Report on Neural Networks for Handwriting Recognition

Like it has been discussed earlier, signature and text both of them are considered to be the recognition system and they can be in the form of online or even the offline as well. This division is being based upon that how a system receives a particular data. The online system is the one that receives the data directly from any kind of the pen device that is being attached to the computer. While on the other side, for the offline system handwriting gets presented already on the paper. For reading out such kind of the handwriting, it will definitely require some sort of the reading devices like a scanner that can help to transform a picture in the digital format. Both of these online as well as the offline systems have the significance interest to the different researchers, but at times online system is the one that gets preferred upon the offline system. Difference between both of them has been shown in the table below. (Bluche T. K., 2015)

	Online	Offline
Applications	Less Applications	Wider Applications
Data	Pen trajectory	Pixel Data
Input Devices	Writing Device	Scanner

 

According to one of the surveys, it has been found out that offline signature verification has always been known to be the difficult approach and it also gives the worse results when it gets compared with the online one. In fact this has also been found out that this approach has also gain a great deal of the interest towards the scientific community. It also gives the high financial impact and particularly on the automated verification in some checks signatures and even some of the signatures on the official documents. Verification for the online signature is way less difficult from the time period of a dynamic information of the signature that is available and also gets captured much easily as well. However too much of the information also caused the different constraints in the selection of best features for the representation of a signature. (Graves A. L., 2008)

Historical Survey of Neural Networks for Handwriting Recognition.

Historical survey for this Neural Networking of handwriting recognition is being followed by the different individuals that have been inserting different kind of the information in the different computers by the help of a keyboard. Here the keyboard is being used as a medium for the inserting of an information through an outsource. This process is not known to be much efficient or useful, inclination related to the automated insertion of the data function has been incremented. Currently this is being known as the “automatic identification” that is the all-encompassing term related to the huge variety of the different procedures that are present. In these procedures, it includes the radio frequency, vision platforms, reading of the different bar codes, magnetic stripe and etc. Neural Network is the one that is also a part of it and gets included in this group. This gets useful to be used along with convenience as well when data is present that is also comprehendible through the people along with the different machines that are being needed and at this stage any other input can’t be presumed. Handwritten recognition or the identification system is being distinguished into four different kind of the stages that are, preprocessing, feature extraction, classification and the last one is post processing. The data has been acquired through the tablet and it constitutes of the different functions that are being organized over occurrence, force, velocity and time. Furthermore timing data along the different structure of the strokes elements leads towards the significant use of the online character identification that is being compared to the recognition of the offline physical writing identification. Velocity along with the force data both of them are convenient for the analysis of personality along with the identification of the separate individuals but in short they don’t come in handy for the character identification that gives a unique force along the writing speed. Strokes are also being evaluated through following course of the movement. These are the strokes that have different breaks that can’t be completely processed by the input of a human. So because of this, preprocessing along with the pattern extraction are being carried out against all these raw strokes. (Doetsch, 2014)

Preprocessing of Neural Networks for Handwriting Recognition.:

Preprocessing is one of the primary step for the character identification or recognition and this phase used to be the critical one for the great recognition rate. One of the main aim or motive behind this preprocessing in any of the system design is to normalize the whole input along with removing different variations as well like noise and etc because in the presence of all these variations rate of recognition gets reduced. There are different kind of the techniques that are being used for the preprocessing are normalization, skew detection, filtering, removal of noise and the correction of slant. If reduction of the noise is not being done at the time of this preprocessing that would obviously lead towards the poor segmentation and at the end recognition rate would be too low. In addition to this, there are different steps of preprocessing which takes place at the first stage just to make sure that handwriting strokes or the characters get back to normalization. Then another phase also takes place by the name of Character Fetching phase, the handwriting is being stored in the form of a complete array of an axel in the form of bitmap image. (Ciresan, 2011)

Feature Extraction of Neural Networks for Handwriting Recognition.:

