atural language processing (NLP) is a field of computer science that deals with understanding and generation of natural languages. Natural language understanding enables computers to understand the natural language and extract meaning from it. Natural language generation involves both spoken and written information. Some applications of both text to speech and speech to text conversion are man-machine interfaces to computers, systems that read and understand printed and hand written text, speed understanding system, text analysis and understanding systems,computer aided instruction systems etc. Ediphones and Dictaphones are examples of speech to text conversion systems.
India is a multi lingual and multi script country with 22 scheduled languages. Every state has its lang-The alphabet of a language is divided into basic characters called vowels and consonants. Two or more basic characters are combined to form compound characters. We recall here the way we have learnt the alphabet and started writing the text in a language, in kindergarten schools, children are made to practice writing the characters and thereby memorize them. The character writing involves combining the primitives. Every character written from an alphabet follows a definite way and depends upon the type of the writer, whether left-hand-writer or right-handwriter. The character generation at once is different from how it is written. Each character in the alphabet has definite way of writing it and is combination of sub parts called primitives. This combination of primitives is a systematic approach in generating or building the characters. The character construction is basic to any medium of learning. While reading, we read the whole character and while writing or constructing a character, we write primitives in an order.
The automation of character construction requires the recognition of primitives from the database of primitives in a language. Many researchers have worked on character recognition, wherein the whole character is considered as one single unit. The focus of the present work is to recognize the images of primitives of Kannada language useful in the construction of characters using syntactic approach. The work is useful for novice learners, multimedia applications, translateration and translation etc. The automated script writing and learning by taking technological leverage is considered a new area of research.
In this paper, we have considered the different font types and font sizes of vowels and consonants characters supported by Kannada language software, namely, Nudi and Baraha. We have identified with the help of language experts the primitives of vowels and consonants and manually separated and their images are stored. These images of primitives are preprocessed through binarization, thinning and resizing. The simple zone based features are obtained for these primitives. Nearest neighbor classification is adopted with Euclidean distance measure for recognition of primitives. We have tested for all the combinations of printed primitives with different fonts" types and sizes.
The remaining part of the paper is organized into four sections. Section 2 deals with detailed survey on automatic primitive recognition. Section 3 deals with the proposed methodology, wherein different stages of the methodology are discussed. The experimental results and discussion are given in section 4. Conclusion and Future work are given in Section 5.
To know the state-of-the-art in automatic primitive recognition, we carried out the literature survey and following isthe gist of cited papers.
[ Leena R Ragha, et. al, 2010] have investigated the moments features on Kannada handwritten basic character set of 49 letters. Four directional images using Gabor wavelets from the dynamically preprocessed original images are found. Then moments features are extracted from them. The comparison of moments features of 4 directional images with original images when tested on Multi Layer Perceptron with Back Propagation Neural Network shows an average improvement of 13% from 72% to 85%. The mean performance of the system with these two features together obtained is 92%.
[ Karthik Sheshadri et.al, 2010] have proposed Kannada Character Recognition method based on kmeans clustering. A segmentation technique to decompose each character into components from three base classes is used to reduce the magnitude of the problem. The k-Means clustering technique provides a natural degree of font independence and this is used to reduce the size of the training data set to about one-tenth of those used in related works. Accuracy comparisons with related work, shows that the proposed method yields a better peak accuracy. The relative merits of probabilistic and geometric seeding in k-means are also discussed.
[ Leena R Ragha, et. al, 2011] have presented the use of moments features on Kannada Kagunita. Four directional images are found using Gabor wavelets from the dynamically preprocessed original image. The Kagunita set is analysed and the regions with vowel and consonant information are identified and cut from the preprocessed original image to form a set of cut images. Moments and statistical features are extracted from original images, directional images and cut images. These features are used for both vowel and consonant recognition using multi-layer perceptron with backpropagation learning. [Sangame S.K, et. al, 2011] have presented an unconstrained handwritten Kannada basic character recognition using invariant moments and chain code features. Invariant moments feature are extracted from zoned images and chain code. A Euclidean distance based K-NN classifier is used to classify the handwritten Kannada vowels. The method is invariant to size, slant, orientation and translation.
[B.V. Dhandra et. al, 2011] have proposed zone based features for recognition of the mixer of handwritten and printed Kannada digits. The kNN and SVM are used to classify the mixed handwritten and printed Kannadadigits. The reported recognition rates are 97.32% and 98.30% for mixed handwritten and printed Kannada digits using KNN and SVM classifiers respectively.
