Character Recognition using Approaches of Artificial Neural Network: A Review Alankrita Aggarwala ,Shivani Gabab and Pardeep Singhc* a,b Panipat Institute of Engineering and Technology, Samalkha, Haryana 132101, India c Computer Science & Engineering, Graphic Era Hill University, Dehradun, Uttarakhand 248001,India Abstract Methods based on learning from examples that have been widely applied to character recognition are considered as one of the classification methods. This class of methods includes statistical methods, artificial neural network, support vector machines, multiple classifier combinations, etc. among these, artificial neural network are quite popular and have gained more attention. The paper presents a survey on ANNs applied for recognizing characters and the features applied as inputs. The principal objective of this paper is to help researchers attempt to apply ANN for character recognition and analysis of the networks and the features extracted to improve classification accuracy.. Keywords 1 Artificial Neural Network, Character Recognition, Neural Classifier 1. Introduction Man-made intellectual competence is the task of giving machines human-like limits. This is conceivably the most troublesome locale in computer programming over the two or three numerous years. The huge endeavors of man-made intellectual competence are offering the ability for the machine to see, interpret, and ability to examine the substance. A large variety of work has been done in the related field, at this point simultaneously; the issue stays puzzling from an all-out viewpoint [1]. Various experts have tried the issue of text affirmation by a wide scope of request moves close. Configuration organizing is maybe the most distorted methodology used conspicuously in the first place period of OCR (optical character affirmation) creative work in this various design giving the base distance in yield. This technique works reasonably with the affirmation of standard content styles anyway gives a dreary appearance composed of hand characters. Feature assessment is another direct procedure, where the quantifiable dissemination of centers is taken apart and balanced properties are isolated [2]. From each picture, a component vector is resolved and taken care of in a database. Affirmation is done by finding the distance of the component vector of the data picture with those set aside in the informational index and yielding the picture with the least deviation. Disregarding the way that this technique gives better results on composed by hand characters moreover; it is outstandingly sensitive to the upheaval and edge thickness. Further, the features removed in this technique will overall be mathematical [3].In the basic assessment, of course, an undertaking is made to remove features that can be successfully translated. These features depend on the real properties, similar to t number of joints, relative position, number of endpoints, length to width extent, etc. This system depends energetically upon the character set and the features isolated for one set are likely not going to work for the other ITTAP’2022: 2nd International Workshop on Information Technologies: Theoretical and Applied Problems, November 22–24, 2022, Ternopil, Ukraine EMAIL:alankrita.agg@gmail.com (A.1); sgsgknl@gmail.com (A. 2),pardeep.maan@gmail.com (A.3) ORCID: 0000-0002-0931-1118 (A. 1); 0000-0002-7919-0702 (A. 2); 0000-0002-0368-4757(A.3) ©️ 2020 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). CEUR Workshop Proceedings (CEUR-WS.org) character set. As tests extend, these clear arrangement procedures are sufficient to oblige the shape variability of tests as can't yield high affirmation precision. To misuse huge model data, the character affirmation neighborhood coordinated focus toward gathering strategies subject to acquiring from the model’s framework especially fake neural associations from the last piece of the 1980s till 1990s. Learning strategies have benefitted character affirmation immensely by conveying the experts from the horrifying situation of organization disclosure and tuning and the affirmation accuracy has incited considering acquiring from colossal model data. Figure 1 shows the process of classification of characters and their recognition [5,6,7]. Classification Feature Pre-Processing Segmentation and Extraction Recognition Figure 1: Classification and Recognition phases The analysis of OCRs in various lingos with their division, feature extraction, and portrayal frameworks is done in this work and focused on the portrayal subject to neural associations and study the various designs embraced [8]. Neural affiliations have been used in a wide degree of the region to deal with a wide level of issues. Not in the least like human characters that can see and hold characters in form of digits or letters; PCs treat such types as twofold plans. Subsequently, appraisals are essential to see and see each character. A neural alliance is dealing with a contraption, either evaluation or genuine stuff, whose approach was empowered by the strategy and working of animal frontal cortexes and parts thereof. The neural affiliations can obtain from models, which makes them genuinely adaptable and wavering. These affiliations are in like manner fitting for consistent plans taking into account the quick response and estimation events are a brief result of the indistinguishable course of action. Any model certificate structure routinely joins a section that portrays and isolates consistent features from a model [9]. Collect features from characters using geometry based character feature. 