the compound characters, a combination of two or more basic characters, were recognized by employing run-based template approach preceded by grouping the characters. The highest score reported was 96% for printed Bangla scripts. Rajasekaran and Deekshatulu [17] employed a two-stage classification system for printed Telugu characters. The curve tracing approach was used to get the primitive shapes in the first stage. These primitive shapes were further used to obtain the Telugu characters. Patterns were obtained by tracing the distinctive points, after removing the primitive shapes from the character, in the second stage. They reported the recognition accuracy as 95.34% for printed Telugu characters. Salimi and Giveki [18] presented recognition system for Arabic handwritten numerals using singular value decomposition (SVD) classifiers and multi-phase particle swarm optimization (PSO). By partitioning the image horizontally and vertically, features were extracted using 2D PCA algorithm. The partitioned images were classified using two simple PSO rules. The combined PSO rule was used to minimize the fitness function. The handwritten numerals were classified using several classifiers such as multi-layer perceptron (MLP), radial basis function (RBF), SVD and artificial neural fuzzy inference system (ANFIS) in their work. They reported recognition accuracies ranging from 63% to 91.7% using various classifiers mentioned. Murthy and Ramakrishnan [19] proposed a hierarchical approach for Online handwritten Tamil and Kannada datasets. In first stage of classification the non-parametric classifier, NNC, with PCA was used to classify the characters. This stage reduces the number of classes and these reduced classes were fed to multiple classifiers at second stage. In second stage using different features separately, with a dynamic time warping (DTW) classifier, the reduced set of confusion classes were classified. The different features on which the work was carried out were quantized slope, dominant points and quartile features. The reported recognition accuracy for Tamil dataset was 81.1% and 90.2%, using single and multistage classifiers respectively. The recognition accuracy (RA) for Kannada basic characters was reported as 76.5%. The RA improved to 92.2% using multistage classifiers. Aradhya et al. [20,21] proposed character identification using combination of FT and PCA. Aradhya et al. [22] presented recognition of numerals and multilingual text using wavelet transforms and wavelet entropy, respectively. Abaynarh et al. [23] contributed work to recognize Amazigh handwritten characters by extracting Legendre moment based features. They reported recognition rate using neural network classifier as 97.46%. Malik and Dixit [24] adopted wavelet transform and Hop-field network to recognize English alphabetical handwritten characters up to a distortion level of 30%. Further they reported that for a few characters having distortion level between 30% and 40% could be recognized. The characters having distortion level above 40% could not be recognized by their method. Das et al. [25] contributed work on recognizing Bangla isolated handwritten compound characters. Using the combination of topological components like longest run and convex hull based
N.S. Panyam et al. / Pattern Recognition Letters 84 (2016) 29–34
31
Fig. 3. Architecture of multi-level feature system.
Fig. 1. Set up to measure (X,Y) co-ordinates (measuroscope).
Fig. 2. Set up to measure Z co-ordinate (depth).
features, they reported recognition accuracy of 81% by employing SVM classifier. 3. Methodology 3.1. Procedure of 3-D data acquisition The physical geometry of a character can be measured using an instrument called co-ordinate measuring machine (CMM) [9,11– 15]. An operator can control manually or can use a computer to measure the X and Y coordinates using a probe which is fixed to this machine. The probes are optical, mechanical, white light or laser based. The coordinates (X,Y) are measured along the contour of a character like starting point, loops, curves and end point [9,11–15]. From the different positions on the palm leaf, the characters are considered for the measurements in the present work. The (X,Y) coordinate machine set up is shown in Fig. 1. When the scriber uses stylus on the palm leaf, the pressure which is applied at various points along the contour is different. This difference in pressure makes different depth of indentation at various pixel points of a character. This is a 3D feature which significantly increases the recognition ratio. This depth is also termed as a Z dimension. The Z-dimension measuring arrangement is shown in Fig. 2, which has a plunger assembly connected with a dial indicator. It has 0.01 microns least count and is a device used
with a mechanical interaction for measuring the depth. This depth ranges from 10 μm to 150 μm [9,11–15]. Along the boundary of the character at certain important points like curves, start point, end point, loops, the Z-dimension is measured. This is considered to be an important feature. A very thin Teflon needle is fabricated which is fixed to the plunger and is used to measure the Z-value. The difference in the distance between the bottom of a pixel point, which is along the boundary of the character, and the adjacent plain surface of the palm leaf gives the Z-value [9,11–15]. This system of measurement is continued for all the pixel points along the boundary of the character. These (X, Y, Z) coordinates are used to obtain the patterns/characters with the help of Microsoft Excel and are stored in the computer. Using the combination of any two coordinates at a time the patterns/characters are obtained i.e. ‘XY’, ‘YZ’ and ‘XZ’ planes of projection [9,11–15]. The total number of classes of Telugu script considered in the proposed work is 28. For every class four different images were used to train the system and one image was used for testing the recognition accuracy. Hence the total number of training images becomes 28 × 4 = 112 and the testing images become 28 × 1 = 28. The data acquisition system for this work is carried out in a unique way compared to the traditional way of scanning a manuscript, which was explained in detail earlier. Now, to increase the number of samples of the training and testing dataset, all these images were rotated by −5°, −4°, −3°, −2°, −1°, 0°, +1°, +2°, +3°, +4° and +5°. This technique to increase the training and testing dataset was reported by Chaudhuri in IEEE Transactions [3] which is successfully utilized in the present work. Hence, the total number of images for training dataset becomes 112 × 11 = 1232. Similarly the testing set is also increased by using the above technique which becomes 28 × 11 = 308 images. All these images are in “XY” plane of projection. This procedure of generating images/patterns is applied for “YZ” and “XZ” plane of projections also. The system is tested using the 308 test images with 1232 images of training dataset. The recognition accuracy is separately calculated for each plane of projection. 3.2. Architecture of the proposed multi-level feature system The architecture of the proposed multi-level system is shown in Fig. 3. Each character is taken as an input image and the size of the image is normalized to 32 × 32. These images are pre-processed and binarized using a threshold value of 0.7. In the first level 2D fast Fourier transform (FFT) is applied to each pixel of the image and the new transformed pixel value of each pixel becomes a feature of the image. The set of these new features are obtained
32
N.S. Panyam et al. / Pattern Recognition Letters 84 (2016) 29–34
for both training and testing images. The Euclidean distance is calculated between the test image and all the training character images. The database image whichever has the smallest Euclidean distance to the test image is considered to be matched with the image under test. Then the proposed algorithm checks whether the test image is correctly identified or not. The wrongly identified test characters in the first level are again given as input to the second level. In the second level 2-D discrete wavelet transform (DWT) is used instead of 2-D FFT to extract the features for both the training and the wrongly identified test images of the first level. Again Euclidean distance is calculated between the test and training characters. If the image is denoted by f(x,y) of size MXN, then the 2-D FFT of the image f(x,y) is depicted in Eq. (1).