Volume XIV Issue IV Version I The result is computed using five-fold cross validation. The mean performance of the recognition reported for the two shape based features together is 98.45% and 93.92%, for numeral characters and vowels, respectively. The mean recognition rate of 95% is obtained for both vowels and characters taken together.
[K S Prasanna Kumar et. al, 2012] have presented an algorithm to optical character recognition (OCR) for Kannada numerals. The segmentation of a numeral into four equal parts and one of these parts i.e., left bottom segment, is used to extract recognition features. A conflict resolution algorithm is proposed to resolve the conflicting features. A minimum number of features are extracted so as to improve the response time.
[ Umapada Pal et. al, 2012] have given a stateof-the-art survey about the techniques available in the area of offline handwriting recognition (OHR) in Indian regional scripts. Various feature extraction and classification techniques associated with the offline handwriting recognition of the regional scripts are discussed in this survey. A separate section is dedicated to the observations made, future scope, and existing difficulties related to handwriting recognition in Indian regional scripts.
[Kauleshwar Prasad, et.al, 2013] [G. G. Rajput et.al, 2013] have proposed a zone based method for recognition of handwritten characters in Kannada language. The normalized character image is divided into 64 zones and each is of size 8x8 pixels. For each zone,from left to right and from top to bottom, the crack code, representing the line between the object pixel and the background (the crack), is generated by traversing it in anticlockwise direction. A feature vector of size 512 is obtained for each character. A multi-class SVM is used for the classification purpose. The data set has 24500 images with 500 samples of each character. Five-fold cross validation is used and yielded 87.24% recognition accuracy. [Swapnil A. Vaidya et. al, 2013] have given an overview of the ongoing research in OCR systems for Kannada scripts. They have provided a starting point for the researchers in the field of OCR. The state-of-the-art OCR techniques used in recognition of Kannada scripts, recognition accuracies and the resources available are discussedin fair detail.
[H. Imran Khan et. al, 2013] have proposed a chain code based feature extraction method for developing HCR system. A eight -neighborhood method is implemented, which allows generation of eight different codes for each character. These codes are used as features of the characters" images. These features are used for training and testing the k-Nearest Neighbor (KNN) classifier. [Nithya E. et.al, 2013] have proposed an OCR system for complex printed Kannada characters. The input to the system is a scanned image of a page of text containing complex Kannada characters and the output is in a machine editable form. The pre-processing step converts the input document into binary form. The lines from the document image are extracted and further segmented into the lines, characters and sub characters. The histogram and connected component methods are used for segmentation and correlation is used for recognition of characters.
[Mamatha.H.R. et, al, 2013] have attempted to measure the performance of the classifier by testing with twodifferent datasets of different sizes. A framework based on the combined concepts of decision fusion and feature fusion for the isolated handwritten Kannada numerals classification is proposed. The combined approach has increased the recognition accuracy by 13.95%. are segmented and the result is displayed along with a vocal output. [Sandhya.N et.al, 2014] have proposed a new classification method for Kannada characters, which is used as a preliminary step for recognition. An analysis of Kannada characters is carried out. The syntactic features are identified. At first, the basic features and their exact positions in the characters are identified and recorded. Further, by using a decision tree, the characters are classified. The experimental results show that the syntactic based method using basic features gives good contribution and reliability for Kannada character classification.
From the literature survey, it is observed that researchers have worked on Kannada character recognition. The feature extraction techniques such as template matching, Zernike moments, geometric moment invariants, directional, positional, Fourier transform, etc are used. The classification techniques such as, neural network, support vector machines, nearest neighbor, etc are used. No specific work is observed on recognition of Kannada language primitives in the light of character construction is cited in the literature [1][2][3][4][5][6][7][8][9][10][11][12][13][14][15][16][17][18]. This is the motivation for the present work on printed and handwritten Kannada language vowels and consonants primitives" recognition.
The proposed methodology consists of four major steps, namely, identifying the primitives in printed and handwritten Kannada vowels and consonants and obtaining their images, preprocessing of primitives" images, primitives" recognition and classification of primitives. The steps are shown in the Figure 1. This step consists of two tasks, namely, identifying primitives of printed as well as handwritten Kannada vowels and consonants and obtaining their images.
