Study On Character Recognition Using Neural Networks Computer Science Essay

This paper briefly explains about Character acknowledgment.

It is the technique to interpret the scanned images utilizing the optical scanners into some other format that can be edited utilizing computing machine editors. Methods used for character acknowledgment and its application are besides discussed. Quite research has been done on English alphabets.

Now a twenty-four hours ‘s OCR machines can acknowledge up to 2500 character per minute


Great demand was required of character acknowledgment package since the usage of cyberspace is turning. More the usage of computing machines and cyberspace increased greater was the demand of on-line information. Digitizing all the documents work was tremendous if one has to two types so the research on character acknowledgment begun work.Character acknowledgment, it is procedure of scanning papers of typed information sometimes handwritten to, and pull outing the characters of the alphabet and therefore doing it possible for us to salvage it in digital editable format. In other words we can state that character acknowledgment is an simple procedure to scan the image from a scanner and so covert it back into the words.All preparation and proving inputs were in bitmap format ( .bmp ) because this is a really common manner to salvage images that have been scanned.

Optical character acknowledgment or merely character acknowledgment provides easy manner of informations entry

Working of OCR

There are fundamentally two techniquesMatrix matchingFeature extractionMost widely used method is matrix fiting. it is simpler and less complex so the other.In Matrix Matching method, it compares the OCR scanned image with standard library of character matrices already stored. When there is an fiting between one of these standard templets of matrices of points within a bound of similarity, the package so marks the image with ASCII character.The 2nd method that is Feature Extraction, it is Optical Character Recognition without any rigorous matching to standard prescribed templet of library. this method is besides called s Intelligent Character acknowledgment, some besides refer to it as Topological Feature Analysis, this truth of this method depends on degree of complexness used by developer. The package scans for characteristics like non closed countries, closed countries, diagonal lines, consecutive lines, line crossings and other graphical characteristics. This method is more dynamic than other method but there is more complexness in this.

Matrix fiting merely works best if there are merely few types of manners, small noise, with small or no fluctuation on characters. For handwriting matching or if characters are of many manners fluctuations feature extraction method is used.

Optical character Recognition founts

Font is a set of characters which is y 0 – 9, A to Z, and it besides includes particular characters. Font consist of a characters which have a defined characteristics that can be used in any size.

For the usage of OCR, there is a standard as defined by ANSI. OCR uses founts that can be easy recognized by the slow velocity, inexpensive systems. These are founts are easy to be read by both scanner and homo. Fonts that have greater truth are used as OCR fountsA fount is in which all the characters are efficaciously the same breadth, irrespective of the existent size of the letters, Numberss or symbols in the fount can be used as OCR fountsPre-processing was needed to turn a scanned image into OCR ready inputs. Preprocessing includes may include following procedure but are non limited to theseImage Idaho reduced to black and whiteOr in a binary signifierand so to a matrix of 1s and nothing, where 1s indicated white pels and zeros the black pels.

OCR scanners

OCR scanners are the reading devices that used for inputting the papers to computing machine these are classified into classs,Text Input andData Capture.

Text input devices are those which are used to read pages or scn paperss or big parts of paperss or even a book. The beginning is scanned with an aim to utilize it for edition intents after it is scanned. These devices have assorted degrees of mechanization from manual eating to holding automatic eating and so, reading, after that sorting, and even stacking capablenesss.The other class of devices that is Data Capture devices used to scan informations that is repeated several times and so make some pre specified arranging on the scanned vitamin E informations as it is being entered.

The information that is delivered from the OCR scanner to the package should be accurate as it is non to be used for redacting subsequently and manual work will be done on it, so it requires more accuracy so text input OCR scanners

Preparation of Datas

The first portion of the undertaking consisted on garnering sample informations and marks to develop the nervous web with. In this undertaking, the 12 platinum. Courier New fount was used to bring forth the capital letters of the alphabet, and besides an empty infinite. The character set in figure 1 was saved in.bmp format and given to the nervous web to utilize for preparation.ABCDEFGHIJKLMNOPQRSTUVWXYZ

Figure 1. Courier New Training Set

Each missive served as an input holding 108 properties. See figure 2 for a sample character from the Courier New font household holding 12*9 properties.

Figure 2. Courier New font SampleA normalized vector from 1 to 27 defined the marks for each of the 27 inputs. Therefore, the end product for the web would be a figure between 0 and 1, with 27 possible values.Next, an ideal word was created and saved in electronic image format for proving the web, merely to do certain Matlab was imitating the web right and that the web was at least working with the preparation informations.The word ‘PERCEPTRON ‘ was used for proving the web to do certain the preparation was successful. Figure 3 shows how the electronic image looked that Matlab received.PERCEPTRON

Figure 3. Ideal trial informations

Then, non-ideal informations from a scanner was used for proving the web.

This non-ideal information was typed and printed out and so scanned back in to imitate the real-world procedure of scanning in a page of text. Figure 4 contains a close-up of a piece of scanned informations.Figure 4. Non-Ideal sampleAfter having a non-ideal input such as the one in figure 4, Matlab has to change over it to a black and white image. After transition to a binary image, much information is lost and the letters besides appear noisy. The scanned informations expressions like that in figure 5.Figure 5.