F (u, v ) =
M−1 N−1 vy ux 1 f (x, y )e− j2π ( M + N ) MN
Fig. 4. Architecture of multi-level system using 2D DCT and 2D DWT.
x=0 y=0
for 0 ≤ u ≤ M − 1 and 0 ≤ v ≤ N − 1
(1)
where M is the number of rows and N is the number of columns of the image. The number of rows for the images is 50 i.e., M = 50 and the number of columns is 50 i.e., N = 50. Here (x,y) are the pixel co-ordinates in the space domain and (u,v) are the co-ordinates in the transform domain. In the first level all the characters are preprocessed and classified by taking the 2D FFT of the training and testing character images. The Euclidean distance between the training and the testing characters are calculated. The database image which has the minimum distance with respect to the test image is identified as the test image. This algorithm checks whether the test image is correctly identified. The unidentified test characters from the first level are given as input to the second level to improve the recognition accuracy. The 2-D DWT is used in the second level of the algorithm. The approximation and detail coefficients of the image f(x,y) are depicted in Eqs. (2) and (3) respectively.
1
W ϕ ( j0 , m, n ) = √ MN
M−1 N−1
f (x, y )ϕ j0,m,n (x, y )
(2)
x=0 y=0
M−1 N−1 1 Wψi ( j, m, n ) = √ f (x, y ) MN x=0 y=0
i ψ j,m,n (x, y )
Fig. 5. Some confusing characters in XY plane of projection.
are extracted using Eq. (2) in the modified architecture shown in Fig. 4. The overall recognition accuracies are reported in the next section.
α p αq
C ( p, q ) =
M−1 N−1
f (x, y ) cos
x=0 y=0
π ( 2x + 1 ) p 2M
cos
for 0 ≤ p ≤ M − 1 and 0 ≤ q ≤ N − 1 where
⎧ ⎨ αp = ⎩
1 √ , p=0 M 2/M, 1 ≤ p ≤ M − 1
⎧ ⎨ αq = ⎩
π ( 2 y + 1 )q 2N (4)
1 √ , q=0 N 2/N, 1 ≤ q ≤ N − 1
4. Results
(3)
where ϕ j0,m,n (x, y ) and ψ ij,m,n (x, y ) are scaled and translated Haar basis functions and j0 = 0; N = M = 2 J ; j = 0, 1, 2 …, J − 1; m = n = 0, 1, 2 …, 2j − 1. The column, row and diagonal variations in the image (detail coefficients) are measured by the directional wavelet for i = H, V and D respectively. The average of the image f(x,y), derived using Eq. (2), is used as a feature vector for the characters in the second level of the proposed algorithm to recognize the wrongly identified character images. Further, the algorithm was tested by applying 2-D DCT in the first level and 2-D DWT in the second level. The architecture of the multi-level system using 2D DCT and 2D DWT is shown in Fig. 4. In the first level the characters are classified by extracting features using 2-D DCT. The 2-D DCT is used to transform the pixel value from space domain to transform domain. These transformed pixel values form the feature set which is depicted in Eq. (4). As discussed earlier the Euclidean distance is calculated between the training and testing characters, the minimum the distance between them is considered as the identified character. After verifying whether they are correctly recognized using the proposed algorithm, the unidentified test images are given as input to the second level. In the second level 2-D DWT is used to change the pixel values, to further use them as feature vector. The 2-D DWT features
There are many characters in Telugu which are highly similar and hence their patterns are always confusing for recognition. These similar characters can be grouped [9,11–15] together for further analysis. Some of the confusing Telugu characters “Ba”, “Bha” and “Ru” from the same group having high similarity are shown in Fig. 5. The characters “Ba”, “Bha” and “Ru” in Telugu are very similar and can be grouped in the same group which have similar pattern and this causes decrease in recognition accuracy for any recognition model. As discussed earlier for each character on palm leaf at selected pixel points the three coordinates (X, Y, Z) are measured. Table 1 shows a sample set of X, Y and Z values for the Telugu character “Ba”. The number of points where measurements are needed vary from character to character based on the size of the character. However this problem is solved to a great extent in the proposed method even for “Ba”, “Bha” and “Ru” characters if we consider the respective ‘YZ’ and ‘XZ’ plane of projection images. The patterns obtained for Telugu characters “Ba”, “Bha” and “Ru” in ‘YZ’ and ‘XZ’ plane of projections are shown in Figs. 6 and 7 respectively. It is very clear from these patterns that these patterns are completely different to each other; thereby recognition accuracy would naturally increase for the proposed approach. In this work, 2D transforms like 2D FFT, 2D DCT and 2D DWT are used for extracting features in Level-1 and Level-2 with different combinations. It is found that the overall recognition accuracy
N.S. Panyam et al. / Pattern Recognition Letters 84 (2016) 29–34 Table 1 X, Y and Z coordinate of a palm leaf character “Ba”. Pixel point
X (mm)
Y (mm)
Z (μm)
1 2 3 4 5 6 7 8 9 10 11
0.243 0.213 1.234 1.123 0.669 1.145 1.549 1.678 2.098 2.256 1.324
1.057 1.387 1.711 1.311 0.777 0.567 0.936 0.379 0.841 1.435 2.333
45 53 69 75 76 85 89 82 74 49 99
Table 3 Two-level feature recognition system 2-D DCT followed by 2-D DWT. Plane of projection
Fig. 6. Confusing characters (shown in Fig. 5) in YZ plane of projection.
Fig. 7. Confusing characters (shown in Fig. 5) in XZ plane of projection.
Table 2 Two-level feature recognition system 2-D FFT followed by 2-D DWT. Plane of projection
XY YZ XZ
Recognition accuracy % Level 1 2-D FFT
Level 2 2-D DWT
25 71.4 50
46.4 96.4 85.7
(RA) of the recognition system can be increased by using Level-2 for the mismatched characters of Level-1. In Level-2, using various transforms the features are extracted for both training character images and also the mismatched testing characters of Level1. These features are further used for recognizing the mismatched characters. Sastry et al. [14] published that when 2D FFT was used in the first level for feature extraction, the recognition accuracy obtained is 71.4% in YZ plane of projection. However, when mismatched characters were further tested using 2D DCT features, the recognition accuracy increased to 92.8% in YZ plane of projection. The experimental results for all planes of projection viz., ‘XY’, ‘XZ’ and ‘YZ’ are shown in Table 2. Initially features are extracted at every pixel of the character image using 2-D FFT in the first level. The Euclidean distance is used to classify the characters as described in methodology section. The wrongly identified (test) characters from first level are further classified by using 2-D DWT features in the second level. It is observed from Table 2 that the best recognition accuracy obtained in level 1 is in ‘YZ’ plane of projection i.e., 71.4% and the overall recognition accuracy after level 2, obtained in ‘YZ’ plane of projection is 96.4%.