The Kannada language characters are classified into Swaras (vowels), Vyanjanas (consonants), Yogavaahakas (partly vowels and partly consonants), Kagunitha (combination of consonants and vowels) and Wothakshara (conjunct consonants) as given in Table 1. Kannada language script consists of more than 250 basic, modified and compound character shapes giving rise to 18511 distinct characters. We have used the word Kannada and Kannada language interchangeably in this paper. Kannada characters are curve shaped with some regions highly denser than others. Some shapes are wider and some are longer than others, as visible in Table1.
We have consulted the Kannada language experts and identified the primitives. These 38 primitives are categorized into basic primitives (BP) and character cum primitives (CcP) as shown in the Figure 2. A single primitive, which is also a complete vowel or complete consonant, is defined as Character cum Primitive. One or more basic primitives are joined at appropriate positions to form the given vowels and consonants. It is also observed that symmetry exists in most of the Kannada characters. Since there is no standard database available for Kannada primitives, we have created a database of primitives for printed and handwritten Kannada vowels and consonants in consultation with language experts. We have identified 38 primitives to construct all the Kannada vowels and consonants and are given in Table 2. For example, the basic primitives required for constructing the two vowels (pronounced as "e") and (pronounced as "aou") are given in Box2. The "+" symbol represents the connection of primitives at appropriate positions for constructing a character. A We have preprocessed the images of the primitives to make them suitable for feature extraction. Preprocessing involves binarization, noise reduction, size normalization and thinning. The binarization is categorized into two main classes, namely, global and local. We have adopted global approach for converting gray scale image to binary image. Image binarization is performed using Otsu's method. The salt and pepper noise present in the image is removed by applying median filter. The process of thinning involves reducing thickness of each line of the pattern to just a single pixel wide is carried out. Size normalization is required as the size of the primitives vary from one vowel to another. In order to bring uniformity among the images of primitives, each image is normalized to the size of 28*28 after finding the bounding box of each image without disturbing the aspect ratio using bilinear standard transformation. The images of primitive obtained after applying all the preprocessing steps to a given sample primitive is shown in Box 3. A total of 49 features values are extracted from each primitive and this will serve as the feature vector. The feature vector for image (i ) denoted by Fi = {z1, z2, z3 ? z49}, where zi denotes ith zone value. There could be some zones, which do not contain any part of the primitive at all; therefore the corresponding zone value in the feature vector is zero. The set of feature vectors obtained from the training samples is used as the Knowledge Base (KB). This knowledge base is used to recognize the test samples. We have used nearest neighbor classifier for recognition. Step 1. Accept and preprocess the input image to eliminate the noise using median filter and perform thinning.
Step 2. Fit the input image in a bounding box and crop the image to resize to 28*28 pixels.
Step 3. Extract 49 zone values, define feature vector and store.
Step 4 . Repeat steps 1 to 3 until the training images are exhausted.
Step 5. Accept the test sample.
Step 6. Compute Euclidean distance of the test sample with all the trained images.
Step 7. Declare the class of the primitive as the class with minimum distance using Nearest neighbor classification.
Step 8. Repeat steps 5 and 7 until test images are exhausted.
Step 9. Obtain the accuracy of classification.
The accuracy of the classifier is evaluated through k-fold cross-validation method. In this method, each time one of the k subsets of images is used as the test set of images and the other (k-1) subsets are put IV.
In experimentation of the methodology, we have considered 60 font styles and 100 font sizes. The different combinations are tried and are as given in Table 3. We have totally five combinations of font styles and font sizes. The bit 0 indicates varying and 1 indicates constant. The experimentation is done on 9120 (38*240) images of printed Kannada vowel primitives. We have considered 240 images with varying font size and font styles for each primitive. The font size and font styles used are given in Table 4. The entire image set is partitioned into training set and test set and classified using K-fold cross validation method. An experiment is carried out for a total of 39 primitives out of which 14 are character cum primitives and 24 are basic primitives. The image data set has 1520 (38*40) images. For example, we have considered 40 images of font size 60 and font style -Nudi 0.1. The zones based features and nearest neighbor classifier have given 100% recognition accuracy for both the types of primitives. Table 5 gives the results obtained for Character cum Primitives and Table 6 gives the results obtained for Basic primitives. 4, are considered. The classifier is subjected to k-fold crossvalidation. We have considered 1520 primitive images in each validation step for training and 760 primitive images for testing. Table 7 and Table 8 give recognition accuracy for this combination of CcP and BP primitives using 3-fold cross validation method. The range of recognition accuracies obtained for both, character cum primitives and basic primitives are given in Table 15. 12 show the results for both varying font sizes and font styles using 5-fold cross validation using Euclidean distance. The recognition accuracy obtained for basic primitives and for character cum primitives is given in Table 15. An experiment is carried out on 3900 (39*100) images. We have considered100 images of 10 font styles and 10 varying sizes, for each font style, as given in Table 4, for each primitive. Table 13 and Table 14 show the results for both varying font sizes and font styles using 5-fold cross validation using Euclidean distance. The recognition accuracy obtained for basic primitives and for character cum primitives is given in Table 15. V.