One of the most important and the primary goal of this phase is to extricate this pattern that is way more suitable for the categorization. All these features can be of different types or categories that includes horizontal features, texture based features, vertical features and so on. Identification for the every segment completely depends on the selection that is being done and it includes length, feature angle, relative positioning and connection of the different angels. Another method was also being used to find out the feature is the extraction of simple cells and then grow these cells one the base of connected component concept. There are some additional useful features that includes the directional feature, writing direction, size, shape, chart and the ending coordinates. This technique of feature extraction algorithm is quite evident from its designation. This all includes the identification of different kind of the characters or even the symbols as well at times by depending upon the features or the aspects that seems to be similar or alike. This is the concept that helps in resembling the humans that how they identify the different characters on the basis or terms of different aspects or the features. Here the developer have to inform about the code related to the specific useful features that are being needed to get recognized through different kind of the procedures and among which one of the procedure is the manual one. Here some of the features that could constitute for the different characters are number of strokes, ratio of the different pixels from the right of vertical half point and the distance from the image center. These kind of the various algorithms are being utilized in the character identification of the algorithm that also includes the recognition of the physical writing. The literature review related to all of these present studies on the character identification, it was completely obvious that there had been a little effort for the improvement of different features. This recognition process could be directly improved by the help of feature enhancement because this helps in improving the misperception regarding different characters that are being identically shaped. This is one of the approach that might provide the programmer to additionally control over the different features that are being used for the identification or recognition. This is the technique that takes longer to be conducted but this helps in giving the exact and accurate results.

The State of the Art of Neural Networks for Handwriting Recognition.

This state of the art of the automatic recognition of a handwriting at the start of the new millennium is that as being a field, this topic is no longer an esoteric topic upon the fringes of an information technology but it is one of the mature discipline that has found number of different commercial uses as well. On-line systems for the recognition of handwriting are easily accessible and available in the computers that are hand held like the personal digital assistants. The performance for all of them is accessible as well as acceptable for the processing of hand printed symbols and when they get to be combined with the entry of a keyboard, a very strong and powerful method for the entry of data has been created. (Rehman, 2014)

Offline systems are the ones that are known to be less accurate as compared to the online systems. There is no doubt in it that they are good enough and they have a useful and a significant effect on some of the specialized domains like the interpretation of the handwritten postal addresses that is being written on the different envelopes along with the reading courtesy amounts on certain different bank checks.

Success for this online system makes it way more interesting as well as attractive for the considering of the developmental different kind of the offline systems that are being used for the first estimate of the trajectory of the writing form for the off-line data and then using the online recognition algorithms. But here the difficulty or problem for the recreating of the temporal data has led this to some of the different extraction systems till now. (Graves A. M., 2013)

There has been done a research on this automated written language of the different recognition dates back to the several decades. But now in these days, all of the cleaned machine printed text documents that have some of the simple layouts can easily be recognized or known reliably through the OCR system of the software. There has also been some kind of the success in the handwriting recognition and especially for the isolated hand printed characters along with the words as well. An example here can be used that is, in the online case all of the recently introduced personal digital assistants have a practical value. In the same manner, some of the online signature verifying systems have also been marketed over some of the last few years along with the instructional tools as well to help out the different children for writing and helping them to teach how to write are considered to be the beginning to emerge. There are many of the offline successes that have come into the constrained domains that includes the postal address, different bank checks along with the census forms as well. Analyzation of all these documents by the complex or complicated layouts, recognition of the degraded printed text along with the recognition of the running handwriting has been continued to remain large enough in the arena of a research. There are also some of the major research challenges in the online as well as offline processing of the handwriting are in the word along with the separation of line, segmentation of different words in characters, recognition of such words when lexicons are large enough and the last thing is that use of all these language models in the processing for aiding along with the recognition. There are several different applications, the performance of the machine are much far from being acceptable although there are some of the potential users which often forget that human subjects are the ones that generally make some of the mistakes related to the reading. (Wang, 2012)

Designing for the human-computer interfaces is completely based upon the handwriting that is the part of some tremendous research effort along with the recognition of the speech as well, processing of the language along with the translation for the facilitation of different levels of communication of people from different groups with the computers. From this point of view, any kind of a success or failure in this field will definitely have a great and major impact on the evolution of different languages.

Case study of Neural Networks for Handwriting Recognition.