33
XY YZ XZ
Recognition accuracy % Level 1 2-D DCT
Level 2 2-D DWT
32.14 85.7 67.8
46.4 96.4 85.7
There is a considerable amount of increase in recognition accuracies with a two-level system compared to a single-level system. Also on comparing the level 1 and level 2 results it is observed from Table 2 that in ‘YZ’ plane of projection there is a good improvement in accuracy compared to the other planes of projection. For Telugu characters the maximum variation is in ‘Y’ direction, which is the inherent characteristic of the Telugu script [9,11–15]. The measured depth at different pixel points of a Telugu character (a 3D feature) which is proportional to the pressure applied by the scriber is an important feature for palm leaf character recognition [9,11–15]. Hence the combination of these ‘Y’ and ‘Z’ features has yielded excellent recognition accuracy of 96.4%. If we consider ‘XZ’ projection plane, ‘Z’ component i.e., the 3D feature has contributed for getting 85.7% recognition accuracy. However, in the absence of this 3D feature ‘XY’ plane of projection yielded only 46.4% of recognition accuracy. The two-level system is also tested by extracting features at every pixel point using 2-D DCT in the first level and 2-D DWT in the second level. The experimental results for these feature sets are shown in Table 3. Even for this combination it is observed that ‘YZ’ plane of projection yielded better results. The overall recognition accuracies are same for both the architectures. From Table 3 it is observed that in ‘YZ’ plane of projection in level 1 the best recognition accuracy obtained is 85.7%. On comparing the level 1 accuracies from Tables 2 and 3 it is observed that for all the plane of projections 2-D DCT has given better performance compared to 2-D FFT. The recognition accuracies for the published and the proposed two-level feature recognition systems are shown in Table 4. Sastry et al. [11] have worked on the same database using PCA and reported the best recognition accuracy as 40%. Sastry et al. [12] also published that, by extracting the 2-D correlation features, the best recognition accuracy as 90%. In both published and proposed methods the best accuracies achieved are in ‘YZ’ plane of projection. The highest recognition accuracy achieved is 96.4% with the proposed two-level approach in ‘YZ’ plane of projection. This is in line with the earlier results. This paper describes two sets of combinations and their results in detail. In the first set, 2D FFT is used for feature extraction of all the images (test and training) in Level-1 and a recognition accuracy of 71.4% is achieved in YZ plane of projection. Further all the mismatched characters of Level-1 are considered for Level-2, where 2D DWT is used to extract features. The recognition accuracy obtained at the end of Level-2 is 96.4% in YZ plane of projection. Hence there is an increment in the recognition accuracy of approximately 4% compared to the published results of Sastry et al. [14]. In the second set, 2D DCT is used for feature extraction in Level-1 and a recognition accuracy of 85.7% is achieved in YZ plane of projection. Further all the mismatched characters of Level-1 are considered for Level-2, where 2D DWT is further used to extract features. The recognition accuracy obtained at the end of Level-2 is 96.4% in YZ plane of projection.
34
N.S. Panyam et al. / Pattern Recognition Letters 84 (2016) 29–34 Table 4 Comparison of proposed two-level method with existing methods. Plane of projection
Recognition accuracy % Published methods
XY YZ XZ
Proposed two-level method
PCA [11]
2D correlation [12]
Two-level method (2D FFT + 2D DCT) [14]
40 40 37
54 90 70
32.2 92.8 67.85
5. Conclusions The published work in this area of palm leaf character recognition is very low, as per the literature. The procedure of data acquisition of these palm leaf characters is unique, without any noise and skew problems in all the images. The total number of 1232 training dataset images and 308 testing images were developed in “XY” plane of projection. Similarly, for both “YZ” and “XZ” planes, training and testing images were developed successfully. The recognition accuracy was found out for all the three different planes of projection, i.e., “XY”, “YZ” and “XZ”. The recognition accuracy for “YZ” plane of projection is found to be highest among all the three planes of projection which is in line with the published results. The best recognition accuracy using the proposed technique yielded 96.4% in “YZ” plane of projection. The RA increased from 85.7% to 96.4% in YZ plane of projection using 2-DCT and 2-D DWT in a two level recognition approach. Some of the confusing characters like “Ba”, “Bha”, “Ru” were recognized correctly in the proposed technique, which in turn increases the overall recognition accuracy of palm leaf character recognition. In future, automatic scanning of the palm leaf characters could be developed for data acquisition, decreasing human interface errors. This would further increase the recognition accuracy. Acknowledgments This work is done as a part of AICTE project titled “Design and Development of Palm Leaf Character Recognition System” under RPS (Research Promotion Scheme) vide 20/AICTE/RIFD/RPS(POLICY1)25/2013-14 dated 5th July 2013. Hence the authors express their sincere thanks to the funding agency, AICTE, New Delhi, India for their support and encouragement. References [1] R. Renuka, V. Suganya, B. Arun Kumar, Online hand written character recognition using Digital Pen for static authentication, in: 2014 International Conference on Computer Communication and Informatics (ICCCI), Jan. 2014, pp. 1–5. [2] H. Swethalakshmi, A. Jayaraman, V. Srinivasa Chakravarthy, C. Chandra Sekhar, Online handwritten character recognition of Devanagari and Telugu characters using support vector machines, in: Guy Lorette, Tenth International Workshop on Frontiers in Handwriting Recognition, Suvisoft, La Baule (France), Oct. 2006. [3] U. Bhattacharya, B.B. Chaudhuri, Handwritten numeral databases of Indian scripts and multistage recognition of mixed numerals, IEEE Trans. Pattern Anal. Mach. Intell. 31 (3) (2009) 444–457. [4] U. Pal, B.B. Chaudhuri, Indian script character recognition: a survey, Pattern Recognit. 37 (9) (Sept. 2004) 1887–1899. [5] R.M.K Sinha, H. Mahabala, Machine recognition of Devanagari script, IEEE Trans. Syst. Man Cybern. 9 (1979) 435–441.