We have proposed a zone features based methodology for recognition of both printed and handwritten 38 primitives of 49 Kannada vowels and consonants together. Zones of 4*4 and images of 28*28 are used. The nearest neighbor classifier is used with Euclidean distance measure. The accuracy of the classifier is verified with k-fold cross validation, for k = 2,3 and 5. We have experimented with four combinations of font sizes and font styles for printed primitives and obtained average recognition accuracy in the range [89%, 94%]. Further, we have obtained accuracy in the range [90%, 94% ]for handwritten primitives. This work is basic to automation of writing of Kannada kagunitha and wothakshara"s.
Combinations Font Size(S) Font Styles(F) | Remarks | ||
1. | 0 | 0 | Both Font Size and Font Style are Fixed |
2. | 0 | 1 | Font Size is Fixed and Font Style is Varied |
3. | 1 | 0 | Font Size is Varied and Font Style is Fixed |
4. | 1 | 1 | Both Font Size and Font Style are Varied |
(Nonuniform mix) | |||
5. | 1 | 1 | Both Font Size and Font Style are Varied |
(Uniform mix) |
Sl No | Font Size | Font Style |
1. | 12,14,16,??,110 (100 sizes) | Baraha 01 |
2. | 12,14,16,??,110 (100 sizes) | Baraha 02 |
..... | 12,14,16,??,110 (100 sizes) | ?.. |
..... | 12,14,16,??,110 (100 sizes) | Baraha 30 |
..... | 12,14,16,??,110 (100 sizes) | Nudi 01 |
..... | 12,14,16,??,110 (100 sizes) | ?.. |
60. | 12,14,16,??,110 (100 sizes) | Nudi 30 |
a) Both Font Size and Font Style Fixed |
Year 2014 | |
48 | |
Volume XIV Issue IV Version I | |
D D D D ) F | Same Size and Same Font |
Global Journal of Computer Science and Technology ( | Primitive P1 P2 P3 P4 P5 P6 P7 P8 P9 P10 P11 P12 P13 P14 Average Image size=28*28 Zone=4*4 No of Samples=38*40=1520 Classification Accuracy (%) 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 |
Same Size and Same Font | |||
Image size=28*28 Zone=4*4 No of Samples=38*40=1520 | |||
Primitive | Classification | Primitive | Classification |
Accuracy(%) | Accuracy(%) | ||
P15 | 100 | P27 | 100 |
P16 | 100 | P28 | 100 |
P17 | 100 | P29 | 100 |
P18 | 100 | P30 | 100 |
P19 | 100 | P31 | 100 |
P20 | 100 | P32 | 100 |
P21 | 100 | P33 | 100 |
P22 | 100 | P34 | 100 |
P23 | 100 | P35 | 100 |
P24 | 100 | P36 | 100 |
P25 | 100 | P37 | 100 |
P26 | 100 | P38 | 100 |
Average | 100 | ||
b) Font Size Fixed and Font Style Varied | |||
An experiment is carried out for 2280 (38*60) | |||
images. |
Year 2014 | |||||
49 | |||||
3-Fold cross validation of CcP's 3_Fold Cross Validation | Volume XIV Issue IV Version I | ||||
Same Size and Different Font using Euclidean Distance | D D D D D D D D ) | ||||
Primitive P1 P2 P3 P4 P5 P6 P7 P8 P9 P10 P11 P12 P13 P14 Average Image size=28*28 Zone size=4*4 No of Samples=38*60=2280 1_Fold 2_Fold 3_Fold Average 70 80 80 76.666667 65 75 50 63.333333 70 100 90 86.666667 100 95 100 98.333333 95 100 100 98.333333 75 85 90 83.333333 60 85 100 81.666667 70 70 85 75 70 75 85 76.666667 75 75 95 81.666667 95 85 65 81.666667 95 85 90 90 70 75 60 68.333333 80 85 85 83.333333 77.857143 83.571429 83.928571 81.785714 Table 8 : 3-Fold cross validation of BP's 3_Fold Cross Validation | Global Journal of Computer Science and Technology ( | ||||
Same Size and Different Font using Euclidean Distance | |||||
Image size=28*28 Zone size=4*4 No of Samples=38*60=2280 | |||||
Primitive | 1_Fold | 2_Fold | 3_Fold | Average | |
P15 | 75 | 80 | 85 | 80 | |
P16 | 85 | 95 | 80 | 86.666667 | |
