Performance Evaluation of Feed-Forward Neural Network Models for Handwritten Hindi Characters with Different Feature Extraction Methods

Performance Evaluation of Feed-Forward Neural Network Models for Handwritten Hindi Characters with Different Feature Extraction Methods

Gunjan Singh (Faculty of Computer Application, Raja Balwant Singh Management Technical Campus, Agra, India), Sandeep Kumar (Department of Computer Science, Faculty of Engineering and Technology Agra College, Agra, India) and Manu Pratap Singh (Department of Computer Science, Institute of Engineering & Technology, Dr. B. R. Ambedkar University, Agra, India)
Copyright: © 2017 |Pages: 20
DOI: 10.4018/IJALR.2017070103


Automatic handwritten character recognition is one of the most critical and interesting research areas in domain of pattern recognition. The problem becomes more challenging if domain is handwritten Hindi character as Hindi characters are cursive in nature and demonstrate a lot of similar features. A number of feature extraction, classification and recognition techniques have been devised and being used in this area; still the efficiency and accuracy is awaited. In this article, performance of various feed-forward neural networks is evaluated for the generalized classification of handwritten Hindi characters using various feature extraction methods. To study and analyze the performance of the selected neural networks, training and test character patterns are presented to each model and their recognition accuracy is measured. It has been analyzed that the Radial basis function network and Exact Radial basis network give highest recognition accuracy while Elman backpropagation neural network gives lowest recognition rate for most of the selected feature extraction methods.
Article Preview

1. Introduction

Handwritten character recognition problem is one of the most explored and complex problem domains in machine learning field (Govindan & Shivprasad, 1990). Based on the type of handwritten data, recognition systems can be categorized as online or offline (Plamondon & Srihari, 2000). Due to the increasing popularity and demand in various application domains, offline handwritten character recognition is gaining a lot of attention in last few decades, but it suffers from generalization. Due to varying writing styles and splines used in handwritten Hindi curve scripts, characters may not have smooth curves or perfectly straight lines all the time. Thus, an offline handwritten character recognition system should be able to efficiently recognize characters with varying size, width, orientation and thickness. It should also be able to identify and recognize the missing data. It has been observed that extraction of significant and meaningful features set direction for proper and accurate classification (Shih, 2010). These features should be independent of the size, orientation and location of the pattern for the smooth functioning of the recognition system. Neural networks are widely used in the field of handwritten character recognition due to their learning, storing and recalling capabilities. A lot of sincere efforts have been done in the field of handwritten character recognition. A few of them are reviewed here.

In (Srivastava & Singh, 2010), an analysis of performance evaluation is presented for the feed-forward neural network trained with three learning algorithms- backpropagation algorithm, Genetic algorithm and Hybrid Genetic algorithm for handwritten English alphabets. The performance of feed-forward neural networks trained with hybrid evolutionary algorithm is analyzed for the recognition of scaled and rotated hand written English alphabets (Pandey, Srivastava, Lavania, & Singh, 2011). An improved backpropagation algorithm is proposed to improve the training efficiency of the backpropagation algorithm for offline handwritten characters. The algorithm works by modifying the gain value, momentum coefficient and learning rate (Kumar & Bhatia, 2013). A performance analysis of multilayer feed forward neural network is presented to perform the classification for the handwritten three letter English words. In this method, the network was trained with modified conjugate descent learning method to minimize the instantaneous local square error with respect to the current weight vector in the weight space in order to improve the performance of feed forward neural network for generalized pattern classification (Parappa & Singh, 2015).

Research work done for handwritten Hindi characters is reviewed next. A feed-forward neural network system with 92% recognition accuracy has been presented in which the preprocessed input character is converted into a Boolean matrix of size 5 X 7 and assigned the matrix to the neural network for classification (Khanale & Chitnis, 2011). A two-layer feed-forward network trained with Levenberg-Marquardt backpropagation algorithm is implemented for Hindi character recognition (Saxena & Gupta, 2012). In this method, the character image is segmented horizontally & vertically into four sub parts. Then pixel value intensities are calculated for each of the segmented part and for full image view. The 82.89% recognition accuracy is obtained with Generalized regression network (Vaidya & Bombade, 2013). A multi-layer neural network with 75% - 80% accuracy for Hindi characters and numerals was reported in (Jaiswal, 2014). A backpropagation neural network with 95% recognition accuracy is presented in (Tanuja, Kumari, & T.M.S., 2015). In this work, canny edge detection method is used to perform character segmentation and distance transformation methods is used to convert the character image into feature & non-feature elements. A study of moment based features is performed on handwritten digit recognition in five Indian scripts: Indo-Arabic, Bangla, Devnagari, Roman and Telugu (Singh, Sarkar, & Nasipuri., 2016). In (Rani, Kakkar, & Rani, 2016), a neural network with 82.5%, 80% and 70% recognition accuracy for Gurumukhi characters, Hindi vowels and English alphabets respectively is presented. In (Dogra & Sehgal, 2017), neural network based approach is followed for the segmentation and recognition of Hindi letters with 100% accuracy for word segmentation, 84.72% accuracy for character segmentation and 89.58% recognition accuracy.

Complete Article List

Search this Journal:
Open Access Articles: Forthcoming
Volume 9: 2 Issues (2019): Forthcoming, Available for Pre-Order
Volume 8: 2 Issues (2018): 1 Released, 1 Forthcoming
Volume 7: 2 Issues (2017)
Volume 6: 2 Issues (2016)
Volume 5: 1 Issue (2015)
Volume 4: 1 Issue (2014)
Volume 3: 4 Issues (2012)
Volume 2: 4 Issues (2011)
Volume 1: 4 Issues (2010)
View Complete Journal Contents Listing