46.4 96.4 85.7
[6] Technology Development for Indian Languages, http://tdil.mit.gov.in/. [7] K. Prabhakar, R. Jagadeesh Kannan, A comparative study of optical character recognition for Tamil script, J. Sci. Res. 35 (4) (2009) 570–582. [8] J. Jomy, K.V. Pramod, B. Kannan, Handwritten character recognition of South Indian scripts: a review, Computing Research Repository, 2011, pp. 1–6. Vol. abs/1106.0107. [9] P.N. Sastry, Modeling of palm leaf scribing – analysis using statistical pattern recognition, ECE Department, JNTUH, India, Sept. 2010 Ph.D. dissertation. [10] F. Pinn, R. Vaidyanathan, D. Heppel, Princely states report, J. Indian States Hist. 2 (2) (2001). https://en.wikipedia.org/wiki/History_of_science_and_technology_ in_the_Indian_subcontinent [11] P.N. Sastry, R. Krishnan, B.V.S. Ram, Classification and identification of Telugu handwritten characters extracted from palm leaves using decision tree approach, J. Appl. Eng. Sci. 5 (3)) (2010) 22–32. [12] P.N. Sastry, R. Krishnan, B.V.S. Ram, Telugu character recognition on palm leaves – a three dimensional approach, Technol. Spectr. 2 (3) (2008) 19–26. [13] P.N. Sastry, R. Krishnan, Isolated Telugu palm leaf character recognition using radon transform, a novel approach, in: World Congress on Information and Communication Technologies (WICT), 2012, pp. 795–802. [14] P.N. Sastry, T.R. Vijaya Lakshmi, R. Krishnan, N.V. Rao, Analysis of Telugu palm leaf characters using multi-level recognition approach, J. Appl. Eng. Sci. 10 (20) (2015) 9258–9264. [15] T.R.V. Lakshmi, P.N. Sastry, R. Krishnan, N.V.K. Rao, T.V. Rajinikanth, Analysis of Telugu palm leaf character recognition using 3D feature, in: International Conference on Computational Intelligence and Networks (CINE), Jan. 2015, pp. 36–41. [16] S. Bag, G. Harit, P. Bhowmick, Recognition of Bangla compound characters using structural decomposition, Pattern Recognit. 47 (3)) (March 2014) 1187–1201. [17] S.N.S. Rajasekaran, B.L. Deekshatulu, Recognition of printed Telugu characters, Comput. Graph. Image Process. 6 (4) (1977) 335–360. [18] H. Salimi, D. Giveki, Farsi/Arabic handwritten digit recognition based on ensemble of SVD classifiers and reliable multi-phase PSO combination rule, Int. J. Doc. Anal. Recognit. 16 (4) (2013) 371–386. [19] V.N. Murthy, A.G. Ramakrishnan, Choice of classifiers in hierarchical recognition of online handwritten Kannada and Tamil aksharas, J. Univers. Comput. Sci. 17 (1) (2011) 94–106. [20] V.M. Aradhya, G.H. Kumar, S. Noushath, Multilingual OCR system for South Indian scripts and English documents: an approach based on Fourier transform and principal component analysis, Eng. Appl. Artif. Intell. 21 (2008) 658–668. [21] V.M. Aradhya, G.H. Kumar, Principal component analysis and generalized regression neural networks for efficient character recognition, in: First International Conference on Emerging Trends in Engineering and Technology, 2008, pp. 1170–1174. [22] V.M. Aradhya, G.H. Kumar, S. Noushath, Robust unconstrained handwritten digit recognition using radon transform, in: International Conference on Signal Processing, Communications and Networking, 2007, ICSCN’07, IEEE, 2007, pp. 626–629. [23] M. Abaynarh, H. El Fadili, L. Zenkouar, Enhanced feature extraction of handwritten characters and recognition using artificial neural networks, J. Theor. Appl. Inf. Technol. 72 (3) (2015) 355–365. [24] P. Malik, R. Dixit, Handwritten character recognition using wavelet transform and Hopfield network, in: International Conference on Machine Intelligence Research and Advancement, 2014, pp. 125–129. [25] N. Das, K. Acharya, R. Sarkar, S. Basu, M. Kundu, M. Nasipuri, A benchmark image database of isolated Bangla handwritten compound characters, Int. J. Doc. Anal. Recognit. 17 (4) (2014) 413–431.
Contents lists available at ScienceDirect
Pattern Recognition Letters journal homepage: www.elsevier.com/locate/patrec
Modeling of palm leaf character recognition system using transform based techniques✩ Narahari Sastry Panyam a,∗, Vijaya Lakshmi T.R. b, RamaKrishnan Krishnan c, Koteswara Rao N.V. a a b c
CBIT, Hyderabad 500075, India MGIT, Hyderabad 500075, India Department of Space, ISRO, Coimbatore 641046, India
a r t i c l e
i n f o
Article history: Received 2 November 2015 Available online 2 August 2016 Keywords: Palm leaf character recognition (PLCR) 3D feature Discrete wavelet transform (DWT) Discrete cosine transform (DCT) Fast Fourier transform (FFT)
a b s t r a c t Optical character recognition (OCR) has been a well-known area of research for last five decades. This is an important application of pattern recognition in image processing. Automatic mail sorting generated interest in the handwritten character recognition (HCR) over a period of time. Palm leaf manuscripts which are very fragile and susceptible to damage caused by insects, contain huge amount of information relating to music, astrology, astronomy etc. Hence it becomes necessary for these manuscripts digitized and stored. These palm leaf manuscripts created interest for the young generation researchers since the last decade. This work exploits a special 3D feature (depth of indentation) which is proportional to the pressure applied by the scriber at that point. This 3D feature is obtained at each of the pixel point of a Telugu palm leaf character. In this work two dimensional Discrete wavelet transform (2-D DWT), two dimensional fast Fourier transform (2-D FFT) and two dimensional discrete cosine transform (2-D DCT) are used for feature extraction. The 3D feature along with the proposed two level transform based technique helps to obtain better recognition accuracy. The best recognition accuracy obtained in this model is 96.4%. © 2016 Elsevier B.V. All rights reserved.
1. Introduction Handwritten character recognition is a computer enabled process that facilitates interpreting and reading legible handwritten inputs provided on old manuscripts like palm leaves, stone inscriptions, etc. [1]. This involves the classification of handwritten characters into appropriate and recognized class [2]. It involves the minute reading of features which define each character. Optical scanning can read the image of a written text from a piece of paper. The most important principle involved in the recognition of handwriting is optical character recognition (OCR). This may be augmented with the help of formatting which will entail correct segmentation into characters and recognize to read the plausible words. Since the advent of digital computers it has become difficult to conduct research on machine simulation for human functions. It is unfortunate that the digital word has yet to devise a powerful
✩ ∗
This paper has been recommended for acceptance by Prof. L. Heutte. Corresponding author. E-mail address: [email protected] (N.S. Panyam).
http://dx.doi.org/10.1016/j.patrec.2016.07.020 0167-8655/© 2016 Elsevier B.V. All rights reserved.
