Mujtaba Husnain

IET Image Processing Review Article Urdu handwritten text recognition: a survey ISSN- Received on 8th April 2019 Revised 26th March 2020 Accepted on 14th April 2020 doi: 10.1049/iet-ipr- www.ietdl.org Mujtaba Husnain1 , Malik Muhammad Saad Missen1, Shahzad Mumtaz1, Mickaël Coustaty2, Muzzamil Luqman2, Jean-Marc Ogier2 1Department of Computer Science & IT, The Islamia University of Bahawalpur 63100, Pakistan Lab, Université of La Rochelle Av. Michel Cŕepeau, 17000 La Rochelle, France E-mail:- 2L3i Abstract: Work on the problem of handwritten text recognition in Urdu script has been an active research area. A significant progress is made in this interesting and challenging field in the last few years. In this study, the authors presented a comprehensive survey for a number of offline and online handwritten text recognition systems for Urdu script written in Nastaliq font style from 2004 to 2019. Following features make their contribution worthwhile and unique among the reviews of a similar kind: (i) their review classifies the existing studies based on types of recognition systems used for Urdu handwritten text, (ii) it covers a very different outlook of the recognition process of the Urdu handwritten text at different granularity levels (e.g. character, word, ligature, or sentence level), (iii) this review article also presents each of surveyed articles in following dimensions: the task performed, its granularity level, dataset used, results obtained, and future dimensions, and (iv) lastly it gives the summary of the surveyed articles according to the granularity levels, publishing years, related tasks or subtasks, and types of classifiers used. In the end, major challenges and tasks related to Urdu handwritten text recognition approaches are also discussed in detail. 1 Introduction Handwriting recognition is an active area of research in the field of pattern recognition and has various applications in industrial and professional applications. Some of these applications include forms processing in government, administrative, health and academic institutes; postal address recognition, processing of bank cheques etc. Handwriting recognition concerns with automatic transforming a source language into its symbolic representation. The source language can be represented either in its spatial (offline) or temporal (online) [1] form in graphical marks. In-depth analysis of handwritten text give rise to a number of useful applications such as author profiling [2], named entity recognition [3], recognition of overlapped characters [4] etc. In the late 1950s, the first optical character recognition (OCR) system was developed for the recognition of Latin text [5, 6] that deals with the recognition of numerals only. With the advancements in OCR, the systems available nowadays are expanded to recognise Latin script, and characters of a variety of other languages such as Chinese, Japanese, Arabic, Persian etc. Optical character recognition of Urdu script started in the late 2000 and the first work on Urdu OCR is published in 2004. The literature review identified the fact that there has been a lack of research efforts in Urdu handwritten text recognition as compared to recognition of script of other language [7–9]. Furthermore, there are few Urdu OCR systems for a printed text that are commercially available [10, 11] but there is no system available for Urdu handwritten text recognising to date. In verbal communication, the Urdu language adopted many dialects across the regions but in formal writing Urdu script uses standard way. Furthermore, it is also observed that Urdu written script shares similarities with other languages such as Arabic and Persian [12]. Therefore, automatic interpretation of Urdu handwritten text would have prevalent and ubiquitous benefits. Urdu handwriting is used in writing official assignments and documents in almost all the organisations in Pakistan. Most of the time, all this data (provided in the form of handwritten applications or forms) is typed into computers for further processing that requires huge man power, processing equipment, time, money, and other resources. For example data entry in National Database and IET Image Process. © The Institution of Engineering and Technology 2020 Registration Authority offices in Pakistan for processing requests of National Identity Cards, student's applications in government institutes, signature on banker cheques etc. require an automatic Urdu handwritten text recognition tool to recognise the text and process the documentation in real-time environment. Furthermore, the system should provide facility to save the information in some appropriate database. This will reduce a significant number of resources in daily official matters considering the nature of this task. The development of a Urdu handwritten recognition system can assist in reading historic Urdu manuscripts to make the content of these manuscripts available. The content of manuscripts is written in a clear and readable way as compared to the handwritten text, which makes the task of recognition of contents of manuscript much simpler. On the other hand, some issues associated with handwritten text make the task of developing the system for recognition of handwritten text more challenging and complicated. Some of these issues are differences in writing style (even from the same author), image degradation due to cursive nature of the script, poor quality or illegible handwriting etc. [5, 6, 13]. This survey article gives a comprehensive survey of Urdu handwritten text recognition. Most recent survey articles on this topic were published in 2013 [7, 14, 15]. As mentioned above, this survey article focus on the literature related to offline Urdu handwritten text recognition only. Furthermore, some recent notable work can be found in [16–19] for recognition of printed Urdu text. This paper is organised as follows: a brief introduction of the Urdu script is given in Section 2. In Section 3, we describe basic steps in intelligent character recognition (ICR) by describing the usual process of Urdu text recognition. Furthermore, datasets associated with Urdu handwritten text are also discussed. Then in Section 4, state-of-the-art discussion on the existing work related to Urdu ICR at different granularity levels is given in detail. Open issues, future directions, and detailed analysis of the existing work are presented in Section 5. The conclusion is given in Section 6. 1 Fig. 1 Basic alphabets and numerals of Urdu script more complex and challenging. Furthermore, the writing style of Urdu is also one of the reasons that make the script even more complex to recognise. Naskh and Nastaliq are two widely used fonts while writing Urdu script. Nastaliq is more complex than Naskh [18, 23, 25] since Nastaliq font has a variety of variations in the shape of an alphabet depending on its position in the word as compared to Naskh, e.g. the alphabet (bay) has different shapes according to its position in a word shown in Fig. 3. Above mentioned issues make the recognition of Urdu handwritten text more challenging as compared to any other script. There are some inherent characteristic features of Urdu alphabets that may help in recognising Urdu's handwritten text. These features include accent and diacritical marks that help in differentiating one character from the other (number and position of dots), number and position of loops and arcs etc. It is pertinent to mention that the non-availability of the appropriate resources and required techniques is one of the reasons that makes the recognition of Urdu handwritten text a complex task. 3 Fig. 2 List of non-joiner and joiner alphabets of Urdu script (a) Non-joiner, (b) Joiner alphabets in Urdu script Fig. 3 Different shapes of Urdu alphabet at a different position in a word 2 Urdu script Urdu is the national language of Pakistan and also considered as one of the two official languages of Pakistan [20] (with the other being English). It is widely spoken and understood as a second language by a majority of people of Pakistan [21, 22] and also being adopted increasingly as a first language by the