108 of feature value ANN Output Measure the accuracy of the recognition Figure 2: An Example figure of designing of artificial neural network Considering those features, the structure can see the given data plan. Dependent upon the given issues, a total number of features and their assembling differ according to the procedures and frameworks for depiction. In various reasonable applications, it’s exceptional to face issues including different features. One can feel that each part is huge for likely a dash of partition. In any case, it has been found over the long haul beyond a particular point, the breaker of further features prompts sadder rather than better execution and fabricates the managing time [10]. As such, the affirmation of features, i.e., keeping suitable features and frustrating insignificant or likely abundance ones, is a fundamental improvement in a model statement structure plan. Character affirmation can be segregated into two giant plans: typewritten and really made. In Typewritten declaration sees a report that has been really made and checked to go before affirmation progress. Such a plan would be used as a way to deal with oversees digitizing books, reports, and papers in libraries, government, or held by affiliations. In translated attestation, the development attempts to see a book that has been outlined by a human [11]. Advantages of character accreditation are: reviewing postal area off envelopes, researching customer-filled turns of events, recording and recuperating substance, digitizing libraries, etc Using OCR, the deciphered and machine-made substance could be managed into PCs to make edifying varieties of existing works without using the control local area. The various techniques for character confirmation have sufficiently been appropriated at any rate the system, blend of phony neural affiliations and inherited computation changes into the critical advantage of the strategy over other existing methodologies. It is advanced than those techniques. Figure 2 shows that using any soft computing methods a network can be built to recognize a character or a group of characters [12]. 2. Properties of Neural Classifier Neural classifiers show different properties in the going with respect. Training complexity: The limits of neural classifiers are generally changed by point dive. By dealing with the readiness tests a fixed number of degrees the planning time taken is forthright with the number of tests. The flexibility of training: The limits of neural classifiers can be changed in string level of configuration level getting ready by point plunge resolved to smooth out overall execution. For the present circumstance, the neural classifier is embedded in the string or plan affirmation for character affirmation [13]. 2.1 Modal Selection: The hypothesis execution of neural classifiers is sensitive to the size of the plan, and the assurance of fitting development rules on cross-endorsement. 2.1.1 Storage & execution complexity: - Neural classifiers have numerous fewer limits and the number of limits isn't hard to control. Thusly, neural association classifiers consume less limit and estimation [14]. 2. Literature Review A Lot of work has been done in this field with the help of phony neural affiliations. ANN joins setting up all characters. ANN can find the most conceivable character by theory [15]. Different techniques for character demand have been analyzed ward on four general philosophies of model affirmation, as proposed [16] configuration organizing, quantifiable methodology, essential structures, and neural affiliations. Secret Markov Model is a completed quantifiable model that endeavors to expect the faint game- plan. Thus, it in like route attempts to see the faint character which is given as data [17]. In case the differentiation between dull data and getting ready data is gigantic, the system may not display well. Likewise, the HMM model doesn't get the connection between letters [18]. A system to make a flexible character affirmation structure using a neural alliance. Back- Propagation Neural Network with one mystery layer is used to make the structure. The development is ready and concentrated with printed and framed by hand English letter sets and showed up in his test outcomes that printed text gives the best exactness in attestation