P17 | 100 | 100 | 100 | 100 | |
P18 | 85 | 80 | 80 | 81.666667 | |
P19 | 80 | 60 | 60 | 66.666667 |
Year 2014 | ||||
50 | ||||
Volume XIV Issue IV Version I D D D D ) F | 4-Fold cross validation of CcP's 2_Fold Cross Validation Different Size and Same Font using Euclidean Distance Image size=28*28 Zone size=4*4 No of | |||
Global Journal of Computer Science and Technology ( | Primitive P1 P2 P3 P4 P5 P6 P7 P8 P9 P10 P11 P12 P13 P14 Average | 1_Fold 100 95 100 100 100 100 100 100 100 100 100 100 100 100 99.64286 Table 10 : 4-Fold cross validation of BP' 2_Fold Average 100 100 100 97.5 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 99.82143 2_Fold Cross Validation Samples=38*40=1520 | ||
Different Size and Same Font using Euclidean Distance | ||||
Image size=28*28 Zone size=4*4 No of | ||||
Samples=38*40=1520 | ||||
Primitive | 1_Fold | 2_Fold | Average | |
P15 | 100 | 100 | 100 | |
P16 | 100 | 100 | 100 | |
P17 | 100 | 100 | 100 | |
P18 | 100 | 100 | 100 | |
P19 | 100 | 100 | 100 |
Year 2014 | ||||||
51 | ||||||
Volume XIV Issue IV Version I | ||||||
D D D D D D D D ) | ||||||
( | ||||||
5-Fold cross validation of BP's (Non -uniform mix) 5_Fold Cross Validation Different Size and Different Font using Euclidean Distance | Global Journal of Computer Science and Technology | |||||
Image size=28*28 Zone size=4*4 No of Samples=38*100=3800 | ||||||
Primitive | 1_Fold | 2_Fold | 3_Fold | 4_Fold | 5_Fold | Average |
P15 | 100 | 100 | 100 | 100 | 90 | 98 |
P16 | 90 | 95 | 100 | 100 | 100 | 97 |
P17 | 100 | 100 | 100 | 100 | 100 | 100 |
P18 | 95 | 90 | 100 | 100 | 100 | 97 |
P19 | 90 | 90 | 95 | 100 | 80 | 91 |
P20 | 90 | 90 | 100 | 100 | 90 | 94 |
5_Fold Cross Validation | ||||||
Different Size and Different Font using Euclidean Distance | ||||||
Image size=28*28 Zone size=4*4 No of Samples=38*100=3800 | ||||||
Primitive | 1_Fold | 2_Fold | 3_Fold | 4_Fold | 5_Fold | Average |
1 | 100 | 100 | 100 | 90 | 100 | 98 |
2 | 95 | 100 | 100 | 90 | 100 | 97 |
3 | 100 | 95 | 100 | 90 | 90 | 95 |
4 | 100 | 95 | 100 | 100 | 100 | 99 |
5 | 100 | 100 | 100 | 100 | 100 | 100 |
6 | 100 | 100 | 95 | 100 | 90 | 97 |
7 | 100 | 100 | 100 | 100 | 95 | 99 |
8 | 100 | 90 | 100 | 90 | 85 | 93 |
9 | 90 | 100 | 100 | 90 | 100 | 96 |
10 | 100 | 90 | 95 | 100 | 95 | 96 |
11 | 100 | 100 | 90 | 100 | 85 | 95 |
12 | 100 | 100 | 90 | 85 | 90 | 93 |
13 | 100 | 100 | 90 | 95 | 95 | 96 |
14 | 95 | 100 | 100 | 95 | 95 | 97 |
Average | 98.571429 | 97.857143 | 97.142857 | 94.642857 | 94.285714 | 96.5 |
Year 2014 | ||||||
52 | ||||||
Volume XIV Issue IV Version I | ||||||
D D D D ) F | ||||||
( | ||||||
Global Journal of Computer Science and Technology | 5-Fold cross validation of BP's(Uniform mix) 5_Fold Cross Validation Different Size and Different Font using Euclidean Distance | |||||
Image size=28*28 Zone size=4*4 No of Samples=38*100=3800 | ||||||
Primitive | 1_Fold | 2_Fold | 3_Fold | 4_Fold | 5_Fold | Average |
P15 | 95 | 85 | 70 | 75 | 60 | 77 |
P16 | 90 | 85 | 95 | 100 | 90 | 92 |
P17 | 100 | 90 | 95 | 100 | 95 | 96 |
P18 | 85 | 95 | 95 | 100 | 95 | 94 |
P19 | 80 | 95 | 100 | 100 | 100 | 95 |
P20 | 100 | 90 | 90 | 100 | 85 | 93 |
Printed Basic | Printed Character cum | |||
Primitives | Primitives | |||
Combinations | Euclidean Distance | Euclidean Distance | ||
Range | Avg | Range | Avg | |
Both Font Size and Font Style are Fixed | 100% | 100% | 100% | 100% |
Font Size is Fixed and Font Style is Varied | 25%--93% | 76% | 63%--98% | 82% |
