Elsevier

Expert Systems with Applications

Volume 134, 15 November 2019, Pages 120-137
Expert Systems with Applications

Recognition of similar characters using gradient features of discriminative regions

https://doi.org/10.1016/j.eswa.2019.05.050Get rights and content

Highlights

  • A feature extraction method for differentiating visually similar characters sets.

  • A class hierarchy describing similar characters of Lanna Dhamma and Thai alphabets.

  • Evaluation and analysis of the performance of the proposed method.

Abstract

One important and challenging issue in handwritten character recognition is the discrimination of visually similar characters. In this paper, we propose a character recognition method for distinguishing similar characters by augmenting commonly used image feature with gradient features from potentially discriminative image regions. The discriminative regions of similar characters sets are automatically detected by analysing the weight vectors of the sparsity promoting logistic fused Lasso method. The histogram of oriented gradients is adopted to compactly represent the gradient features. Additionally, the locality preserving projection method is employed to alleviate the high dimensional nature of the resulting feature vectors. Experimental results on handwritten Lanna Dhamma and Thai characters datasets demonstrate the capability of the proposed method in discriminating visually similar characters. The method also outperforms existing character recognition methods by considerable margins. It has a great potential for character recognition of other alphabets.

Introduction

Optical Character Recognition (OCR) is a well-known automatic method for transforming digital document images to machine-encoded texts. Several techniques of OCR have been developed during the past decade (Cilia, Stefano, Fontanella, di Freca, 2019, Kamble, Hegadi, 2015, Olszewska, 2015, Pramanik, Bag, 2018, Sarkhel, Das, Das, Kundu, Nasipuri, 2017, Zhou, Zhang, Yin, Liu, 2016). The common module of the majority of OCR pipelines is the character recognition phase where image actually get converted into text and it is the phase where existing methods were unable to provide satisfactory result. Of particular interest is the recognition of visually similar characters found in several Asian alphabets. The confusion arising during the recognition of similar characters is a challenging issue in OCR. Several literature reported that the issue exists in OCR of Chinese (Jin, Gao, Liu, Li, Ding, 2011, Wang, Lu, 2017), Hangul (Ryu & Kim, 2014), Japanese (Suzuki, Kato, Nemoto, & Ichimura, 2002), Bangla (Surinta, Karaaba, Schomaker, & Wiering, 2015), Devanagari (Jangid & Srivastava, 2016), Thai (Surinta et al., 2015) and Lanna Dhamma (Inkeaw, Charoenkwan, Huang, Marukatat, Ho, Chaijaruwanich, 2017, Inkeaw, Chueaphun, Chaijaruwanich, Klomsae, Marukatat, 2015), for example.

In recent years, techniques for improving the recognition rate of similar characters have been investigated in many studies. We can categorize existing methods into two streams: classifier-centric and feature-centric. The classifier-centric approach focuses on developing a complex and highly discriminative classifier for better discriminating similar characters. Recently, Wang and Lu used Convolutional Neural Networks (CNN) organized in a hierarchical structure to deal with groups of similar characters (Wang & Lu, 2017). Shao et al. presented a special Artificial Neural Network (ANN) named Connection Reduced Network (CRN) for similar characters discrimination (Shao, Gao, & Wang, 2016). Although discriminative classifiers proposed in these studies performed well on similar character recognition, they usually require large training sample due to the classifier’s high complexity.

On the other hand, feature-centric approach aims at engineering high quality features to be learned by generic classifiers such as Modified Quadratic Discriminant Function (MQDF) (Gao, Liu, 2008, Ryu, Kim, 2014, Shao, Wang, Xiao, Zhang, Zhang, 2011), ANN (Chaudhary, Shikkenawis, Mitra, & Goswami, 2012), Minimum Euclidean Distance (MED) (Tao, Liang, Jin, & Gao, 2014), Bayes classifier (Leung & Leung, 2010) and Support Vector Machine (SVM) (Jangid & Srivastava, 2016). Some studies adopted dimensionality reduction techniques such as Linear Discriminant Analysis (LDA) (Gao & Liu, 2008), Local Preserving Projection (LPP) (Chaudhary et al., 2012) and Kernel Discriminative Locality Alignment (KDLA) (Tao et al., 2014) to obtain low dimensional representation of characters. Although, low dimensional features are often easier to work with, one must be aware that some subtle information can be lost during the dimensionality reduction process. As a result, the approach is generally less appealing.