over-deciphered characters [19]. The back-propagation appraisal changes the schematic of the agreement by using a sigmoidal limit. The advantage of quite far is that beyond what many would consider possible is differentiable works decently on principal masterminding issues. Regardless, as the badly arranged whimsies grow. The introduction of back-propagation tumbles off rapidly under the way that magnificent spaces have essentially by and large minima which are little among the close by minima. Evaluation search procedures will, overall, get captured at neighborhood minima [21]. Furthermore, BPN encounters a scaling issue. Neural relationship with Back Propagation learning showed results through searching for various kinds of cutoff focuses. In any case, the choice of beyond what many would consider possible routinely as of now picks the achievement of the planning correspondence [22]. The decision of these cutoff focuses proceeds in the sensible use of general principles, yet their value is everything considered shocking. Since the chief undertakings to solidify GA and NN started in the last piece of the 1980s, various researchers have joined the new turn of events and made a flood of journal articles, explicit reports, etc [23]. An expansive arrangement of issues has been explored by different GANN approaches, for instance, face validation, animates [24], the portrayal of the normality of the thyroid organ, covering recipe assumption [25], and some more. In like manner, a wide degree of encoding methods has been executed. The various procedures made for character certificates. Comparative observation of ANN and the one proposed technique for breast cancer recognition and detection is done [26]. Along with the design, an edge detection Algorithm by Artificial Neural Network (ANN) for cancer detection [27] and providing security features of cryptography recognition can be done[28]. Different approaches of risk management in recognition some frameworks are defined [29] and by using the machine learning algorithms and random forest approach the risk in recognition of errors can be minimized. [30]. 4. Specific Studies on various OCRs using Neural Networks Iyswarya, M. R et al., (2012)[2] proposed Tamil character affirmation using an MNN classifier using a discretized sigmoid limit. The database for setting up the MNN contains 64*64 pictures of twofold data for the game plan of 18 Tamil fundamental characters. The Tamil characters are formed using the data contraption, the mouse. After smoothing, the features are taken out from the equal data. These features are the subsequent invariant features, which are non-direct limits invariant to translation, scaling, heading and are described on the numerical previews of the image. Therefore, a total of 108 sets, is used as the features for setting up the neural association. The MNN has 7 information units, with 2 mystery layers. The association is worked with a discrete sigmoid limit. The proposed technique gave an arrangement period of 433seconds with a typical %age botch of 0.68. Since BPNN encounters issues of the slow planning measure, picking suitable characteristics for the limits and moving away from close by minima is a fundamental endeavor. Thusly the maker further proposes an optical back-propagation computation. Two three-layered associations are completed, each with a different design. The assessments were made in seeing characters from A-J. Twofold photos of the characters are tended to with combined characteristics (- 1,1). Dull pixels are addressed by 1's and given to the data layer.2 sorts of pictures are used in the readiness set; an 8*6 twofold picture for the principle NN, while 12*8 for the second NN. The inception work used is the sigmoid limit; twofold vectors of the size 4 location the yield regard and is tended to by 0000, B by 0001, while J by1001. The size of the mysterious layer for the chief organization is 8producing 48X 8X 4 associations, while the ensuing association has 16 in its mysterious layer, consequently 96*16*4 designing. Little characteristics for the learning rate were used to avoid close by minima, the value comes from from0.1 to 0.4. The planning measures were finished when the mean square goof isn't actually or identical to 0.0001. The OBP hopes to speed up the arrangement cycle and break from neighborhood minima. Werbos, P. J. (1994) [22] uses a 3 NN plan, Backpropagation, transient Flow, and the spooky common dull NN. BPN has been for the most part used in plan affirmation issues. In the BPN Architecture, the models are learned as a limit between the data sources and the directed yields. The association has 1 information and 2 mystery layers getting subsets of features. The heaps and cutoff points for the neurons are affected together by a bordering social affair of data incorporates instead of a singular worth, therefore getting the interrelationships of the information vector in a restricted way. 2 redundant plans are