For a case study, an image has been selected that contains a text “THE QUICK BROWN FOX JUMPS OVER THE LAZY DOG” has been selected. Font that was being used for this writing was Calibri in style and the text size was 20 having space of 2.2. For the matter of simplicity along with the better understanding that what has been written, all of the letters were capitalized. Here the training set is the one that consisted of same font along with the same size as well from the alphabets A to Z having a full stop at the end of it. As it has also been discussed earlier that custom training is the one that helps in making all of this way more accurate. In the below figure it helps in explaining that how all of the alphabets are being included in it. The word DOG in the sentence has been moved purposely to the next line just to check out the line detection in a program. Different stimulation studies are also being carried out,

                               THE QUICK BROWN FOX JUMPS OVER A LAZY DOG.

                                    Figure shows the printed characters for the case study.

Analysis of the case study:

Training set, there is a specific idea behind this targeted OCR that gets the tailored training set. This training set contains all of the different kind of alphabets that include 26 alphabets at the end of a punctuation. As it has been seen that only capital letters are being used, this training set don’t need any of the small letters at all.

                            ABCDEFGHIJKLMNOPQRSTUVWXYZ

                             This shows the training set for the character recognition.

Training of Neural Networks for Handwriting Recognition.

There are different number of the nodes that can now get fixed as soon as the training set gets ready. In this set, 35 different kind of the input nodes are present that further represents each pixel for the character of 7*5. There are almost 27 output nodes that means the hidden nodes and are half of 27 in number. All of the hidden nodes were being fixed at 11. For the training of Neural Network model, there are different parameters that need to remain fixed. There are different kind of the optimal values that are also found out by the stimulation analysis. Some of the important parameters are the Learning rate, Momentum constant. As soon as it gets trained, the model can further be used for the recognizing of a number for the different images having same set of the building blocks.

Optimum value of Neural Networks for Handwriting Recognition.

This is the parameter for the model which decides that how fast and accurate the result will be. It is very important to do and carry out some of the analysis before fixing up the different values. This optimum value can be found out through looking at the performance that is being attained as well as time taken.

Analysis of the learning rate of Neural Networks for Handwriting Recognition.

For the analyzation of a learning rate, momentum constant along with the maximum epochs had been kept constant specifically.

           Analysis of the Learning rate

LR	0.01	0.1	0.25	0.5	0.75	1
Epoch	6000	3681	4137	5264	2846	3740
Goal Met	No	Yes	Yes	Yes	Yes	Yes
Time (s)	29	26	23	27	16	21

 

From the table it is clear that LR at the rate of 0.75 that has produced the best results in the terms of time along with the performance as well. There has been an average time taken is very low as compared to the rest part. Average performance was 0.444 and it has been clearly one of the most suitable choice. In the further increase of a learning rate also means in one way or the other that Neural Network is the model that would deal with all of the minor changes as a different input.

Analysis for the momentum cost of Neural Networks for Handwriting Recognition.

After the optimization of learning rate, time of the training completely depends upon the momentum constant. Same like the learning rate, it helps in determination that how fast it gets converged. This Momentum Cost MC reduces the different chances of giving out the wrong inputs in the process of learning when any kind of the unusual inputs are being processed. In a same way this MC eases the function for the Learning Rate.

Result of Neural Networks for Handwriting Recognition.:

Training was being done or carried out and it was successful as well. In addition to meeting of the performance goal, all this process just took 21 seconds to finish out the training. Problem was then presented further to the trained model for the recognization of different characters and the result came out with 100 percent accuracy without any of the fault or problem. All of the characters were being decoded correctly. Detection of the spacing along with the line detection was completely accurate. For the recognizing of characters, training set was being created from the set of same handwriting. When all of the steps were being followed exactly, result obtained was 90 percent accurate.

Future prospects:
Integration of the different fuzzy logic into the Neural Networks

This Fuzzy logic is a kind of logic that helps in recognizing more than something true and false values and therefore better stimulation of the real world. This fuzzy logic along with the neural networks have been integrated for the uses as a diverse along with the automotive engineering, different applicants screening for the jobs, control for the crane and monitoring as well.