Some studies determine discriminative regions before feature extraction. Features from the discriminative regions are additionally used to improve the recognition performance of similar characters. Fisher’s Linear Discriminant (FLD) (Jangid, Srivastava, 2016, Leung, Leung, 2010), Average Symmetric Uncertainty (ASU) (Xu, Huang, & Liu, 2010), AdaBoost (Shao et al., 2011) and regional importance measure (Ryu & Kim, 2014) are the main methods used to analyze the discriminative regions. The characteristics of existing similar character recognition methods were summarized and are presented in Table 1. As one can see from the table, the existing discriminative region based methods (Jangid, Srivastava, 2016, Leung, Leung, 2010, Ryu, Kim, 2014, Shao, Wang, Xiao, Zhang, Zhang, 2011) can deal only with similar character pairs. In practice, there can be more than two characters that have very similar shapes: for example, similar characters sets in Lanna Dhamma and Thai alphabets are illustrated in Fig. 1. In addition, the methods in Leung and Leung (2010), Ryu and Kim (2014) and Jangid and Srivastava (2016) divide an image into grid with predefined cell size, and the discriminative regions are determined upon the divided cells. Defining an optimal cell size depends on characteristics of considering character images, which is not always straightforward.

In this work, we follow feature-centric approach and propose a character recognition technique which works well not only on easily distinguishable characters but also on visually similar characters. This is accomplished by augmenting gradient features extracted from the whole image with gradient features from “discriminative regions”. The proposed method consists of three phases: (1) determination of discriminative regions, (2) discriminative feature extraction and (3) character recognition. In the first phase, determining discriminative regions of similar characters sets is formulated as a feature selection problem upon pixel information. We incorporate the fused Least Absolute Shrinkage and Selection (fused Lasso) (Tibshirani, Saunders, Rosset, Zhu, & Knight, 2005) into a logistic regression classifier and consider the learnt weight vector as a discriminative map of similar character pairs. Since there are several pairs of similar characters, we compute discriminative maps of all pairs and merge them into one map. The local maximum regions on the merged map are taken as the discriminative regions. We postulate that the information from discriminative regions will be useful for discriminating similar characters. All similar characters sets are individually analyzed to obtain their own discriminative regions maps. In the second phase, we extract Histogram of Oriented Gradients (HOG) (Dalal & Triggs, 2005) from the discriminative regions as well as from the whole image and form a feature vector of the image. Additionally, the supervised LPP (He & Niyogi, 2003) is used to reduce the dimensionality of the feature vector. In the third phase, standard classifiers can be used to classify character images using the resulting feature vectors.

We evaluated the proposed method using two handwritten Lanna Dhamma and Thai datasets. The similar characters sets of each alphabet are described by a hierarchy designed on character shape. We perform experiments using both cross-validation and writer-independent tests. We consider the following points to be our key contributions:

  • A new feature extraction method aiming at differentiating visually similar characters sets. The method combines three well-established techniques. The fused Lasso is extended with logistic regression in order to determine discriminative regions. The HOG is adopted as feature description, and the LPP is applied to reduce the dimensionality of the final feature vector.

  • A characters hierarchy to describe similar characters sets of Lanna Dhamma and Thai alphabets.

  • Evaluation and analysis of the performance of the proposed method through experiments on datasets of handwritten Lanna Dhamma and Thai characters.

We believe that the proposed method has a great potential for character recognition of other alphabets, especially in feature engineering approach. In addition, it can be also applied to other image recognition problems with appropriate modifications.

The rest of the paper is organized as follows. Section 2.1 gives brief descriptions of fused Lasso and LPP which are utilized in this work; Section 3 describes the feature extraction method we propose to address the problem of similar characters, Section 4 describes the datasets used in this study, establishes similar character sets and provides the discussions of the empirical results; finally, Section 5 concludes the study.

Section snippets

Backgrounds

In this section, we briefly describe the techniques utilized in this work namely fused Lasso and locality preserving projections.

Proposed method

The diagram of the proposed character recognition technique is shown in Fig. 2. It consists of three phases: (1) determination of discriminative regions, (2) discriminative feature extraction and (3) character recognition. We firstly determine discriminative regions of similar characters sets. All similar characters sets are individually analyzed to obtain their own discriminative regions. In the second phase, we extract HOG features from the discriminative regions as well as from the whole

Experiment and results

In this section, we demonstrate the performance of the proposed method for handwritten Lanna Dhamma and Thai characters. We organize the experiment into five parts. The first part presents the datasets used in this section. In the second part, we establish character class hierarchies for the two alphabets. The third part describes the experimental setting of baseline classifier. In the fourth part, we describe evaluation methodology and evaluation measurements. The last part shows the

Conclusion and future work

We have proposed a character recognition method to distinguish similar characters by gradient feature from the discriminative regions. The proposed method consists of three phases: (1) determination of discriminative regions, (2) discriminative feature extraction and (3) character recognition.