used. In the transient stream association, the concealed layer has a copy back layer, which is used to remember and reuse the previous sign-on time to learn common characteristics of a model. In the Spectro transient association plans, despite the copy back layer to get comfortable with a momentary progression, input center points are grouped in little social occasions to gather the spooky scattering of the component vectors too. The information layer has interesting affiliations, which match the supernatural scattering of a component of the character. The shadow feature is suitable to address the spooky and common nature of isolated physically composed character data. The path toward removing this segment incorporates sifting the character in a left to right and beginning to end way. The shadow feature tended to a character as a grouping of its parts. Each segment of this vector tends to the distance between the restrictions of the character's shape. It has 4 specific parts containing 15 numbers each. The first is the game plan of 1 number tending to the distances between the cutoff points in a left to right way while the second is the course of action of those in a completely the manner. Stripping the left and right constraints of the character commercial getting the 15 numbers to get the third piece of the vector. These numbers address the distance between the restrictions of the character and a while later the 15 numbers tending to the distance between the constraints of inside structure in a top and way are resolved. Dutta, A., et al., (2021) [6] proposed two-part extraction techniques. Similarly, two neural classifiers were used for experimentation. The essential technique called course feature arrangements the line segment in each character into 4 sorts’s a vertical line, level line, right cockeyed, and left-slanting and delegate a number to each heading. The second component extraction technique relies upon the assessment and space of progress features from the establishment to cutting edge pixels in the vertical and level manner. The neural classifier picked for character affirmation was backpropagation and extended reason work associations. Two separate neural associations were ready for lower case and uppercase letter sets. Results showed that the prompt component defeated the advancement feature. Affirmation rates above 80% were represented. Lincy, R. B., & Gayathri, R. (2021) [9] examined the current ring-based, the new region-based, and the mix of these i.e., the blending procedure for the affirmation of interpreted English capital letters. Essential features like endpoints, convergence centers, and the number of branches are used for the pre-gathering of characters, the local features, for instance, normalized vector length and focuses got from either ring or region approaches are used in the planning using the reference characters of the test Character. Taking into account the pre-gathering, input centers and yield center points are picked for the backpropagation association. The overall affirmation rate is generally 85%. Singh, S., & Garg, N. K. (2021) [3] discussed the recognition of Devanagari characters in a noisy environment. After preprocessing, the segmentation process generates the basic units in the Devanagari script. For each unit, the co-ordinate of the bounding rectangle of the unit and its type (core character, upper or lower modifier) are stored. Further structural features are extracted from the segmented images. The features include the line in various orientations and the circle of various radii similar to those in the characters. Hencea 23-bit binary string is generated for each character unit, which consists of sufficient information to classify the character. The network accepted 23 inputs, corresponding to the 23 structured features in the fearer vector, 11 hidden layers, and 3 output neurons where each output corresponds to a character in the core character subset of the Devanagari subscript. The system gave a recognition rate of 76%. Another approach for seeing the Devanagari character is discussed by R. Bajaj et al. (2002) [12]. The characters are separated from words. Then, the shirorekha (level line top of the word) is made plans to use projection profile and run length. Plans above shirorekha are detached as ascenders while those under the character stature are the descenders. The point incorporates further gathering the partitioned pictures into 74 classes: 4 ascenders, 2 descenders, and 68 focus parts. Focus parts are portrayed into 4 social affairs subject to the presence of a vertical bar no vertical bar, and vertical bar at the center, right, or at various regions. Four neural associations are used for portrayal inside these social events. Neural association classifier contrasting with focus fragments has an exactness of 84.77%. The neural association classifier used was a standard backpropagation network with coordination activation work. The breaking point sureness for each class is directed by looking at the dull executive characteristics. Lately, a structure for translated numeral affirmation of Devanagari characters. The numerals have been tended to using two kinds of features. The essential kind gives a coarse shape plan of the numeral and is respectably pitiless toward minor changes in character shapes. The less than the ideal of features endeavors to give emotional portrayals of the character those that are depended upon to be invariant across making styles and text styles. Multi- layer wisdom is used for the request for the numerals. The Neural Network approach is similarly used for the upgrade of Bangla characters. Dutta et al.