people living in urban areas of Pakistan. Urdu script is written from right to left while numerics are written from left to right, this is the reason Urdu can be considered as one of the bidirectional languages. Urdu script consists of 38 basic letters and 10 numerics as shown in Fig. 1. This alphabet set is also considered as a superset of all Urdu script-based languages alphabets, i.e. the Arabic script contains 28 while the Persian script contains 32 alphabets [23]. Furthermore, the Urdu script also contains some additional alphabets to express the Hindi phonemes. Both the Hindi and Urdu languages [23] have the same phonology with the only difference in the writing style of the script. All the Urdu-script-based languages such as the Arabic and Persian have some unique characteristics, i.e. (i) the script of these languages is written from right to left in cursive style and (ii) the script of these languages is context-sensitive, i.e. written in the form of ligatures, which is a combination of a single or many alphabets. Owing to this context-sensitivity, most of the alphabets have different shapes depending on their position and the adjoining character in the word [10]. The connectivity of alphabets [24] has enriched the Urdu vocabulary with almost 24, 000 ligatures. In Urdu script, alphabets are classified into two groups: joiner and non-joiner [21–23]. The joiner characters join to other characters on the initial, middle, and end position in the ligature. While non-joiner appears in isolated form. Fig. 2 shows a list of the joiner and non-joiner alphabets of Urdu. It is observed that the authors may write the words of the Urdu script in which the individual letters overlap with each other. This kind of overlapping makes the segmentation process of splitting the words to characters 2 Intelligent character recognition (ICR) In the field of computer vision and pattern recognition, ICR is the process of recognition of given handwritten text. However, a complete handwriting recognition system also includes correct segmentation of words into characters, formatting of the extracted segments, and finding the most reasonable words. One of the significant issues in processing the handwritten forms and applications is that the data samples processed by the ICR systems do not necessarily contain the characters in isolated form. In this situation, the segmentation process is used to extract the specific area of interest from the whole image. To resolve this issue and to reduce the segmentation errors, the handwritten forms and applications are designed in a way that the correct position of each character can be calculated easily. The work reported in [26] investigated the scenario that to which extent the reliability provided by the ICR system allows the user to make use of reject option, i.e. to reject the text images having more than one character overlapping due to cursive nature of handwriting, as well as the images containing single characters that are not correctly recognised. This activity helps in designing a tool for the system designer to enhance the efficiency of the system as the volume of forms containing less constrained data fields increases. For the rest of the paper, the term ICR will be used for handwritten text recognition. There are some issues in ICR such as a change in font, the slope of the line, different writing style even from a single writer, overlapping joining letters, missing placement of dots and diacritics aka secondary strokes etc. that make the process of ICR more challenging than the recognition of printed text. Furthermore, the cursive nature of Urdu script makes these issues even more complex and challenging while recognising the handwritten Urdu text. These issues are discussed in detail in subsequent subsections below. 3.1 Urdu-based ICR systems Urdu-based ICR systems can be divided into two types, i.e. online and offline [7, 15, 23]. In online ICR, real-time text recognition is performed using sensors to detect and analyse the pen tip movement, stroke position, baseline detection etc. While in offline, character recognition implicates the automatic conversion of handwritten text from a scanned image of the paper. It is also observed by the researchers that offline character recognition is a complex process than online character recognition [27–29]. In practice, the textual data is given as input in the form of a scanned image analysed and recognised as machine-readable characters. The text data may have different fonts and handwriting styles that need to be preprocessed to produce an ideal and clear view of the input data. ICR is different from OCR in the sense that ICR is associated with handwritten text recognition while OCR works on recognition of printed text. A typical ICR system comprised three phases: (i) preprocessing, (ii) segmentation, and (iii) recognition. In the first phase, a set of operations are IET Image Process. © The Institution of Engineering and Technology 2020 Table 1 Details of Urdu datasets Dataset Statistics Fig. 4 Basic geometrical strokes of Urdu script [12] performed to reduce the ink-noise ratio because the input set of handwritten documents may include inconsistent text because of having different writing styles. These operations include skewness smoothing, chain coding, and baseline removal [7, 8]. In the segmentation phase, the scanned image is segmented at three levels mentioned in [27, 28, 30]. In the first level, the text image is segmented for extracting the baseline by using a profile along the Y -axis, known as vertical projection. In the second phase, the horizontal projection profile is calculated for every row as the sum of all column pixel values inside the row. In the last level, the image segments are further segmented to extract the basic graphical and geometrical components of the text image. To get the finer results from segmentation, one must have to move through all the three levels. In the last phase, recognition is performed in which the segmented text data is scanned and matched with the stored training set data. The application of ICR has increased its efficacy towards automatic recognition of real-world handwritten documents to make them useful for various business and academic applications. 3.2 Datasets related to Urdu handwritten text One of the biggest issues that an artificial intelligence expert encounters—besides designing a model to solve a specific problem —is having an appropriate dataset that directly relates to the problem at hand. Furthermore, the dataset should be processed in a way so that the designed model can make sense of the information. In this subsection, we discuss the datasets associated with Urdu handwritten text and are used in this survey article are discussed in detail. Up to our knowledge, there are five datasets available for Urdu handwritten text namely, Urdu Nastaliq Handwritten Dataset (UNHD) [31], Urdu Printed Text Image Database (UPTI) [12], Centre for Pattern Recognition and Machine Intelligence (CENPARMI) [32], Cursive and Language Adaptive Methodology (CALAM) [33] and Prince Mohammad Bin Fahd University-Urdu/ Arabic Database (PMU-UD) [34]. Among these datasets, UNHD [31] and UPTI [12] are the promising datasets since both the corpus contain an appropriate amount of the respective data instances at the word and character level. These basic instances can help the researchers to perform recognition and classification operations using state-of-the-art machine learning approaches such as deep networks etc. Furthermore, these two datasets may be identified as benchmark datasets since there is no other dataset having such enough data points. Table 1 summarises the detail of the Urdu handwritten data corpus discussed in the literature. 