Offline Script Identification from multilingual Indic-script documents: A state-of-the-art

Abstract

Offline Script Identification (OSI) facilitates many important applications such as automatic archiving of multilingual documents, searching online/offline archives of document images and for the selection of script specific Optical Character Recognition (OCR) in a multilingual environment. In a multilingual country like India, a document containing text words in more than one language is a common scenario. A state-of-the-art survey about the techniques available in the area of OSI for Indic scripts would be of a great aid to the researchers. Hence, a sincere attempt is made in this article to discuss the advancements reported in the literature during the last few decades. Various feature extraction and classification techniques associated with the OSI of the Indic scripts are discussed in this survey. We hope that this survey will serve as a compendium not only for researchers in India, but also for policymakers and practitioners in India. It will also help to accomplish a target of bringing the researchers working on different Indic scripts together. Taking the recent developments in OSI of Indian regional scripts into consideration, this article will provide a better platform for future research activities.

Introduction

With the advancement in computer technology and the availability of low cost high capacity storage devices, storing of documents in electronic form has become a common practice. A document either in handwritten or printed form may contain writings in different scripts, graphics and images. For example, museum archives contain old fragile documents having scientific or historical or artistic value and written in different scripts with many graphic illustrations. OCR is a type of software designed to translate images of text into machine editable text. However, most OCR systems are script-specific in the sense that they can read characters written in a particular script only. Script is defined as the graphic form of the writing system. A script class refers to a particular style of writing and the set of characters used in it. Languages throughout this world are typeset in many different scripts. A script may be used by only one language or may be shared by many languages, sometimes with slight variations from one language to other. For example, Devnagari script is used for writing a number of Indic languages like Sanskrit, Hindi, Konkani, Marathi, Nepali, etc. It is perhaps impossible to design a single recognizer which can identify a large number of scripts/languages. So, for a multilingual text document, identification of different scripts and extraction of portions written in the same script is a pressing need so that script-specific OCR systems can be employed. However, manual identification of the mixed-script document may be too monotonous and imperceptible. So, automatic OSI techniques are necessary to identify the scripts in the given input document which can further be sent to their corresponding OCR engines. Fig. 1 shows some examples of mixed-script documents.

Script identification is a key step in document image analysis especially when the environment is multi-script and multilingual. It also serves as an essential precursor for recognizing the language in which a document is written. This is necessary for further processing of the document, such as searching, indexing or translation. For scripts used by only one language, script identification itself accomplishes language identification. For scripts shared by many languages, script recognition acts as the first level of classification followed by language identification within the script.

India is a highly multilingual country with 22 constitutionally recognized languages. Besides these, hundreds of other languages are used in India, each one with a number of dialects. The officially recognized languages are Hindi, Bengali, Punjabi, Marathi, Gujarati, Oriya, Sindhi, Assamese, Nepali, Urdu, Sanskrit, Tamil, Telugu, Kannada, Malayalam, Kashmiri, Manipuri, Konkani, Maithali, Santhali, Bodo, and Dogari. Hindi, written in Devnagari script, is India’s official language and has the most speakers, estimated to be more than 500 million. Indic scripts are a logical composition of individual script symbols and follow a common logical structure. This can be referred to as the “script composition grammar” which has no counterpart in any other set of scripts in the world. Indic scripts are written syllabically and are usually visually composed in three tiers where constituent symbols in each tier play specific roles in the interpretation of that syllable  [1].

Automatic script identification in a multilingual environment is a challenging research problem over the last two decades  [2]. Researchers have investigated OCR for a number of Indic scripts. However, most of these researches have been confined to the identification of isolated characters rather than the script. Unlike simple concurrence in Roman script, the Indic scripts are a composition of the constituent symbols in two dimensions. This implies that researchers first segment an Indic script word into its composite characters and then each composite character is decomposed into the constituent symbols or strokes that are finally recognized. Fig. 2 shows the general block diagram of the script identification system.