Pulsed Networks of Neural Networks for Handwriting Recognition.:

Most of the practical applications of the artificial neural networks are completely based upon the computational modeling that involves the propagation of continuous variables through one of the processing unit to the next one. In past years, different kind of the data from neurobiologists experiments have made it much clear that such biological neural networks communicate through the different pulses and there is a timing for the pulses for the transmitting of an information along with the performance of computation. This is the realization that has stimulated significant and impactful research on the pulsed neural networks that includes the theoretical analysis along with the model development, modeling of neurobiology along with the hardware implementation. (Fujisawa, 2008)

Hardware Specialized for the Neural Networks:

There are some of the networks that have been hardcoded into the chips or even the analog devices, this is the technology that will be much more useful as being the networks that are being used gets more complexed or complicated. Another primary advantage of the direct encoding neural networks onto the different chips or the specialized analog devices is the SPEED. NN hardware is the one that currently runs in some of the areas like especially those areas where high performance is being required along with the hardwired networks as well. When different NN algorithms get developed to the point where number of different useful things can be done, high performance for the NN hardware will ultimately become really important and essential for the practical operation. (Nahar, 2012)

Improvement of certain existing technologies:

All of the current Neural Network technologies will definitely get improved. Everything starting from the handwriting along with the speech recognition and it will even become more sophisticated as all of the researchers develop better training methods along with the network architectures.

In future this might allow,

·         Different robots that may see, feel and predict the whole world around them.

·         Helps in improving the stock prediction.

·         Common usage of all the self-driving cars.

·         Composition of the music.

·         Some of the handwritten documents that gets transformed automatically into the word processing formatted documents.

·         Self-diagnosis of the different medical problems by the use of neural networks.

Conclusion of Neural Networks for Handwriting Recognition.:

Handwriting is one of the special art of human being that has influenced the innovation of today’s technology. There are many different kind of the systems out of which pen based system like the PDA that means personal digital assistance have also been developed. There are different users that also enter the input through writing on the screen and the system will capture and then recognize the characters further. Recognizing the single character would be quite easy and simple but while dealing with the cursive or the mixed cursive word, the recognition of the different characters can be tough and challenging task.

There have been done many approaches for the recognition of the handwriting and among which one of them is the Neural Network NN. This NN is being motivated by the capability of the human learning that is in other words is the brain function is one of the promising as well as the potential approach in the recognizing of the handwritten characters and the different words. Different studies have been done and it has also shown that Neural Networking is the one that can perform very well and it also yields much encouraging results as well. In some problems, NN is the one that could even get combined with different other approaches like the statistics or NN models for the production of better results. In a complete nutshell, there are more researchers on the Neural Networking particularly in the handwriting recognition and this domain needs to be continued as this is the promising as well as the potential technology to be explored.

References of Neural Networks for Handwriting Recognition.:

Bluche, T. K. (2015). Where to apply dropout in recurrent neural networks for handwriting recognition?. In 2015 13th International Conference on Document Analysis and Recognition. 681-685.

Bluche, T. N. (2015). Framewise and CTC training of neural networks for handwriting recognition. In 2015 13th international conference on document analysis and recognition. 81-85.

Ciresan, D. C. (2011). Convolutional neural network committees for handwritten character classification. In 2011 International Conference on Document Analysis and Recognition. 1135-1139.

Doetsch, P. K. (2014). Fast and robust training of recurrent neural networks for offline handwriting recognition. In 2014 14th International Conference on Frontiers in Handwriting Recognition . 279-284.

Fujisawa, H. (2008). Forty years of research in character and document recognition—an industrial perspective. Pattern Recognition,. 2435-2446.

Graves, A. &. (2009). Offline handwriting recognition with multidimensional recurrent neural networks. In Advances in neural information processing systems. 545-552.

Graves, A. L. (2008). Unconstrained on-line handwriting recognition with recurrent neural networks. In Advances in neural information processing systems . 577-584.

Graves, A. M. (2013). Speech recognition with deep recurrent neural networks. In 2013 IEEE international conference on acoustics, speech and signal processing . 6645-6649.

Liwicki, M. G. (2012). Neural networks for handwriting recognition. In Computational intelligence paradigms in advanced pattern classification. 5-24.

Nahar, K. (2012). Artificial neural network. COMPUSOFT, An international journal of advanced computer technology. 25-27.

Pham, V. B. (2014). Dropout improves recurrent neural networks for handwriting recognition. In 2014 14th International Conference on Frontiers in Handwriting Recognition. 285-290.

Rehman, A. &. (2014). Neural networks for document image preprocessing: state of the art. Artificial Intelligence Review, . 253-273.

Wang, T. W. (2012). End-to-end text recognition with convolutional neural networks. In Proceedings of the 21st International Conference on Pattern Recognition. 3304-3308.