4 Exploratory analysis As mentioned earlier, the current survey is conducted in four broad dimensions namely the designated task completed by the Urdu ICR system; its granularity level; dataset used; and the quality of results obtained. Our aim is to assess in detail the use of different approaches used in the above-mentioned dimensions in the domain of Urdu ICR. In general, we categorised the tasks and issues discussed above based on different levels, as given below. • Character level recognition • Word level recognition • Ligature level recognition IET Image Process. © The Institution of Engineering and Technology 2020 Price total number of writers 500 text lines per page 8 UNHD [31] total number of text lines 10, 000 public total Number of words 312, 000 total number of characters 187, 200 total number of writers 250 UPTI [12] total number of text images 10, 000 US250 text lines per page 6 total number of characters 970, 650 total number of writers 343 CENPARMI [32] total number of text images 19, 432 US500 total samples of Urdu handwritten digits 180 total number of writers 725 total number of images 1200 CALAM [33] total number of text lines 3, 043 US400 total number of words 46, 664 total number of ligatures 101, 181 total number of writers 70 PMU-UD [34] total number of text images 5, 180 public total number of Urdu digits 10 • Sentence level recognition 4.1 Character level recognition It is noteworthy that the efficacy of the approach used in Urdu ICR is dependent on the performance of the method proposed. Therefore, an effective combination of the methods used in Urdu ICR can eventually help in recognising the alphabets correctly. The following two analysis tasks come under the umbrella of Urdu ICR at character level recognition: (i) Urdu alphabet recognition and (ii) Urdu numerals recognition. This section will cover the approaches that focus on the above-mentioned tasks. 4.1.1 Urdu alphabet recognition: Handwritten text recognition at the character level is a challenging task because of having a large number of variations in writing styles (even from a single author). It is observed from the literature related to character level recognition in Urdu script, artificial neural network (ANN) and its different variants are widely used. An ANN [35] is a collection of nodes (aka artificial neurons) linked with each other. These links between artificial neurons are enabled to transmit a signal from one to another within the network. These neurons can process the signals received and then propagate to the neurons connected in subsequent layers. The structure of the ANN may be affected by the kind of information flowing through it because a neural network usually trains itself using the input and labelled output. The problem of developing a generic type of ICR that can resolve the issues associated with any language is challenging since different languages exhibit different characteristic features and thus generalising this type of system is not possible. To overcome this problem, a novel approach is proposed in [29] exploring that the character set of any language can be represented by primitive geometrical strokes. One of the promising features of the approach is that the recogniser (artificial neural network) has to be trained only once. The data structure of the character set should be represented in the form of geometrical strokes in some XML file. This file helps in training the neural network only for once for each word in the language. Fig. 4 shows a set of 13 basic geometrical strokes. For evaluation purposes, a set of 25 handwritten Urdu text samples were tested and achieved the success rate of 75–80%. One of the limitations of this approach is that it does not apply to the words having dots and diacritics. Owing to having a large character (or alphabet) set, there is an inherent similarity among some major strokes as shown in Fig. 4. This similarity is one of the challenging issues in Urdu ICR. 3 Fig. 5 Urdu alphabets (a) Redundancy of strokes among Urdu alphabets, (b) Grouping of characters based on the number of strokes [36] Fig. 6 Features (a) Initial trend for Alif, (b) Character-box-slope for seen, (c) Cusp in hey, (d) Feature vector for intersections of Daal with axes (U: up, D: down, R: right, L: left), (e) Anticlockwise finishing trend of Aen [36] Fig. 7 Alphabets ‘De’ (left) and ‘Daal’ (right) were repeatedly misrecognised Keeping in view the above-mentioned fact, in [36], the authors divided the Urdu alphabet set in four groups according to the number of strokes as shown in Fig. 5b. The authors performed an online Urdu ICR considering singlestroke characters only. Some novel features (shown in Fig. 6) are extracted and then fed to three different classifiers namely backpropagation neural network (BPNN), probabilistic neural network (PNN), and correlation-based classifier. The proposed approach is tested on 85 instances of single stroke characters taken from 35 writers of different age groups. The results showed that the PNN classifier achieved a higher accuracy of 95% as compared to the other two classifiers. Unlikely BPNN, the PNN-based classifiers require no initial training. This is the reason PNN-based classifiers achieved higher accuracy than BPNN. For isolated character recognition, the authors in [37] proposed a technique in which a feature vector is built by analysing the primary and secondary strokes while writing Urdu characters in isolated form. Some of the stroke features that were used to train the classifier are the length of the bounding box diagonal; the angle of the bounding box diagonal; the distance between the first and last point; the sine and cosine of the angle between the first and last point; the total length of the primary stroke; and the total angle traversed. A linear classifier is applied to the data set of five samples each of 38 Urdu alphabets, i.e. a total of 190 characters are provided by two different writers who can write Urdu characters smoothly. The classifier recognised the characters with an error rate of almost 6% because there are some characters that share quite similar shapes (see Fig. 7) and are not correctly recognised. A similar work is reported in [38] by considering the initial half of different Urdu characters. In this work, only those characters were considered that change their shapes with respect to their position and context in a word. Fig. 8 depicts the Urdu alphabets in the initial half forms and are classified on the basis of the number of strokes. Almost 100 native Urdu writers and speakers were invited to write in Urdu script. The writers were provided with stylus and digitising tablet to get the dataset of 3600 instances of 4 Fig. 8 Classification of initial half forms based on the number of strokes [38] Urdu letters in the initial half form. A combination of multilevel one-dimensional wavelet analysis with Daubechies wavelet [39– 42] is applied to extract features from these instances. A number of neural networks with different configuration were trained for recognition purposes. Among these networks, BPNN provided a maximum recognition rate of 92%. In [43], the authors introduced a novel similar character discrimination method for recognition of handwritten Urdu character in online mode. The proposed methodology is a three-step model namely the pre-classification, feature extraction, and fine classification process. In the preclassifier phase, the discrimination of similar characters is performed by making smaller subsets based on stroke number and diacritics. Secondly, the structural features and wavelet features are extracted manually. Finally, a group of different machine learning classifiers such as support vector machines (SVMs), ANNs, and recurrent neural network (RNN) classifiers are applied separately and the obtained results are compared for fine classification within subsets. The results showed that the RNN classifier, when applied without using the proposed pre-classifier and features, can be applied to check the end-to-end the capability of the RNN classifier. Furthermore, the experimental results depict that the proposed method is efficient and achieves an overall accuracy of 96% on a large-scale self-collected dataset. It is pertinent to mention that the proposed approach is also feasible to apply to other Arabic-based scripts. Multidimensional long short term memory (MDLSTM) neural network is one of the RNN that is implicitly used for sequence learning and segmentation in a multidimensional environment [44– 46]. This model was used for the first time in the