(2014) [24] uncovered work on the affirmation of isolated Bangla alphanumeric physically composed characters using neural associations. The characters have been tended to the extent the locals and basic objectives introduced by the convergence present in the characters. A two-stage feed-forward neural association, arranged by a back-propagation computation has been used for affirmation. The fundamental prerequisites constrained by the convergences have been encoded in the topography of the real association. Using Kohonen's neural association, Montana, D. J. et al.(1989) [23] proposed a structure to see the letters arranged by the Oriya script. The information pixels are dealt with to the neurons in the Kohonen layer where the neurons choose the yield according to a weighted condition. The character is portrayed by the greatest yield Obtain from the neurons. Here the maker assessed five Oriya characters and consequently the immovable nature of the system isn't set up. Sukhaswami MB et al.(1995) [15] tended to a system that affirms for printed Telugu characters by neural association method no work is yet given a record of composed by hand Telugu characters. From the start, use of Hop field, a neural association model, for affirmation purposes. As a result of limitations in the limit furthest reaches of the Hop field neural association, they later purposed another arrangement called diverse neural association subsidiary memory (MNNAM). These associations can be arranged similarly as they work on ordinarily disjoint courses of action of planning plans. Frameworks organization of these more unobtrusive associations into MNNAM overhauls as far as possible. 5. Conclusion Here is the fundamental finish of this exploration paper that creative methodologies of fake neural organization in character acknowledgment having a significant job and to take care of the issue of character acknowledgment and this paper some imaginative methodologies applied by the various analysts for transcribed character acknowledgment utilizing counterfeit neural organizations. Even though multi-facet feed-forward neural organization design is all the more normally utilized, numerous specialists have attempted to investigate another engineering to accomplish higher acknowledgment correctness. A ton of issues is should have been contemplated concerning the acknowledgment exactness. The division, highlight extraction, and proper choice of reasonable neural organization design oversee the arrangement precision. The highlights extricated ought to be revolution and size invariant and sufficiently able to arrange the character proficiently for the given order conspire. More accentuation is to be given to complex issues identified with over-divided and covering characters to deal with fluctuation associated with the composing style of various people; strong engineering of neural organization ought to be thought of. Further acknowledgment of Indian written by hand script acknowledgment needs to confront additional difficult errands because vowel modifiers are present leads to improvement in transformation and composite characters. 6. Future Scope Here the techniques for seeing interpreted character font are presented. The projected strategy can be applied to various dull characters. Artificial Neural Network assists the framework with seeing the character whether the specific model can be opened in the data set. The neural affiliation-based strategy gives an exactness of 85 %. made for proposed assessments can't be applied to see cursive penmanship Recognition. Later on, we can endeavor essentially indistinguishable tests over different characters and for certain additional or fresh cutoff points to improve the precision. 7. References [1] Liu, C. L., & Fujisawa, H. (2005, August). Classification and learning for character recognition: comparison of methods and remaining problems. In Int. Workshop on Neural Networks and Learning in Document Analysis and Recognition. [2] Iyswarya, M. R., Deepak, S., Jagathratchagan, P., & Kailash, J.