computer that can substitute for the functions of optical senses of the humans. In recent times research focused on machine simulation that will mimic reading as the human being does. However, the recognition accuracy is very low for handwritten characters [3]. The process of character recognition can be structured into two broad categories Offline and Online character recognitions [4]. The Offline character recognition process involves the scanning of the document, digitalization of the data and storing it in a computer for the process of character recognition. Contrary to this, characters are first extracted simultaneously while they are created for online character recognition systems. Obviously, Offline character recognition is not inhibited by external factors like speed and stroke in the movement of writing. The Online system of recognition will be impacted by these features. Pattern recognition techniques till recent times resorted to template or feature based approach [4]. In the template based recognition, the desired pattern for recognition is superposed upon a template pattern that is considered ideal. The extent of correlation between the pattern for recognition and ideal pattern help to decide the degree of match. This was the approach in the earlier OCR systems. In recent times, the feature based approach is clubbed with the template approach for better results [4]. For example, the
30
N.S. Panyam et al. / Pattern Recognition Letters 84 (2016) 29–34
Bangla OCR system incorporated feature based approach for basic characters and template based approach for compound characters [4]. The more refined feature based approach extracts special and specific properties from the test pattern and implies them in a more sophisticated model. This can be two types, [4,5] spatial domain and transform domain. The spatial domain approach depends on deriving features directly from the pixel image representation individual to a pattern. The transform domain is more elaborate. It involves the transformation of image into another space using any one of the transform techniques. Today a modern set of techniques are available to derive any feature from any pattern to a degree of clarity. Nevertheless a prominent graph based work for popular Indian language script is not reported in the literature [4]. Though Sinha and Mahabala [4,5] employed embedded picture language for Devanagari (a script similar to Hindi) OCR studies. India is a diverse nation and rich in literature. As of today there are 33 languages and 20 0 0 dialects of which 22 are recognized under the constitution. Thus, it is a multilingual nation – Assamese, Hindi, Kannada, Bengali, Konkani, Malayalam, Manipuri, Tamil, Telugu, Punjabi, Sanskrit and Urdu [6] are the languages officially recognized. Most of these use 12 scripts in various forms. Hindi, Konkani and Sanskrit languages use Devanagari script, Bangla, Assamese and Manipuri use Bengali script and Punjabi language use the Gurumukhi script. A common trait of these languages is that there is no upper case and lower case letters and they are all descendants of the Brahmi script [7,8]. All of them are phonetic i.e., the symbols used in the script relate to a phoneme or the sound of the language. Another prominent feature of these languages is the vowel is not explicitly written when it follows a consonant in a word, which leads to writing a composite character. Consonants sometimes combine with the vowels or another consonant to form a complex character [7,8]. The most common features of these languages is that they are scripted from left to right except for Urdu which is written from right to left [8]. Chaudhuri and co-workers reported that there is no standard database for Indian languages and the research contributed in this area is low [3]. The ancient literature of Southeast Asia including India is preserved as written on palm leaves [9]. This was the main writing material for many centuries. Popular literature, scientific treaties and history are found since fifth century B.C. [9,10] on palm leaves. One of the best preserved oldest existing documents is recognized to be recorded in the second century A.D. Palm leaves were used as writing material to record art, medicine, astronomy, etc., and were preserved and passed through generations [9,11–15]. Telugu script which is an offshoot of Brahmi script has complex structural characteristics, which are difficult for character recognition [16]. It has 16 vowels and 36 consonants [14]. The challenges for Telugu character recognition are as follows. •
•
•
Compared to English, Indian languages have more number of basic and composite characters. With smaller number of users, languages like Telugu have not attracted equivalent efforts for character recognition. Grammatical operation of sandhi, which literally means “junction” or “union”.
2. Related work Sastry et al. [11,12] have reported that Principal Component Analysis (PCA) using nearest neighborhood classifier (NNC) classification method got very low recognition accuracy in all the three planes “XY”, “YZ” and “XZ” ranging between 37% and 40%. They further published that using 2-D correlation features and NNC classification method the recognition accuracy obtained was 90% in
“YZ” plane of projection. The confusing characters like “Ya”, “Ma”, “Na”, and “Va” in “XY” plane, could give higher recognition accuracy in “YZ” plane of projection using their algorithm. Pal and Chaudhuri [4] devised the completely capable system for the first time to perform OCR of the printed Bangla scripts. Pre-processing in this system starts with skew correction, noise reduction is next and gradation of the images (lines, zones and characters) follow suit. The blend of salient features and pattern similarity is engaged for identification using the combination of feature based and template based approaches. For recognizing the basic characters, feature-based classifier was employed whereas the compound characters, a combination of two or more basic characters, were recognized by employing run-based template approach preceded by grouping the characters. The highest score reported was 96% for printed Bangla scripts. Rajasekaran and Deekshatulu [17] employed a two-stage classification system for printed Telugu characters. The curve tracing approach was used to get the primitive shapes in the first stage. These primitive shapes were further used to obtain the Telugu characters. Patterns were obtained by tracing the distinctive points, after removing the primitive shapes from the character, in the second stage. They reported the recognition accuracy as 95.34% for printed Telugu characters. Salimi and Giveki [18] presented recognition system for Arabic handwritten numerals using singular value decomposition (SVD) classifiers and multi-phase particle swarm optimization (PSO). By partitioning the image horizontally and vertically, features were extracted using 2D PCA algorithm. The partitioned images were classified using two simple PSO rules. The combined PSO rule was used to minimize the fitness function. The handwritten numerals were classified using several classifiers such as multi-layer perceptron (MLP), radial basis function (RBF), SVD and artificial neural fuzzy inference system (ANFIS) in their work. They reported recognition accuracies ranging from 63% to 91.7% using various classifiers mentioned. Murthy and Ramakrishnan [19] proposed a hierarchical approach for Online handwritten Tamil and Kannada datasets. In first stage of classification the non-parametric classifier, NNC, with PCA was used to classify the characters. This stage reduces the number of classes and these reduced classes were fed to multiple classifiers at second stage. In second stage using different features separately, with a dynamic time warping (DTW) classifier, the reduced set of confusion classes were classified. The different features on which the work was carried out were quantized slope, dominant points and quartile features. The reported recognition accuracy for Tamil dataset was 81.1% and 90.2%, using single and multistage classifiers respectively. The recognition accuracy (RA) for Kannada basic characters was reported as 76.5%. The RA improved to 92.2% using multistage classifiers. Aradhya et al. [20,21] proposed character identification using combination of FT and PCA. Aradhya et al. [22] presented recognition of numerals and multilingual text using wavelet transforms and wavelet entropy, respectively. Abaynarh et al. [23] contributed work to recognize Amazigh handwritten characters by extracting Legendre moment based features. They reported recognition rate using neural network classifier as 97.46%. Malik and Dixit [24] adopted wavelet transform and Hop-field network to recognize English alphabetical handwritten characters up to a distortion level of 30%. Further they reported that for a few characters having distortion level between 30% and 40% could be recognized. The characters having distortion level above 40% could not be recognized by their method. Das et al. [25] contributed work on recognizing Bangla isolated handwritten compound characters. Using the combination of topological components like longest run and convex hull based
N.S. Panyam et al. / Pattern Recognition Letters 84 (2016) 29–34
31
Fig. 3. Architecture of multi-level feature system.
Fig. 1. Set up to measure (X,Y) co-ordinates (measuroscope).