work of [47] for Urdu script recognition. One of the promising features of the model IET Image Process. © The Institution of Engineering and Technology 2020 Table 2 Comparison of the results obtained on UPTI dataset using different methods Reference Features Approach UPTI dataset Accuracy, % Fig. 9 Tagged sample from UPTI dataset (a) Text line image, (b) Ground truth or transcription [18] [47] pixels MDLSTM [49] pixels BLSTM [50] pixels BLSTM [51] statistical features MDLSTM 68% training 16% validation 16% 46% 44% validation 10% 46% training 34% validation 20% 46% 16% validation 16% test - 94.85 94.7 Table 3 Accuracy reported on common datasets Dataset Reference Accuracy, % Approaches UPTI [12] Fig. 10 Different samples of input images with noise and without noise [31] Fig. 11 Grouping of Urdu characters according to shape similarity. Each group is numbered from right to left [55] is that it can scan the input image in all four directions thus reducing the chance of ambiguity. For evaluation purposes, UPTI (Urdu Printed Text Image) dataset [12] is used that contains 10, 000 scanned images of both Urdu handwritten and printed text. MDLSTM is one of the supervised techniques, therefore, each input sample in the dataset is tagged and labelled with the appropriate information, shown in Fig. 9. The dataset is further divided according to the following ratio: 68% for training, 16% for both testing and validation purposes. To evaluate the accuracy of the proposed approach, Levenshtein edit distance [48] is computed between the output text and base-line results and achieved the accuracy of 98% as compared to results reported in the work of [49, 50] reporting 88.94 and 89.00% accuracy, respectively. Table 2 shows a comparison of the proposed approach on the UPTI dataset [12] with other techniques. A promising work is reported in [31] in which Urdu handwritten text is recognised using UNHD [52]. This dataset https://sites.google.com/site/researchonurdulanguageresearch on the Urdu language can be accessed publicly. The dataset contains 312, 000 words (including both Urdu script and Urdu numerals) written on a total of 10, 000 lines by 500 writers of different age groups. The writers are directed to write on white pages of size A4. Each individual is provided six blank pages labelled with author ID and the page number. One of the samples of written pages is shown in Fig. 10. Furthermore, to maintain the uniformity in data, the writers were asked to write the provided printed text. To recognise the text, the bidirectional long short term memory (BLSTM)-based approach is proposed that is based on a RNN capable of restoring the previous sequence information. For evaluation purposes, the dataset is divided into 50% for training, 30% for validation, and 20% for testing and achieved 6-8% error IET Image Process. © The Institution of Engineering and Technology 2020 UNHD CENPARMI [32] [53] [54] [31] [9] - MDLSTM (Leven. Dist) MDLSTM (CTC) BLSTM SVM rate that can be improved using two-dimensional BLSTM as proposed by the authors. Table 3 gives the summary of accuracy reported on common datasets in the Urdu domain. In [53], the authors proposed a novel approach for Urdu text recognition at the character level, written in Nastaliq font by combining the convolution neural network (CNN) and MDLSTM (convo-recursive deep learning model). In the first phase, CNN is deployed to extract the characteristic features which are then fed to MDLSTM in the second phase. This approach outperformed the state-of-the-art systems on the UPTI data set. In a very recent work [55], the authors proposed the use of the convolutional neural network to recognise the multi-font offline Urdu handwritten characters in an unconstrained environment. They also introduced a novel dataset of Urdu handwritten characters and numerals by inviting a number of native Urdu speaking people from different age groups and academic level. They grouped the Urdu isolated characters based on the shape similarity, as shown in Fig. 11. A series of experiments were performed on their proposed dataset. The accuracy achieved for character recognition is among the best while comparing with the ones reported in the literature for the same task. 4.1.2 Urdu numeral recognition: It is quite easy for a human being to recognise the handwritten digit data but for the computer system, there is a need for an intelligent approach based on some machine learning algorithms developed for this kind of job. The digit writing stroke, length, width, orientation, and other geometrical features tend to change while writing the same digit even by the same author. These different writing styles may introduce shape variations of Urdu numerals that may break the strokes primitives and also change their topology. These issues make Urdu handwritten digit recognition one of the active research areas in the field of image processing. Unfortunately, there is no commercially available standard dataset of Urdu numerals. Owing to this lack of resource, the researchers developed their own dataset and concluded the results. This section covers some notable work related to handwritten digit recognition in the Urdu domain. In [56], different transformations of Daubechies wavelet [39– 42] are applied for features extraction from the dataset of about 2150 samples of handwritten Urdu digits. A sample of the dataset is shown in Fig. 12 above. For evaluation purposes, 2000 samples were used for training the neural network and 150 instances for testing. To decompose the images into different frequency bands, both the low-pass and high-pass filtering are applied at each phase 5 Fig. 14 Grayscale and gradient strength of Urdu ligature at different level (a) Greyscale image of size 128 × 128, (b) Gradient strength, (c) Gradient direction, (d) Division of gradient image into 9 × 9 blocks [9] Table 4 Experimental results with different image sizes and gradient features [9] Image size Feature size Results Fig. 12 Handwritten Urdu digit samples [56] 64 × 64 64 × 128 128 × 64 64 × 128 - 94.96% (3580/-% (3590/-% (3585/-% (3621/3770) Fig. 15 Sample sentence in Urdu script. Underlined words are showing named entities 4.2 Word level recognition Fig. 13 Recognition rate using different Daubechies wavelets for handwritten digits [56] of Daubechies wavelet [39–42] filtering. For classification purposes, BPNN is used and achieved an average recognition rate of 92.05% as shown in Fig. 13. The recognition time (in seconds) are 1.0247, 1.1015, 2.003, 1.4659, 1.501, 1.1163, 1.4792, 1.508, 1.5725, and 1.4754, respectively, for each of the ten wavelet filters, thus achieving the average training time of 1.4245. In [25, 57], the authors presented the similarities and dissimilarities between Urdu and Arabic script while recognition of handwritten numeric data. A hybrid technique of hidden Markov model (HMM) and the fuzzy rule is used to recognise the handwritten digit of both Arabic and Urdu script. The dataset is prepared by inviting 30 trained users to write both the Urdu and Arabic numerals and collected 900 samples in total. The system obtained 97, 96, and 97.8% recognition rates using fuzzy rule, HMM, and hybrid approach, respectively. The authors also conclude that the separation of numerals from Urdu text in a handwritten text is still a challenging issue due to having shape similarity, e.g. first alphabet of Urdu script (Alif) and Urdu numeric (One) both are represented by the same symbol of . Similarly, in [34], a multi-language handwritten numeral recognition system is proposed using novel structural features. A total of 65 local structural features are extracted and several classifiers are used for testing numeral recognition. Random forest was found to achieve the best results with an average recognition of 96.73%. The proposed method is tested on six different popular languages, including Arabic Western, Arabic Eastern, Persian, Urdu, Devanagari, and Bangla. It is pertinent to mention that in this study, only those digits in the languages are chosen that do not resemble each other. Yet using the novel feature extraction method, a high recognition accuracy rate is achieved. The experiments are performed on well-known available datasets of each language. A dataset for the Urdu language is also developed in this study and introduced as PMU-UD (https://archive.ics.uci.edu/ml/datasets/ PMU-UD). The results indicate that the proposed method gives high recognition accuracy than other methods. 