(2021) Handwritten Tamil Character Recognition Using Convolution Neural Network by Adam Optimizer. International Journal of Advanced Research in Science, Communication and Technology (IJARSCT) Volume 6, Issue 1,pp:40-45 [3] Singh, S., & Garg, N. K. (2021). Review of Optical Devanagari Character Recognition Techniques. In Intelligent System Design (pp. 97-106). Springer, Singapore. [4] Inunganbi, S., Choudhary, P., & Manglem, K. (2021). Meitei Mayek handwritten dataset: compilation, segmentation, and character recognition. The Visual Computer, 37(2), 291-305. [5] Pradhan, S. K., Thirupathi, J., Rao, N. K., Aluru, J., & Suraparaju, K. MANUAL CHARACTER RECOGNITION WITH OCR. Turkish Journal of Physiotherapy and Rehabilitation, 32, 3. [6] Dutta, A., Garai, A., Biswas, S., & Das, A. K. (2021). Segmentation of text lines using multi- scale CNN from warped printed and handwritten document images. International Journal on Document Analysis and Recognition (IJDAR), 1-15. [7] Singh, H., Sharma, R. K., Singh, V. P., & Kumar, M. (2021). Recognition of online handwritten Gurmukhi characters using recurrent neural network classifier. Soft Computing, 25(8), 6329-6338. [8] Simayi, W., Ibrayim, M., & Hamdulla, A. (2021). Character type based online handwritten Uyghur word recognition using recurrent neural network. Wireless Networks, 1-11. [9] Lincy, R. B., & Gayathri, R. (2021). Optimally configured convolutional neural network for Tamil Handwritten Character Recognition by improved lion optimization model. Multimedia Tools and Applications, 80(4), 5917-5943. [10] Pradhan, S. K., Thirupathi, J., Rao, N. K., Aluru, J., & Suraparaju, K. MANUAL CHARACTER RECOGNITION WITH OCR. Turkish Journal of Physiotherapy and Rehabilitation, 32, 3. [11] Santosh, K. C., & Gawali, B. W. (Eds.). (2021). Recent Trends in Image Processing and Pattern Recognition: Third International Conference, RTIP2R 2020, Aurangabad, India, January 3-4, 2020, Revised Selected Papers, Part I (Vol. 1380). Springer Nature. [12] R. Bajaj, L Dey, S. Chaudhary Devanagari numeral Recognition by Combining Decision of Multiple Connectionist Classifier, Sadhana 2002. [13] SENTHILSINGH, C., SUJIHELEN, L., SASIKUMAR, B., & LIVINGSTON, L. M. HANDWRITTEN TEXT CLASSIFICATION USING DEEP LEARNING. Turkish Journal of Physiotherapy and Rehabilitation, 32, 2. [14] Das, D., Nayak, D. R., Dash, R., & Majhi, B. (2020). A Multi-Stage Hybrid Model for Odia Compound Character Recognition. In Applied Intelligent Decision Making in Machine Learning (pp. 53-70). CRC Press. [15] Sukhaswami MB, Seetharamulu P, Pujari AK. Recognition of Telugu characters using neural networks. Int J Neural Syst. 1995 Sep;6(3):317-57. doi: 10.1142/s0129065795000238. PMID: 8589867. [16] Jieun Kim, Ho-sub Yoon, Graph Matching Method for Character Recognition in Natural Scene Images, INES 2011, pp 347-350, 978-1-4244-8956-5/11/$26.00 ©2011 IEEE [17] Saha, S., & Som, T. (2010). Handwritten character recognition by using Neural-network and Euclidean distance metric. IJCSIC-International Journal of Computer Science and Intelligent Computing, 2(1). [18] Singh, R., Yadav, C. S., Verma, P., & Yadav, V. (2010). Optical character recognition (OCR) for printed devnagari script using artificial neural network. International Journal of Computer Science & Communication, 1(1), 91-95. [19] Hewavitharana, S., Fernando, H. C., & Kodikara, N. D. (2002, December). Off-Line Sinhala Handwriting Recognition Using Hidden Markov Models. In ICVGIP. [20] Krevat, E., & Cuzzillo, E. (2006). Improving off-line handwritten character recognition with hidden markov models. Transaction on Pattern Analysis and Machine Learning, 33. [21] Faaborg, A. J. (2002). Using neural networks to create an adaptive character recognition system. Cornell University, Ithaca, NY, 23. [22] Werbos, P. J. (1994). The roots of backpropagation: from ordered derivatives to neural networks and political forecasting (Vol. 1). John Wiley & Sons. [23] Montana, D. J., & Davis, L. (1989, August). Training feedforward neural networks using genetic algorithms. In IJCAI (Vol. 89, pp. 762-767). [24] Dutta, D., Chowdhury, A. R., Bhattacharya, U., & Parui, S. K. (2014, September). Stroke level user-adaptation for stroke order free online handwriting recognition. In 2014 14th International Conference on Frontiers in Handwriting Recognition (pp. 250-255). IEEE. [25] Chatha, D., Aggarwal, A., & Kumar, R. (2020). Comparative Analysis of Proposed Artificial Neural Network (ANN) Algorithm With Other Techniques. International Journal of Security and Privacy in Pervasive Computing (IJSPPC), 12(1), 29-35. [26] Aggarwal, A., & Chatha, D. (2019). To Design a Mammogram Edge Detection Algorithm Using an Artificial Neural Network (ANN). International Journal of Distributed Artificial Intelligence (IJDAI), 11(1), 34-43. [27] Nagpal, S. (2016). Quantum Cryptography Integrated Effective Communication Approach for WPAN. International Journal of Enhanced Research in Management & Computer Applications, 5(9), 1-5. [28] Aggarwal, A., Dhindsa, K. S., & Suri, P. K. (2020). A Pragmatic Assessment of Approaches and Paradigms in Software Risk Management Frameworks. International Journal of Natural Computing Research (IJNCR), 9(1), 13-26.