Font Size is Varied and Font Style is Fixed | 68%-100% | 97% | 98%-100% | 99% |
Both Font Size and Font Style are Varied(Non-uniform | 38%-100% | 90% | 93%-100% | 97% |
Both Font Size and Font Style are Varied(Uniform mix) | 44%-100% | 88% | 93%-100% | 87% |
Average | 90% | 93% | ||
Handwritten Basic | Handwritten Character cum | |||
Primitives | Primitives | |||
Euclidean Distance | Euclidean Distance | |||
Average | 93% | 94% |
Data fusion based framework for the recognition of Isolated Handwritten Kannada Numerals. International Journal of Advanced Computer Science and Applications 2013. 2013. 4 (6) .
Implementing Kannada Optical Character Recognition on the. Android Computer and Communication Engineering 2013. 2 (1) .
A Novel Approach of Handwritten Character Recognition using Positional Feature Extraction. International Journal of Computer Science and Mobile Computing 2013. June-2013. 2 p. .
Zone Based Features for Handwritten and Printed Mixed Kannada Digits Recognition. International Conference on VLSI, Communication & Instrumentation, 2011. p. 2011. (ICVCI)
OCR System for Complex Printed Kannada Characters. International Journal of Advanced Research in Computer Science and Software Engineering 2013. June -2013. 3 (6) p. .
Shape Descriptors based Handwritten Character Recognition Engine with Application to Kannada Characters. International Conference on Computer & Communication Technology (ICCCT)-2011, 2011.
Zone based Handwritten Kannada Character Recognition Using Crack code and SVM. 2013 International Conference on Advances in Computing, 2013. (ICACCl)-2013)
Isolated Kannada Character Recognition using Chain Code Features. International Journal of Science and Research 2319-7064. 2013. August 2013. IJSR. (8) .
Optical Character Recognition (OCR) for Kannada numerals using Left Bottom 1/4th segment minimum features extraction. International Journal of Computer Techology & Applications :2229-6093. 2012. 3 (1) p. .
Character Recognition Using Matlab"s Neural Network Toolbox. International Journal of u-and e-Service 2013. 6 (1) . (Science and Technology)
A Comparative Study of Different Feature Extraction and Classification Methods for Recognition of Handwritten Kannada Numerals. International Journal of Database Theory and Application 2013. 6 (4) .
Recent Trends and Tools for Feature Extraction in OCR Technology. International Journal of Soft Computing and Engineering (IJSCE) 2231- 2307. 2013. January 2013. (2) .
Feature Based Kannada Character Classification Method of Kannada Character Recognition. International Journal of Scientific & Engineering Research 2229-5518. 2014. February-2014 729. 5 (2) .
Operating System for Kannada Sign Boards. International Journal of Advanced Research
Using Moments Features from Gabor Directional Images for Kannada Handwriting Character Recognition. International Conference and Workshop on Emerging Trends in Technology (ICWET 2010) -TCET, (Mumbai, India
Feature Analysis for Handwritten Kannada Kagunita Recognition. International Journal of Computer Theory and Engineering 2011. 3 p. .
Recognition of isolated handwritten Kannada characters using invariant moments and chain code. world journal of science and technology 2231 -2587. 2011. 1 (8) p. .
Handwriting Recognition in Indian Regional Scripts: A Survey of Offline Techniques. ACM Transactions on Asian Language Information Processing 2012. March 2012. 11 (1) . (Publication date)