6 Recognition of cursive script such as Urdu is performed by analysing the fundamental units, i.e. the individual letters that combine to form words. Several techniques are used so far in this field to enhance recognition results. This section covers notable work related to the recognition of Urdu handwritten text at the word level. In [9], a set of gradient and structural features were extracted and fed to SVM to recognise Urdu handwritten words from a given text. Robert's filter mask [58] is applied for extracting gradient features from a 128 × 128 grey level image of handwritten text. To overcome noise and spatial resolution in the image, Gaussian filters [59] are used to normalise image as 9 × 9 scales as shown in Fig. 14. Structural features of an image including projection profile, topological features, upper and lower profiles etc. are extracted by a calculating distance of pixel positions from the upper right and upper left corner of the image. These calculations are then normalised using Gaussian-like estimations. For evaluation purpose, CENPARMI dataset [32] is divided into training (60%), validation (30%), and testing (20%). Table 4 depicts the experimental results with different image sizes and gradient features. Name-entity recognition (NER) is a task to extract and identify real-world objects such as persons, organisations, locations etc. that can be represented with a proper name [60]. For example in Fig. (Bahawalpur: name of city) and 15, the underlined words (Punjab: Name of the state of Pakistan) are the named entities in Urdu. Two approaches are used for the development of NER systems in the Urdu domain [61] namely (i) rule-based approach; (ii) statistical approaches. Rule-based approaches are based on gazetteer lists manually written by linguistics. This approach is language-dependent, i.e. the techniques used to develop a rulebased system for one language cannot be applied to other languages. The authors also performed online recognition of Urdu handwritten text using the tree-based dictionary approach and achieved a 96% accuracy rate. In [62], the authors used a rule-based system and achieved an accuracy rate of 75% in extracting 13 different named entities from two different sets of Urdu documents collected from news resources. A similar approach is proposed in [63] in which about 2000 Urdu documents were analysed for extracting different named entities thus achieving 73% accuracy. IET Image Process. © The Institution of Engineering and Technology 2020 In statistical approaches for handwritten text recognition, different machine learning models such as HMM, SVMs, and Bayesian decision theory etc. are also used. These approaches are language independent and easily applicable to multiple languages as well. In [64], the authors used two different models maximum entropy (ME) and conditional random field (CRF) for information extraction. For evaluation purposes, a corpus of 50, 000 tokens was carefully analysed to extract a number of different named entities. F-measure for ME-based approach was 55.30% and that of the CRF-based approach was 68.90%. A multilingual approach is proposed in [65] in which the authors developed a statistical CRF model for extracting named entities from South and Southeast Asian languages particularly for Bengali, Hindi, Telugu, Oriya, and Urdu. For Bengali and Hindi languages, the gazetteer lists were also used. The system achieved F-measure of 59.39% for Bengali, 33.12% for Hindi, 28.71% for Oriya 4.749% for Telugu and 35.52% for Urdu. Table 5 summarises different approaches used in developing the Urdu NER system. It is observed that modern mobile phones are equipped with pointing devices to write the script in any language instead of typing. This activity attracts the researchers to develop a system that can read such a type of text in real-time. To develop such a system to recognise the Urdu handwritten text, the authors proposed a mobile-based system [66] using a number of various classifiers such as HMM, fuzzy logic, k-nearest neighbour (k-NN), hybrid HMM fuzzy, hybrid k-NN fuzzy, and CNN. The statistical parameters are checked and validated to examine the performance of the classifiers. The experimental results concluded that by applying the proposed preprocessing algorithms using a rule-based approach, the recognition rates are impressively enhanced. The vector quantisation method provided better results for recognition of online Urdu handwritten characters while the hybrid k-NN and fuzzy classifier showed promising results for recognition of ligatures and words in mobile phones. In [67], the authors presented a quite different approach for reading the cursive handwriting. The method includes the process of finding patterns in order of writing from the static images of handwriting. In other words, they designed a novel segmentation algorithm decomposing the ‘unfolded’ ink into strokes. A novel approach is introduced to compare the ink of and unknown handwriting with the pattern set of reference words and a graph search algorithm to search for the best interpretation among the possible ones. Then compare the ink-matching patterns to the ink of the unknown handwriting with the set of reference words (or terms). These reference words are written by whom the transcripts are given. Then the graph search algorithm is applied to search for the best interpretation among the possible ones. The novelty lies in the fact that the proposed method does not involve any feature extraction, nor a classification stage and may benefit from a linguistic context, if available. The authors reported the results of experiments on 8000 samples written in different cursive languages and draw some conclusions and outline further developments. The process of recognition at word level requires unicode formatting and usage of proper font due to the cursive nature of Urdu script. Furthermore, there is a need for new techniques such as deep neural networks with enhanced mapping functions to upgrade the existing ones to resolve the issues for recognition of Urdu text at the word level. 4.3 Ligature level recognition As mentioned earlier, ligatures are formed by combining alphabets in Urdu script. Some examples of the ligature are given in Fig. 3. While writing ligature in Urdu script, alphabets may change their shape based on the the position in the word (or ligature). In this section, work done in Urdu ICR at the ligature level is reviewed in detail. In [68], the authors proposed a novel segmentation free approach for online recognition of Urdu handwritten script by using different features of ligatures. These features include position and number of the loop(s), number of intersections, the direction of ligature etc. The ligatures can be written by using primary and secondary strokes. For stroke identification, the authors used the backpropagation neural network. The proposed system was trained on a set of 240 ligatures written by different native Urdu writers. A total of 860 ligatures were tested, for evaluation purposes, with an accuracy rate of 93% (base ligatures) and 98% (ligature of secondary strokes). A similar task is performed in [69] using two different methods, the HMM and fuzzy logic classifier. Input strokes were initially divided into 62 classes based on the starting and ending style of ligature, using fuzzy rules. For the base stroke ligatures, 26 temporal features were extracted. To reduce time complexity, these features were truncated to 16 characteristic features. About 1800 ligatures are collected from 15 trained and native Urdu writers and were classified with an accuracy rate of 87.65%. A novel approach is proposed in a recent work reported in [70] in which the bottom-up technique is used to group the connected components of ligatures written on the same line of the cursive text document. The traced sequence