Fig. 2. Set up to measure Z co-ordinate (depth).
features, they reported recognition accuracy of 81% by employing SVM classifier. 3. Methodology 3.1. Procedure of 3-D data acquisition The physical geometry of a character can be measured using an instrument called co-ordinate measuring machine (CMM) [9,11– 15]. An operator can control manually or can use a computer to measure the X and Y coordinates using a probe which is fixed to this machine. The probes are optical, mechanical, white light or laser based. The coordinates (X,Y) are measured along the contour of a character like starting point, loops, curves and end point [9,11–15]. From the different positions on the palm leaf, the characters are considered for the measurements in the present work. The (X,Y) coordinate machine set up is shown in Fig. 1. When the scriber uses stylus on the palm leaf, the pressure which is applied at various points along the contour is different. This difference in pressure makes different depth of indentation at various pixel points of a character. This is a 3D feature which significantly increases the recognition ratio. This depth is also termed as a Z dimension. The Z-dimension measuring arrangement is shown in Fig. 2, which has a plunger assembly connected with a dial indicator. It has 0.01 microns least count and is a device used
with a mechanical interaction for measuring the depth. This depth ranges from 10 μm to 150 μm [9,11–15]. Along the boundary of the character at certain important points like curves, start point, end point, loops, the Z-dimension is measured. This is considered to be an important feature. A very thin Teflon needle is fabricated which is fixed to the plunger and is used to measure the Z-value. The difference in the distance between the bottom of a pixel point, which is along the boundary of the character, and the adjacent plain surface of the palm leaf gives the Z-value [9,11–15]. This system of measurement is continued for all the pixel points along the boundary of the character. These (X, Y, Z) coordinates are used to obtain the patterns/characters with the help of Microsoft Excel and are stored in the computer. Using the combination of any two coordinates at a time the patterns/characters are obtained i.e. ‘XY’, ‘YZ’ and ‘XZ’ planes of projection [9,11–15]. The total number of classes of Telugu script considered in the proposed work is 28. For every class four different images were used to train the system and one image was used for testing the recognition accuracy. Hence the total number of training images becomes 28 × 4 = 112 and the testing images become 28 × 1 = 28. The data acquisition system for this work is carried out in a unique way compared to the traditional way of scanning a manuscript, which was explained in detail earlier. Now, to increase the number of samples of the training and testing dataset, all these images were rotated by −5°, −4°, −3°, −2°, −1°, 0°, +1°, +2°, +3°, +4° and +5°. This technique to increase the training and testing dataset was reported by Chaudhuri in IEEE Transactions [3] which is successfully utilized in the present work. Hence, the total number of images for training dataset becomes 112 × 11 = 1232. Similarly the testing set is also increased by using the above technique which becomes 28 × 11 = 308 images. All these images are in “XY” plane of projection. This procedure of generating images/patterns is applied for “YZ” and “XZ” plane of projections also. The system is tested using the 308 test images with 1232 images of training dataset. The recognition accuracy is separately calculated for each plane of projection. 3.2. Architecture of the proposed multi-level feature system The architecture of the proposed multi-level system is shown in Fig. 3. Each character is taken as an input image and the size of the image is normalized to 32 × 32. These images are pre-processed and binarized using a threshold value of 0.7. In the first level 2D fast Fourier transform (FFT) is applied to each pixel of the image and the new transformed pixel value of each pixel becomes a feature of the image. The set of these new features are obtained
32
N.S. Panyam et al. / Pattern Recognition Letters 84 (2016) 29–34
for both training and testing images. The Euclidean distance is calculated between the test image and all the training character images. The database image whichever has the smallest Euclidean distance to the test image is considered to be matched with the image under test. Then the proposed algorithm checks whether the test image is correctly identified or not. The wrongly identified test characters in the first level are again given as input to the second level. In the second level 2-D discrete wavelet transform (DWT) is used instead of 2-D FFT to extract the features for both the training and the wrongly identified test images of the first level. Again Euclidean distance is calculated between the test and training characters. If the image is denoted by f(x,y) of size MXN, then the 2-D FFT of the image f(x,y) is depicted in Eq. (1).
F (u, v ) =
M−1 N−1 vy ux 1 f (x, y )e− j2π ( M + N ) MN
Fig. 4. Architecture of multi-level system using 2D DCT and 2D DWT.
x=0 y=0
for 0 ≤ u ≤ M − 1 and 0 ≤ v ≤ N − 1
(1)
where M is the number of rows and N is the number of columns of the image. The number of rows for the images is 50 i.e., M = 50 and the number of columns is 50 i.e., N = 50. Here (x,y) are the pixel co-ordinates in the space domain and (u,v) are the co-ordinates in the transform domain. In the first level all the characters are preprocessed and classified by taking the 2D FFT of the training and testing character images. The Euclidean distance between the training and the testing characters are calculated. The database image which has the minimum distance with respect to the test image is identified as the test image. This algorithm checks whether the test image is correctly identified. The unidentified test characters from the first level are given as input to the second level to improve the recognition accuracy. The 2-D DWT is used in the second level of the algorithm. The approximation and detail coefficients of the image f(x,y) are depicted in Eqs. (2) and (3) respectively.
1
W ϕ ( j0 , m, n ) = √ MN
M−1 N−1
f (x, y )ϕ j0,m,n (x, y )
(2)
x=0 y=0
M−1 N−1 1 Wψi ( j, m, n ) = √ f (x, y ) MN x=0 y=0
i ψ j,m,n (x, y )
Fig. 5. Some confusing characters in XY plane of projection.
are extracted using Eq. (2) in the modified architecture shown in Fig. 4. The overall recognition accuracies are reported in the next section.