We evaluated the proposed method using datasets of handwritten Lanna Dhamma and Thai characters which were divided into training and test datasets. The training samples were collected from writers distinct

CRediT authorship contribution statement

Papangkorn Inkeaw: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Data curation, Writing - original draft, Visualization. Jakramate Bootkrajang: Formal analysis, Writing - review & editing. Sanparith Marukatat: Writing - review & editing. Teresa Gonçalves: Resources, Writing - review & editing. Jeerayut Chaijaruwanich: Writing - review & editing, Supervision, Project administration, Funding acquisition.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgements

This study was funded under the Royal Golden Jubilee Ph.D. Program by the Thailand Research Fund (Grant No. PHD/0185/2556). We would like to thank Chiang Mai University, Thailand, for financial support and collection of digital Lanna archives. We also thank Department of Informatics, University of Évora, Portugal, for supporting this work in a laboratory.

References (45)

  • S. Barua et al.

    Bangla handwritten city name recognition using gradient-based feature

  • S. Boughorbel et al.

    Generalized histogram intersection kernel for image recognition

    IEEE international conference on image processing 2005

    (2005)
  • M. Chaudhary et al.

    Similar looking gujarati printed character recognition using locality preserving projection and artificial neural networks

    2012 third international conference on emerging applications of information technology

    (2012)
  • N. Dalal et al.

    Histograms of oriented gradients for human detection

    2005 IEEE computer society conference on computer vision and pattern recognition (cvpr’05)

    (2005)
  • T.-F. Gao et al.

    High accuracy handwritten chinese character recognition using lda-based compound distances

    Pattern Recognition

    (2008)
  • X. He et al.

    Locality preserving projections

    Proceedings of the 16th international conference on neural information processing systems

    (2003)
  • J. Heaton

    Introduction to neural networks for Java

    (2008)
  • M.R. Hestenes

    Multiplier and gradient methods

    Journal of Optimization Theory and Applications

    (1969)
  • P. Inkeaw et al.

    Recognition-based character segmentation for multi-level writing style

    International Journal on Document Analysis and Recognition (IJDAR)

    (2018)
  • P. Inkeaw et al.

    Recognition of handwritten lanna dhamma characters using a set of optimally designed moment features

    International Journal on Document Analysis and Recognition (IJDAR)

    (2017)
  • P. Inkeaw et al.

    Lanna dharma handwritten character recognition on palm leaves manuscript based on wavelet transform

    2015 IEEE international conference on signal and image processing applications (ICSIPA)

    (2015)
  • M. Jangid et al.

    Similar handwritten devanagari character recognition by critical region estimation

    2016 international conference on advances in computing, communications and informatics (ICACCI)

    (2016)
  • Cited by (11)

    • Density based semi-automatic labeling on multi-feature representations for ground truth generation: Application to handwritten character recognition

      2021, Knowledge-Based Systems
      Citation Excerpt :

      Each character sample is a 54 × 54 pixel square. Confusion arising during the recognition of similar characters is a notably challenging issue in OCR of Devnagari [39], Thai [36] and Lanna Dhamma alphabets [28,40]. The properties of the datasets used in our experiments are summarized in Table 2.

    • Offline recognition of handwritten Indic scripts: A state-of-the-art survey and future perspectives

      2020, Computer Science Review
      Citation Excerpt :

      The feature extraction technique transforms large and redundant input into a reduced collection of parameters called a feature vector. Feature vectors hold great significance as they efficiently and unambiguously recognize the patterns and thereby improve the recognition rate of OCR [44,45]. Features are broadly classified into two categories, namely statistical and structural features.

    • Burnt-in Text Recognition from Medical Imaging Modalities: Existing Machine Learning Practices

      2024, Journal of Advanced Computational Intelligence and Intelligent Informatics
    • Thai Handwritten Character Recognition Using Deep Convolutional Neural Network

      2023, 2023 8th International Conference on Computer and Communication Systems, ICCCS 2023
    View all citing articles on Scopus
    View full text