of the connected components is stored as a feature vector and labelled with appropriate indices. The proposed approach is tested on various samples of the dataset generated by inviting native Urdu writers. Accuracy of about 96% is achieved from experimental results in recognising ligatures. In a notable work of [4], the authors performed a similar task on Urdu printed text using a cross-correlation technique on a small subset of Nastaliq text. In a recent work [71], a holistic OCR system is proposed for printed Urdu Nastaliq font using statistical features and hidden Markov models to recognise Urdu ligatures. For training purposes, 1525 unique high-frequency Urdu ligature clusters from the standard UPTI database are selected. The overall ligature dataset is first to split into primary and secondary ligatures and is recognised separately. In the second phase, the heuristics approach is applied to recognise the complete ligature by associating secondary ligatures with the primary ligature. For evaluation purposes, the model is assessed through various experiments thus achieving 92.20% accuracy compared to the recent studies on this problem. In a noteworthy work [72], the authors developed an Urdu OCR system that claims to recognise the Urdu handwritten text at the ligature level. The important achievement in this approach is that it resolves the issues associated with the segmentation problems while pre-processing the Urdu handwritten text having overlapping characters. The proposed approach is solely based on a semisupervised multi-level clustering to categorise and recognise the ligatures. The classification process is then performed using four widely used machine learning techniques, i.e. decision trees, linear discriminant analysis, naive Bayes, and k-nearest neighbour (kNN). The obtained results showed an overall accuracy of 62, 61, 73, and 90% for the decision tree, linear discriminant analysis, naive Bayes, and k-NN, respectively. 4.4 Sentence level recognition Table 5 Summary of approaches used for developing Urdu NER system Reference Approach used F-measure, % Corpus [62] [63] [64] [65] rule-based rule-based ME, CRF CRF -,-,- 1,62,275 tokens 36,000 tokens 55,000 tokens 36,000 tokens IET Image Process. © The Institution of Engineering and Technology 2020 In [54], the authors utilised the implicitly based recognition technique for Urdu text in the Nastaliq style. Multidimensional RNN (MD-RNN) in a combination with long short term memory (LSTM) and connectionist temporal classification (CTC) output layers is trained on the features extracted by the sliding window on Urdu text lines. The reason behind using MD-RNNs is to replace the single recurrent connection found in conventional RNNs with as many recurrent connections as the number of dimensions in the data. During the forward pass at each point in the data sequence, 7 Fig. 16 MD-RNN-based recognition system for Urdu Nastaliq [54] Table 6 Distribution of the articles published year-wise Year References No. of articles- total [29] [61] [17, 18, 68] [4, 74] [25, 36, 37, 57] [9, 19, 63–65, 69] [3, 24, 62] [38, 60, 70] [56] [47, 54, 73] [11, 31, 53, 71] [43, 72, 75] [55, 66] - the hidden layer of the network receives both an external input and its own activation from one step back along all dimensions. Fig. 16 depicts the overall system, which is divided into three stages: input, MD-LSTM, and output layers. In the first phase, the overlapped sliding windows are organised to extract the features on a greyscale image of the Urdu text line. As a result, a sequence of feature vectors is generated for each text line. Secondly, the sequence of the feature vector of each text line is given to MD-LSTM for training purposes. This sequence is finally assigned to appropriate labels using the CTC layer. This methodology is applied to UPTI dataset [12] and acquired an accuracy rate of 96.4% as compared with some other recognition systems using the same technique. A similar task is performed in [73] by introducing the features of zones (a variant of implicit segmentation approach) that are extracted using an n × n sliding window of uniform size. These features are extracted by calculating the density of pixels, statistical features of the pixels that represent relevant information about the character etc. To get the same number of zoning regions of the same size, the greyscale images of Urdu text lines were normalised to fix height and width size (30, 50, 70, and 90). 8 Table 7 Distribution of articles based on classifier used Classification Ref. No. of technique articles neural networks rule based on HMM SVM others [31, 36, 47, 53–56, 68, 73] [43, 66, 75] [25, 57, 61–64, 69, 76] [9, 65] [29, 37, 70, 72, 74] 11 8 2 5 For each zone i, two features f 1i and f 2i are computed. The value of f 1i is the ratio between the sum of all pixel intensity values in the ith zones (ni) and the sum of all pixel intensity in the image (∑ ni) as given in (1). Similarly, the value of f 2i is the ratio between the sum of all pixel intensity values in the ith zones (ni) and the number of pixels in the ith zone ( f size), as described in (2) f 1i = ni ∑ ni (1) f 2i = ni f size (2) The authors also conclude that the values of zoning features are prone to distortion and noise due to the cursive property of the Nastaliq style. Equation (3) depicts the feature vector that is based on top-to-bottom and right-to-left scan values using the LSTM sequence learning property. Where FV is the feature vector and Fx, y is the zoning feature of the zone x,y. Experiments were performed on UPTI dataset [12] achieving a recognition rate of 93.38%. Tables 6 and 7 depict the distribution of articles based on the publishing year and classification techniques discussed in this study, respectively FV = F1, 1F1, 2, …F1, n; F2, 1, F2, 2, …F2, n (3) In a very recent work reported in [75], a group of researchers proposed a novel gated BLSTM (GBLSTM) model for recognition of both the printed and handwritten Urdu Nastaleeq text based on IET Image Process. © The Institution of Engineering and Technology 2020 Table 8 Distribution of articles based on the Urdu handwritten recognition tasks performed at different granular levels Level of References No. of analysis articles character level word level ligature level sentence level alphabet: [29, 31, 36–38, 47, 49, 50, 53] [43, 55] numeric: [25, 55–57] [9, 60–66] [4, 68–72] [54, 73, 75] 11 4 8 6 3 ligature information. The novelty lies in incorporating the raw pixel values as the characteristic features instead of human crafted features since the latter being more error-prone. For evaluation purposes, the proposed model is trained and tested on un-degraded and tested on a novel but artificially degraded versions of Urdu printed and handwritten text images dataset. The recognition accuracy of the proposed GBLSTM model is 96.71%, i.e. higher than the prevalent Urdu OCR systems. Table 8 shows the distribution of articles based on the granularity level discussed in this study. 5 Discussion This study reviews a number of interesting and useful works that primarily focused on challenging issues associated with Urdu handwritten text recognition. Table 6 depicts the year-wise distribution of articles surveyed for this state-of-the-art review and it is quite obvious that researchers showed more interest in Urdu ICR during-. If we consider granule-based Urdu handwritten text recognition, most of the studies focused on character and word level as shown in Table 8. Moreover, it was observed that a very limited work done on sentence and ligature level. It is evident from Table 7 that ANN is the most widely used approach to resolve various tasks since the ANN approach is more efficient and coherent for solving the problems having high-dimensional data set as compared to other techniques. 