α p αq
C ( p, q ) =
M−1 N−1
f (x, y ) cos
x=0 y=0
π ( 2x + 1 ) p 2M
cos
for 0 ≤ p ≤ M − 1 and 0 ≤ q ≤ N − 1 where
⎧ ⎨ αp = ⎩
1 √ , p=0 M 2/M, 1 ≤ p ≤ M − 1
⎧ ⎨ αq = ⎩
π ( 2 y + 1 )q 2N (4)
1 √ , q=0 N 2/N, 1 ≤ q ≤ N − 1
4. Results
(3)
where ϕ j0,m,n (x, y ) and ψ ij,m,n (x, y ) are scaled and translated Haar basis functions and j0 = 0; N = M = 2 J ; j = 0, 1, 2 …, J − 1; m = n = 0, 1, 2 …, 2j − 1. The column, row and diagonal variations in the image (detail coefficients) are measured by the directional wavelet for i = H, V and D respectively. The average of the image f(x,y), derived using Eq. (2), is used as a feature vector for the characters in the second level of the proposed algorithm to recognize the wrongly identified character images. Further, the algorithm was tested by applying 2-D DCT in the first level and 2-D DWT in the second level. The architecture of the multi-level system using 2D DCT and 2D DWT is shown in Fig. 4. In the first level the characters are classified by extracting features using 2-D DCT. The 2-D DCT is used to transform the pixel value from space domain to transform domain. These transformed pixel values form the feature set which is depicted in Eq. (4). As discussed earlier the Euclidean distance is calculated between the training and testing characters, the minimum the distance between them is considered as the identified character. After verifying whether they are correctly recognized using the proposed algorithm, the unidentified test images are given as input to the second level. In the second level 2-D DWT is used to change the pixel values, to further use them as feature vector. The 2-D DWT features
There are many characters in Telugu which are highly similar and hence their patterns are always confusing for recognition. These similar characters can be grouped [9,11–15] together for further analysis. Some of the confusing Telugu characters “Ba”, “Bha” and “Ru” from the same group having high similarity are shown in Fig. 5. The characters “Ba”, “Bha” and “Ru” in Telugu are very similar and can be grouped in the same group which have similar pattern and this causes decrease in recognition accuracy for any recognition model. As discussed earlier for each character on palm leaf at selected pixel points the three coordinates (X, Y, Z) are measured. Table 1 shows a sample set of X, Y and Z values for the Telugu character “Ba”. The number of points where measurements are needed vary from character to character based on the size of the character. However this problem is solved to a great extent in the proposed method even for “Ba”, “Bha” and “Ru” characters if we consider the respective ‘YZ’ and ‘XZ’ plane of projection images. The patterns obtained for Telugu characters “Ba”, “Bha” and “Ru” in ‘YZ’ and ‘XZ’ plane of projections are shown in Figs. 6 and 7 respectively. It is very clear from these patterns that these patterns are completely different to each other; thereby recognition accuracy would naturally increase for the proposed approach. In this work, 2D transforms like 2D FFT, 2D DCT and 2D DWT are used for extracting features in Level-1 and Level-2 with different combinations. It is found that the overall recognition accuracy
N.S. Panyam et al. / Pattern Recognition Letters 84 (2016) 29–34 Table 1 X, Y and Z coordinate of a palm leaf character “Ba”. Pixel point
X (mm)
Y (mm)
Z (μm)
1 2 3 4 5 6 7 8 9 10 11
0.243 0.213 1.234 1.123 0.669 1.145 1.549 1.678 2.098 2.256 1.324
1.057 1.387 1.711 1.311 0.777 0.567 0.936 0.379 0.841 1.435 2.333
45 53 69 75 76 85 89 82 74 49 99
Table 3 Two-level feature recognition system 2-D DCT followed by 2-D DWT. Plane of projection
Fig. 6. Confusing characters (shown in Fig. 5) in YZ plane of projection.
Fig. 7. Confusing characters (shown in Fig. 5) in XZ plane of projection.
Table 2 Two-level feature recognition system 2-D FFT followed by 2-D DWT. Plane of projection
XY YZ XZ
Recognition accuracy % Level 1 2-D FFT
Level 2 2-D DWT
25 71.4 50
46.4 96.4 85.7
(RA) of the recognition system can be increased by using Level-2 for the mismatched characters of Level-1. In Level-2, using various transforms the features are extracted for both training character images and also the mismatched testing characters of Level1. These features are further used for recognizing the mismatched characters. Sastry et al. [14] published that when 2D FFT was used in the first level for feature extraction, the recognition accuracy obtained is 71.4% in YZ plane of projection. However, when mismatched characters were further tested using 2D DCT features, the recognition accuracy increased to 92.8% in YZ plane of projection. The experimental results for all planes of projection viz., ‘XY’, ‘XZ’ and ‘YZ’ are shown in Table 2. Initially features are extracted at every pixel of the character image using 2-D FFT in the first level. The Euclidean distance is used to classify the characters as described in methodology section. The wrongly identified (test) characters from first level are further classified by using 2-D DWT features in the second level. It is observed from Table 2 that the best recognition accuracy obtained in level 1 is in ‘YZ’ plane of projection i.e., 71.4% and the overall recognition accuracy after level 2, obtained in ‘YZ’ plane of projection is 96.4%.
33
XY YZ XZ
Recognition accuracy % Level 1 2-D DCT
Level 2 2-D DWT
32.14 85.7 67.8
46.4 96.4 85.7
There is a considerable amount of increase in recognition accuracies with a two-level system compared to a single-level system. Also on comparing the level 1 and level 2 results it is observed from Table 2 that in ‘YZ’ plane of projection there is a good improvement in accuracy compared to the other planes of projection. For Telugu characters the maximum variation is in ‘Y’ direction, which is the inherent characteristic of the Telugu script [9,11–15]. The measured depth at different pixel points of a Telugu character (a 3D feature) which is proportional to the pressure applied by the scriber is an important feature for palm leaf character recognition [9,11–15]. Hence the combination of these ‘Y’ and ‘Z’ features has yielded excellent recognition accuracy of 96.4%. If we consider ‘XZ’ projection plane, ‘Z’ component i.e., the 3D feature has contributed for getting 85.7% recognition accuracy. However, in the absence of this 3D feature ‘XY’ plane of projection yielded only 46.4% of recognition accuracy. The two-level system is also tested by extracting features at every pixel point using 2-D DCT in the first level and 2-D DWT in the second level. The experimental results for these feature sets are shown in Table 3. Even for this combination it is observed that ‘YZ’ plane of projection yielded better results. The overall recognition accuracies are same for both the architectures. From Table 3 it is observed that in ‘YZ’ plane of projection in level 1 the best recognition accuracy obtained is 85.7%. On comparing the level 1 accuracies from Tables 2 and 3 it is observed that for all the plane of projections 2-D DCT has given better performance compared to 2-D FFT. The recognition accuracies for the published and the proposed two-level feature recognition systems are shown in Table 4. Sastry et al. [11] have worked on the same database using PCA and reported the best recognition accuracy as 40%. Sastry et al. [12] also published that, by extracting the 2-D correlation features, the best recognition accuracy as 90%. In both published and proposed methods the best accuracies achieved are in ‘YZ’ plane of projection. The highest recognition accuracy achieved is 96.4% with the proposed two-level approach in ‘YZ’ plane of projection. This is in line with the earlier results. This paper describes two sets of combinations and their results in detail. In the first set, 2D FFT is used for feature extraction of all the images (test and training) in Level-1 and a recognition accuracy of 71.4% is achieved in YZ plane of projection. Further all the mismatched characters of Level-1 are considered for Level-2, where 2D DWT is used to extract features. The recognition accuracy obtained at the end of Level-2 is 96.4% in YZ plane of projection. Hence there is an increment in the recognition accuracy of approximately 4% compared to the published results of Sastry et al. [14]. In the second set, 2D DCT is used for feature extraction in Level-1 and a recognition accuracy of 85.7% is achieved in YZ plane of projection. Further all the mismatched characters of Level-1 are considered for Level-2, where 2D DWT is further used to extract features. The recognition accuracy obtained at the end of Level-2 is 96.4% in YZ plane of projection.