The process of identification of printed/handwritten script includes preprocessing and segmentation, feature extraction and recognition or classification processes. Preprocessing is necessary when the data or image is captured for further processing. Preprocessing is a method of enhancing the quality of an image for better understanding of the image. The choice of preprocessing method to be adopted on a document image depends on the type of application for which the image is used. The noise gets introduced in the document image during its acquisition and/or transmission over wired/wireless channels, as well as because of changing of some parameters of acquisition system in the OCR. Skew is an unavoidable distortion that is often introduced during scanning or copying of a document. There are many techniques that are generally available to accomplish preprocessing on images; however, several experiments on script identification suggest that preprocessing methods have got to be customized to suit the requirements of script identification. Preprocessing techniques generally include noise removal, scaling, binarization, skew and slant correction including the header line removal, etc. Binarization is the process of converting 256 levels of gray scale information into two levels (black and white) image information. To binarize a gray scale image, at first, threshold value(s) has to be determined. If a pixel value is less than the threshold value, then the pixel value of the corresponding output image is set as 1 (black) otherwise it is set as 0 (white). The morphological opening and closing operators  [3] are frequently used for removing noise from the document images but it also connects discontinuities that are caused during the thresholding stage. The opening and closing operators are defined as follows:

$$A\circ B=(A\ominus B)\oplus B$$$$\text{and}A\ast B=(A\oplus B)\ominus B$$

where, $\oplus$ and $\ominus$ are the morphological erosion and dilation operators respectively and $B$ is the related structure element.

Next, the text lines and then the words are segmented from the document images. For printed Indic scripts, it has been found that the horizontal and vertical projection profiles are useful for segmenting text lines and words respectively from document images. The horizontal projection profile method computes sum of all black pixels on every row and constructs the corresponding horizontal pixel density histogram. Based on the peak/valley points of the histogram, individual text lines are separated as shown in Fig. 3(a). When the density of data pixels in a row is zero, it denotes a boundary between two consecutive lines. After the text line segmentation, each text line is scanned vertically for word segmentation as shown in Fig. 3(b). For vertical projection profile, the number of data pixels corresponding to each column is calculated to construct a vertical pixel density histogram. So, by analyzing the vertical pixel density histogram of a text line, words are easily separated through the columns having data pixel density equal to zero.

Feature extraction is a vital part of any recognition system. The main intention behind feature extraction is to depict the pattern by means of a bare minimum number of attributes. One significant job in the design of a pattern recognition system is to develop an algorithm to extort characteristics or features of pattern from initial measurement. Feature selection is a process of minimizing the number of features and maximizing the discriminating property of the feature set. This process aims to identify an optimal subset of relevant features from a large number of features collected from the patterns in the dataset, such that the overall accuracy of classification is increased. Recognition or classification of a particular pattern depends on the selection of the features and classifiers which can classify or recognize a particular pattern belonging to a particular class. The classifier is sometimes called the ‘heart’ of the pattern recognition system. It takes a feature vector and assigns to it a decision where the decision is the label of pattern class labels which the pattern recognizer has decided.

From the literature survey, it can be seen that only a few researchers have taken the burden of writing a review article on the script identification techniques. Some of such reviews described in  [4], [5], [6], [7] are worth mentioning. U. Pal et al.  [4] presented a review of an OCR system for Indic language scripts. A survey of script identification techniques for multi-script document images is also described in  [5]. A report on the key technologies in multilingual OCR of handwritten Devnagari script can also be found in  [6]. In one of the earlier works  [7], a review of offline handwritten script identification techniques is also described.

All existing works on automatic script/language identification are broadly classified into either local approach or global approach  [7]. In local approach, the features are extracted from a list of connected components such as line, word and character, which are obtained only after segmenting the underlying document image. So, the success rate of classification depends on the effectiveness of the pre-processing steps. But, it is difficult to find a common segmentation method that best suits for all the script classes. Due to this limitation, local approaches cannot meet the criterion as a generalized scheme. In contrast, global approaches employ analysis of regions comprising of at least two text lines and hence fine segmentation of the underlying document into lines, words and characters, is not necessary. Consequently, the script classification task is simplified and performed faster with the global approach than the local approach. In this paper, we present a comprehensive survey of different script identification techniques developed mainly for identification of the major scripts of India.

The organization of the survey is as follows: Section  2 covers the properties and evolution of Indic scripts whereas Section  3 discusses the methods used for OSI of Indic scripts. Sections  4 Structure-based script identification, 5 Visual appearance-based script identification describes the different script identification techniques related to OSI. Section  6 presents a comparative analysis of some of the benchmark work for OSI of Indic scripts. Section  7 provides a snapshot of the benchmark databases available for researchers in this domain. Section  8 describes the scope of future work followed by some of the existing difficulties in this domain whereas Section  9 concludes the review.