5.1 Future directions and open challenges In this section, we will discuss some open issues related to Urdu ICR. The data set developed and collected from a number of resources is so much noisy and unstructured that it causes maximum time consumption while preprocessing. One of the major challenges is the validation of the proposed methods given in the surveyed articles. The researchers must have access to some standard benchmarks dataset results to validate their proposed methods. Furthermore, very few attempts have been made to extract the optimised features that might help in achieving higher recognition accuracy. Therefore, various hybrids of machine learning and optimisation techniques can be developed for feature subset selection. To extract the characteristic features, both structural and statistical pixel-based approaches are not utilised. To excerpt the correct and targeted features, various machine learning and optimisation techniques can be combined. Some more open issues are recognition of ambiguous, overlapped, and illegible words and separation of numeric data from the script in handwritten text. 6 Conclusion Urdu script is categorised as one of the cursive and bidirectional script derived from Arabic, i.e. the reason the Urdu script shares with the Arabic almost similar challenges and issues but with higher complexity. This complexity is due to the usage of diacritics, which changes the alphabet from the Arabic alphabet set IET Image Process. © The Institution of Engineering and Technology 2020 to the Urdu alphabet set, e.g. (Kaaf) can be misinterpreted and recognised as (Gaaf) because of having a quite similar shape. The character of (Gaaf) is not in the Arabic alphabet set. Similar issues are associated with the recognition of Urdu alphabets, which share similar shapes both in Urdu and Arabic character sets. This study presents an all-inclusive and state-of-the-art review of the research published in various aspects of Urdu handwritten text recognition from 2004 to 2019. The survey is reviewed at four granularity levels namely character level, word level, ligature level, and sentence level. It has also been observed that the literature presented in the study covered the significant progress made in the field of recognition of printed Urdu text only rather than recognition of handwritten Urdu script. This is due to the inherent cursive nature and different writing styles of Urdu handwritten text. One of the main contributions of the survey is to highlight a list of available public datasets for Urdu handwritten text recognition. Furthermore, Urdu character recognition in Nastaliq script is a two-level process; first, the text is segmented into lines and then the lines are combined to make ligatures and isolated characters. Segmentation of ligatures into basic character shapes is more challenging and still an open field for researchers in this area. We also conferred some significant conclusions from the review regarding the techniques and approaches applied. Apart from ANN and SVM; we analysed that some of the ensemble learning approaches and linear-time algorithms like random forest, connected components, pixel-based structural, and statistical feature extraction using formal concept analysis. Furthermore, the deep network and its different variations must be applied to address the challenging issues in a disciplined way. Some of these issues are uncertainty in recognition, limited number of training samples etc. Therefore, there must be an attempt to apply the abovementioned approaches to resolve these issues. 7 [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] References Bahlmann, C.: ‘Directional features in online handwriting recognition’, Pattern Recognit., 2006, 39, (1), pp. 115–125 Anwar, W., Bajwa, I.S., Ramzan, S.: ‘Design and implementation of a machine learning-based authorship identification model’, Sci. Program., 2019, 2019, pp. 98–110 Jahangir, F., Anwar, W., Bajwa, U.I., et al.: ‘N-gram and gazetteer list based named entity recognition for Urdu: a scarce resourced language’. Proc. 10th Workshop on Asian Language Resources, Mumbai, India, 2012, pp. 95–104 Sattar, S.A., Haque, S., Pathan, M.K., et al.: ‘Implementation challenges for Nastaliq character recognition’. Int. Multi-Topic Conf., Jamshoro, Pakistan, 2008, pp. 279–285 Mori, S., Nishida, H., Yamada, H.: ‘Optical character recognition’ (John Wiley & Sons Inc., USA, 1999) Herbert, H.F.: ‘The history of OCR, optical character recognition’ (Recognition Technologies Users Association, Manchester Center, VT, 1982) Khan, N.H., Adnan, A., Basar, S.: ‘An analysis of off-line and on-line approaches in Urdu character recognition’. Proc. 15th Int. Conf. on Artificial Intelligence, Knowledge Engineering and Data Bases (AIKED'16), Venice, Italy, 2016, pp. 29–31 Jan, Z., Shabir, M., Khan, M.A., et al.: ‘Online urdu handwriting recognition system using geometric invariant features’, Nucleus, 2016, 53, (2), pp. 89–98 Sagheer, M.W., He, C.L., Nobile, N., et al.: ‘Holistic Urdu handwritten word recognition using support vector machine’. 2010 20th Int. Conf. on Pattern Recognition, Istanbul, Turkey, 2010, pp- Wali, A., Hussain, S.: ‘Context sensitive shape-substitution in Nastaliq writing system: analysis and formulation’. Innovations and Advanced Techniques in Computer and Information Sciences and Engineering, Bridgeport, CT, USA, 2007, pp. 53–58 Akram, Q.U.A., Hussain, S.: ‘Ligature-based font size independent OCR for Noori Nastalique writing style’. 2017 1st Int. Workshop on Arabic Script Analysis and Recognition (ASAR), Nancy, France, 2017, pp. 129–133 Sabbour, N., Shafait, F.: ‘A segmentation-free approach to Arabic and Urdu OCR’ in ‘Document recognition and retrieval XX’, vol. 8658 (International Society for Optics and Photonics, USA, 2013), pp. p 86580N Baldominos, A., Saez, Y., Isasi, P.: ‘A survey of handwritten character recognition with MNIST and EMNIST’, Appl. Sci., 2019, 9, (15), p. 3169 Pal, U., Jayadevan, R., Sharma, N.: ‘Handwriting recognition in indian regional scripts: a survey of offline techniques’, ACM Lang. Inf. Process. (TALIP), 2012, 11, (1), p. 1 Jameel, M., Kumar, S., Karim, A.: ‘A review on recognition of handwritten Urdu characters using neural networks’, Int. J. Adv. Res. Comput. Sci., 2017, 8, (9), pp. 727–730 Pal, U., Sarkar, A.: ‘Recognition of printed urdu script’. 2003 Proc. Seventh Int. Conf. on Document Analysis and Recognition, Edinburgh, UK, 2003, pp- Shamsher, I., Ahmad, Z., Orakzai, J.K., et al.: ‘OCR for printed Urdu script using feed forward neural network’, Proc. World Acad. Sci. Eng. Technol., 2007, 23, pp. 172–175 9 [18] [19] [20] [21] [22] [23] [24] [25] [26] [27] [28] [29] [30] [31] [32] [33] [34] [35] [36] [37] [38] [39] [40] [41] [42] [43] [44] [45] [46] [47] 10 Ahmad, Z., Orakzai, J.K., Shamsher, I., et al.: ‘Urdu Nastaleeq optical character recognition’, Proc. World Acad. Sci. Eng. Technol., 2007, 26, pp. 249–252 Javed, S.T., Hussain, S., Maqbool, A., et al.: ‘Segmentation free Nastalique urdu ocr’, World Acad. Sci. Eng. Technol., 2010, 46, pp. 456–461 Simons, G.F., Fennig, C.D.: ‘Ethnologue: languages of asia’ (Sil International, USA, 2017) Rahman, T.: ‘Language and politics in pakistan’, Language, 1998, 133, p. 9 Mahboob, A., Jain, R.: ‘Bilingual education in india and pakistan’ in ‘Bilingual and multilingual education’ (Springer, Cham, USA, 2016), pp. 1– 14 Razzak, M.I.: ‘Online urdu character recognition in unconstrained environment’. PhD thesis, International Islamic University, Islamabad, 2011 Lehal, G.S.: ‘Choice of recognizable units for Urdu OCR’. Proc. Workshop on document analysis and recognition, Mumbai, India, 2012, pp. 79–85 Razzak, M.I., Hussain, S.A., Sher, M.: ‘Numeral recognition for Urdu script in unconstrained environment’. 2009 Int. Conf. on Emerging Technologies, Islamabad, Pakistan, 2009, pp. 44–47 De-Stefano, C., Fontanella, F., Marcelli, A., et al.: ‘Rejecting both segmentation and classification errors in handwritten form processing’. 2014 14th Int. Conf. on Frontiers in Handwriting Recognition, Heraklion, Greece, 2014, pp. 569–574 dos Santos, R.P., Clemente, G.S., Ren, T.I., et al.: ‘Text line segmentation based on morphology and histogram projection’. 