34
N.S. Panyam et al. / Pattern Recognition Letters 84 (2016) 29–34 Table 4 Comparison of proposed two-level method with existing methods. Plane of projection
Recognition accuracy % Published methods
XY YZ XZ
Proposed two-level method
PCA [11]
2D correlation [12]
Two-level method (2D FFT + 2D DCT) [14]
40 40 37
54 90 70
32.2 92.8 67.85
5. Conclusions The published work in this area of palm leaf character recognition is very low, as per the literature. The procedure of data acquisition of these palm leaf characters is unique, without any noise and skew problems in all the images. The total number of 1232 training dataset images and 308 testing images were developed in “XY” plane of projection. Similarly, for both “YZ” and “XZ” planes, training and testing images were developed successfully. The recognition accuracy was found out for all the three different planes of projection, i.e., “XY”, “YZ” and “XZ”. The recognition accuracy for “YZ” plane of projection is found to be highest among all the three planes of projection which is in line with the published results. The best recognition accuracy using the proposed technique yielded 96.4% in “YZ” plane of projection. The RA increased from 85.7% to 96.4% in YZ plane of projection using 2-DCT and 2-D DWT in a two level recognition approach. Some of the confusing characters like “Ba”, “Bha”, “Ru” were recognized correctly in the proposed technique, which in turn increases the overall recognition accuracy of palm leaf character recognition. In future, automatic scanning of the palm leaf characters could be developed for data acquisition, decreasing human interface errors. This would further increase the recognition accuracy. Acknowledgments This work is done as a part of AICTE project titled “Design and Development of Palm Leaf Character Recognition System” under RPS (Research Promotion Scheme) vide 20/AICTE/RIFD/RPS(POLICY1)25/2013-14 dated 5th July 2013. Hence the authors express their sincere thanks to the funding agency, AICTE, New Delhi, India for their support and encouragement. References [1] R. Renuka, V. Suganya, B. Arun Kumar, Online hand written character recognition using Digital Pen for static authentication, in: 2014 International Conference on Computer Communication and Informatics (ICCCI), Jan. 2014, pp. 1–5. [2] H. Swethalakshmi, A. Jayaraman, V. Srinivasa Chakravarthy, C. Chandra Sekhar, Online handwritten character recognition of Devanagari and Telugu characters using support vector machines, in: Guy Lorette, Tenth International Workshop on Frontiers in Handwriting Recognition, Suvisoft, La Baule (France), Oct. 2006. [3] U. Bhattacharya, B.B. Chaudhuri, Handwritten numeral databases of Indian scripts and multistage recognition of mixed numerals, IEEE Trans. Pattern Anal. Mach. Intell. 31 (3) (2009) 444–457. [4] U. Pal, B.B. Chaudhuri, Indian script character recognition: a survey, Pattern Recognit. 37 (9) (Sept. 2004) 1887–1899. [5] R.M.K Sinha, H. Mahabala, Machine recognition of Devanagari script, IEEE Trans. Syst. Man Cybern. 9 (1979) 435–441.
46.4 96.4 85.7
[6] Technology Development for Indian Languages, http://tdil.mit.gov.in/. [7] K. Prabhakar, R. Jagadeesh Kannan, A comparative study of optical character recognition for Tamil script, J. Sci. Res. 35 (4) (2009) 570–582. [8] J. Jomy, K.V. Pramod, B. Kannan, Handwritten character recognition of South Indian scripts: a review, Computing Research Repository, 2011, pp. 1–6. Vol. abs/1106.0107. [9] P.N. Sastry, Modeling of palm leaf scribing – analysis using statistical pattern recognition, ECE Department, JNTUH, India, Sept. 2010 Ph.D. dissertation. [10] F. Pinn, R. Vaidyanathan, D. Heppel, Princely states report, J. Indian States Hist. 2 (2) (2001). https://en.wikipedia.org/wiki/History_of_science_and_technology_ in_the_Indian_subcontinent [11] P.N. Sastry, R. Krishnan, B.V.S. Ram, Classification and identification of Telugu handwritten characters extracted from palm leaves using decision tree approach, J. Appl. Eng. Sci. 5 (3)) (2010) 22–32. [12] P.N. Sastry, R. Krishnan, B.V.S. Ram, Telugu character recognition on palm leaves – a three dimensional approach, Technol. Spectr. 2 (3) (2008) 19–26. [13] P.N. Sastry, R. Krishnan, Isolated Telugu palm leaf character recognition using radon transform, a novel approach, in: World Congress on Information and Communication Technologies (WICT), 2012, pp. 795–802. [14] P.N. Sastry, T.R. Vijaya Lakshmi, R. Krishnan, N.V. Rao, Analysis of Telugu palm leaf characters using multi-level recognition approach, J. Appl. Eng. Sci. 10 (20) (2015) 9258–9264. [15] T.R.V. Lakshmi, P.N. Sastry, R. Krishnan, N.V.K. Rao, T.V. Rajinikanth, Analysis of Telugu palm leaf character recognition using 3D feature, in: International Conference on Computational Intelligence and Networks (CINE), Jan. 2015, pp. 36–41. [16] S. Bag, G. Harit, P. Bhowmick, Recognition of Bangla compound characters using structural decomposition, Pattern Recognit. 47 (3)) (March 2014) 1187–1201. [17] S.N.S. Rajasekaran, B.L. Deekshatulu, Recognition of printed Telugu characters, Comput. Graph. Image Process. 6 (4) (1977) 335–360. [18] H. Salimi, D. Giveki, Farsi/Arabic handwritten digit recognition based on ensemble of SVD classifiers and reliable multi-phase PSO combination rule, Int. J. Doc. Anal. Recognit. 16 (4) (2013) 371–386. [19] V.N. Murthy, A.G. Ramakrishnan, Choice of classifiers in hierarchical recognition of online handwritten Kannada and Tamil aksharas, J. Univers. Comput. Sci. 17 (1) (2011) 94–106. [20] V.M. Aradhya, G.H. Kumar, S. Noushath, Multilingual OCR system for South Indian scripts and English documents: an approach based on Fourier transform and principal component analysis, Eng. Appl. Artif. Intell. 21 (2008) 658–668. [21] V.M. Aradhya, G.H. Kumar, Principal component analysis and generalized regression neural networks for efficient character recognition, in: First International Conference on Emerging Trends in Engineering and Technology, 2008, pp. 1170–1174. [22] V.M. Aradhya, G.H. Kumar, S. Noushath, Robust unconstrained handwritten digit recognition using radon transform, in: International Conference on Signal Processing, Communications and Networking, 2007, ICSCN’07, IEEE, 2007, pp. 626–629. [23] M. Abaynarh, H. El Fadili, L. Zenkouar, Enhanced feature extraction of handwritten characters and recognition using artificial neural networks, J. Theor. Appl. Inf. Technol. 72 (3) (2015) 355–365. [24] P. Malik, R. Dixit, Handwritten character recognition using wavelet transform and Hopfield network, in: International Conference on Machine Intelligence Research and Advancement, 2014, pp. 125–129. [25] N. Das, K. Acharya, R. Sarkar, S. Basu, M. Kundu, M. Nasipuri, A benchmark image database of isolated Bangla handwritten compound characters, Int. J. Doc. Anal. Recognit. 17 (4) (2014) 413–431.