2009 10th Int. Conf. on Document Analysis and Recognition, Barcelona, Spain, 2009, pp. 651–655 Marinai, S., Nesi, P.: ‘Projection based segmentation of musical sheets’. Proc. of the Fifth Int. Conf. on Document Analysis and Recognition (ICDAR'99) (Cat. No. PR00318), Bangalore, India, 1999, pp. 515–518 Ali, A., Ahmad, M., Rafiq, N., et al.: ‘Language independent optical character recognition for hand written text’. Proc. 8th Int. Multitopic Conf. 2004 (INMIC 2004), Lahore, Pakistan, 2004, pp. 79–84 Malik, S.A., Maqsood, M., Aadil, F., et al.: ‘An efficient segmentation technique for Urdu optical character recognizer (OCR)’. Future of Information and Communication Conf., San Francisco, CA, USA, 2019, pp. 131–141 Ahmed, S.B., Naz, S., Swati, S., et al.: ‘Handwritten urdu character recognition using one-dimensional BLSTM classifier’, Neural Comput. Appl., 2019, 31, (4), pp- Sagheer, M.W., He, C.L., Nobile, N., et al.: ‘A new large urdu database for off-line handwriting recognition’. Int. Conf. on Image Analysis and Processing, Vietri sul Mare, Italy, 2009, pp. 538–546 Choudhary, P., Nain, N.: ‘Calam: linguistic structure to annotate handwritten text image corpus’, ‘Computational intelligence in data mining’, vol. 2 (Springer, India, 2015), pp. 449–460 Alghazo, J.M., Latif, G., Alzubaidi, L., et al.: ‘Multi-language handwritten digits recognition based on novel structural features’, J. Imaging Sci. Technol., 2019, 63, (2), pp- Van Gerven, M., Bohte, S.: ‘Artificial neural networks as models of neural information processing’, Front. Comput. Neurosci., 2017, 11, p. 114 Haider, I., Khan, K.U.: ‘Online recognition of single stroke handwritten Urdu characters’. 2009 IEEE 13th Int. Multitopic Conf., Islamabad, Pakistan, 2009, pp. 1–6 Shahzad, N., Paulson, B., Hammond, T.: ‘Urdu qaeda: recognition system for isolated urdu characters’. Proc. UI Workshop on Sketch Recognition, Sanibel Island, Florida, 2009 Safdar, Q.-T.-A., Khan, K.U.: ‘Online urdu handwritten character recognition: initial half form single stroke characters’. 2014 12th Int. Conf. on Frontiers of Information Technology, Islamabad, Pakistan, 2014, pp. 292–297 Meurant, G.: ‘‘Wavelets: a tutorial in theory and applications’, vol. 2 (Academic Press, USA, 2012) Yelampalli, P.K.R., Nayak, J., Gaidhane, V.H.: ‘Daubechies wavelet-based local feature descriptor for multimodal medical image registration’, IET Image Process., 2018, 12, (10), pp- Daubechies, I.: ‘The wavelet transform, time-frequency localization and signal analysis’, IEEE Trans. Inf. Theory, 1990, 36, (5), pp- Daubechies, I.: ‘Ten lectures on wavelets’, vol. 61 (SIAM, USA, 1992) Safdar, A., Ullah Khan, K., Peng, L., et al.: ‘A novel similar character discrimination method for online handwritten urdu character recognition in half forms’, Sci. Iranica, 2018, to appear Kalchbrenner, N., Danihelka, I., Graves, A.: ‘Grid long short-term memory’, arXiv preprint arXiv:-, 2015 Hochreiter, S., Schmidhuber, J.: ‘Long short-term memory’, Neural Comput., 1997, 9, (8), pp- Sak, H., Senior, A., Beaufays, F.: ‘Long short-term memory recurrent neural network architectures for large scale acoustic modeling’. Fifteenth Annual Conf. of the Int. Speech Communication Association, Singapore, Singapore, 2014 Naz, S., Umar, A.I., Ahmed, R., et al.: ‘Urdu Nasta'liq text recognition using implicit segmentation based on multi-dimensional long short term memory neural networks’, SpringerPlus, 2016, 5, (1), p. 2010 [48] [49] [50] [51] [52] [53] [54] [55] [56] [57] [58] [59] [60] [61] [62] [63] [64] [65] [66] [67] [68] [69] [70] [71] [72] [73] [74] [75] [76] Yujian, L., Bo, L.: ‘A normalized Levenshtein distance metric’, IEEE Trans. Pattern Anal. Mach. Intell., 2007, 29, (6), pp- Ahmed, S.B., Naz, S., Razzak, M.I., et al.: ‘Evaluation of cursive and noncursive scripts using recurrent neural networks’, Neural Comput. Appl., 2016, 27, (3), pp. 603–613 Ul-Hasan, A., Ahmed, S.B., Rashid, F., et al.: ‘Offline printed Urdu Nastaleeq script recognition with bidirectional LSTM networks’. 2013 12th Int. Conf. on Document Analysis and Recognition, Washington, DC, USA, 2013, pp- Naz, S., Umar, A.I., Shirazi, S.H., et al.: ‘Segmentation techniques for recognition of Arabic-like scripts: a comprehensive survey’, Educ. Inf. Technol., 2016, 21, (5), pp- Gosselin, B.: ‘Multilayer perceptrons combination applied to handwritten character recognition’, Neural Process. Lett., 1996, 3, (1), pp. 3–10 Naz, S., Umar, A.I., Ahmad, R., et al.: ‘Urdu Nasta'liq text recognition system based on multi-dimensional recurrent neural network and statistical features’, Neural Comput. Appl., 2017, 28, (2), pp. 219–231 Naz, S., Umar, A.I., Ahmad, R., et al.: ‘Offline cursive Urdu-Nastaliq script recognition using multidimensional recurrent neural networks’, Neurocomputing, 2016, 177, pp. 228–241 Husnain, M., Saad Missen, M.M., Mumtaz, S., et al.: ‘Recognition of Urdu handwritten characters using convolutional neural network’, Appl. Sci., 2019, 9, (13), p. 2758 Borse, R., Ansari, I.: ‘Offline handwritten and printed urdu digits recognition using Daubechies wavelet’ (ER Publication, New Delhi, India, 2015) Razzak, M.I., Hussain, S.A., Belaïd, A., et al.: ‘Multi-font numerals recognition for urdu script based languages’ (International Journal of Recent Trends in Engineering (IJRTE), Academy publisher, Inria-, 2009) Davis, L.S.: ‘A survey of edge detection techniques’, Comput. Graph. Image Process., 1975, 4, (3), pp. 248–270 Young, I.T., Van Vliet, L.J.: ‘Recursive implementation of the Gaussian filter’, Signal Process., 1995, 44, (2), pp. 139–151 Naz, S., Umar, A.I., Shirazi, S.H., et al.: ‘Challenges of urdu named entity recognition: a scarce resourced language’, Res. J. Appl. Sci., Eng. Technol., 2014, 8, (10), pp- Malik, S., Khan, S.A.: ‘Urdu online handwriting recognition’. 2005 Proc. IEEE Symp. on Emerging Technologies, Islamabad, Pakistan, 2005, pp. 27– 31 Singh, U.P., Goyal, V., Lehal, G.S.: ‘Named entity recognition system for Urdu’. Proc. COLING 2012, Mumbai, India, 2012, pp- Riaz, K.: ‘Rule-based named entity recognition in urdu’. Proc. 2010 Named Entities Workshop, Stroudsburg, PA, USA, 2010, pp. 126–135 Mukund, S., Srihari, R., Peterson, E.: ‘An information-extraction system for urdu—a resource-poor language’, ACM Trans. Asian Lang. Inf. Process. (TALIP), 2010, 9, (4), p. 15 Ekbal, A., Bandyopadhyay, S.: ‘Named entity recognition using support vector machine: a language independent approach’, Int. J. Electr. Comput. Syst. Eng., 2010, 4, (2), pp. 155–170 Anwar, F.: ‘Online urdu handwritten text recognition for mobile devices using intelligent techniques’. PhD thesis, International Islamic University, Islamabad, 2019 De Stefano, C., Marcelli, A., Parziale, A., et al.: ‘Reading cursive handwriting’. 2010 12th Int. Conf. on Frontiers in Handwriting Recognition, Kolkata, India, 2010, pp. 95–100 Husain, S.A., Sajjad, A., Anwar, F.: ‘Online urdu character recognition system’. Machine Vision Applications (MVA), Tokyo, Japan, 2007, pp. 98– 101 Razzak, M.I., Anwar, F., Husain, S.A., et al.: ‘HMM and fuzzy logic: a hybrid approach for online urdu script-based languages’ character recognition’, Knowl.-Based Syst., 2010, 23, (8), pp. 914–923 Panwar, S., Ahamed, M., Nain, N.: ‘Ligature segmentation approach for urdu handwritten text documents’. Proc. 2014 Int. Conf. on Information and Communication Technology for Competitive Strategies, Rajasthan, India, 2014, p. 1, Din, I.U., Siddiqi, I., Khalid, S., et al.: ‘Segmentation-free optical character recognition for printed urdu text’, EURASIP J. Image Video Process., 2017, 2017, (1), p. 62 Khan, N.H., Adnan, A., Basar, S.: ‘Urdu ligature recognition using multi-level agglomerative hierarchical clustering’, Cluster Comput., 2018, 21, (1), pp. 503–514 Naz, S., Ahmed, S.B., Ahmad, R., et al.: ‘Zoning features and 2DLSTM for urdu text-line recognition’, Proc. Comput. Sci., 2016, 96, pp. 16–22 Sattar, S.A., Haque, S., Pathan, M.K.: ‘Nastaliq optical character recognition’, Cluster Comput., 2008, pp. 329–331 Ahmad, I., Wang, X., Mao, Y.H., et al.: ‘Ligature based Urdu Nastaleeq sentence recognition using gated bidirectional long short term memory’, Cluster Comput., 2018, 21, (1), pp. 703–714 Choudhary, P., Nain, N.: ‘A four-tier annotated urdu handwritten text image dataset for multidisciplinary research on urdu script’, ACM Trans. Asian LowResource Lang. Inf. Process. (TALLIP), 2016, 15, (4), p. 26 IET Image Process. © The Institution of Engineering and Technology 2020