Non-Ideal black and white sampleThen, Matlab converts the black and white image to a matrix of 1s and nothings. For illustration, the missive ‘Q ‘ can be spotted in the matrix after being converted:1 1 1 0 0 0 1 1 11 1 0 1 1 1 0 1 11 0 1 1 1 1 1 0 11 0 1 1 1 1 1 0 11 0 1 1 1 1 1 0 11 0 1 1 1 1 1 0 11 0 1 1 1 1 1 0 11 0 1 1 1 1 1 0 11 1 0 1 1 1 0 1 11 1 1 0 0 0 1 1 11 1 1 1 0 0 1 1 11 1 0 0 1 1 0 0 1Figure 6. Binary Matrix Representing Q

Architectures Tested

For all the architectures used, there were 27 input vectors each holding 108 properties.

Linear Associator With Pseudoinverse Rule

The first architecture that was used to try character acknowledgment was the Linear Associator utilizing the Pseudoinverse regulation. The Pseudoinverse was used alternatively of the Hebb regulation because the paradigm forms were non extraneous.

The Pseudoinverse regulation was preferred over other acquisition regulations because of its simpleness. The weight matrix for the additive associator utilizing the Pseudoinverse regulation can be found utilizing the undermentioned matrix equation:


Where P+ is the pseudoinverse defined by P+= ( PTP ) -1PTAfter organizing the input matrix P, and matching mark matrix T, the weight matrix was easy to cipher. Because of the regulation ‘s simpleness, altering the weight matrix for a new set of founts would be speedy plenty to make on-the-fly.

The Linear Associator gave better consequences than any other web tested, so this was the one chosen in the concluding version of the undertaking.

4-Layer Networks With Backpropagation Algorithm

Several different architectures were experimented with, get downing with a 4-layer web holding 12 nerve cells in the first 2 beds, 2 nerve cells in the 3rd bed, and 1 nerve cell in the 4th bed. With all the transportation maps as tangent sigmoid, the ideal informations was loaded and the web converged to a minimal mistake after about 50 eras. The web was tested with the ideal informations, and found to decently place the letters, but with the non-ideal informations, the web could non place any of the characters.The web was likely over-learning the paradigm information set, so the figure of nerve cells in each bed was changed a twosome times. Even with mean squared mistakes ( MSE ) under.01, the web could non decently place the non-ideal informations.

5-Layer Network With Backpropagation Algorithm

Of the few 5-layer webs tested, the 1 with the best consequences had 2 nerve cells in the first bed and 5 nerve cells in the 2nd, 3rd, and 4th beds, and 1 nerve cell in the fifth bed.

The tangent sigmoid map was used on the first 4 beds, and a pure additive map was used on the 5th bed.Upon preparation, the web reached an MSE of virtually zero. When tested with non-ideal informations, the public presentation was much better than with the 4-layer web, but still non every bit good as with the Linear Associator.

Consequences and Analysis

Using an ideal paradigm informations set, the best consequences for the 3 types of webs used are as follows:Figure 7. % Accuracy Using Ideal PrototypesNote: These per centums do non include the infinites in the sentences that each web easy recognized. If taken into history, these per centums would be much higher.Since the truth is evidently excessively hapless, assorted steps were taken to seek to better public presentation. These included:Using border sensing on non-ideal informationsUsing different strategies for the marksScreening the paradigm letters by similar form and sizeNone of these efforts perceptibly affected the public presentation.The chief ground the public presentation was so low was because of a character-offset consequence that occurred when Matlab reduced the scanned image to black and white.

See figure 8. The in-between image is the ideal paradigm, centered about its 9-pixel breadth, and the outer two images are what the scanned character may look like. Even though all the characters are indistinguishable, the beginning makes it barely possible for the nervous web to place it right.Figure 8. Offset consequenceFollowing, efforts to redact the paradigm forms were made because the paradigm forms should fit ( every bit good as possible ) the non-ideal information that will be gathered.For proving the effects, the Linear Associator was used because it had already been giving better consequences than the other webs tested.The first edit to the paradigm patterns involved adding noise in topographic points the scanned images looked noisy.

For the scanned images in this undertaking, most of the noise appeared at the top of the letters, so that is where the noise was added to the paradigm patterns. This method increased the truth of the Linear Associator to 12 % .Then, non-ideal paradigm forms were created utilizing the same method the non-ideal information was gathered.

This greatly improved the public presentation. The Linear Associator gave an optimal truth of 21 % .

Figure 9. % Accuracy Using Different Prototypes

Advantages of OCR

There are assorted grounds for utilizing OCR scanning method so other methods of informations entry like saloon codification. Advantages include but non limited toLesser informations entry mistake in comparing to manual entryTo fall in several Data Entry in digitized signifierTo expeditiously Handle Peak LoadsMake it Human Readable and editable signifierIt can Be easy Used with to publish once moreCan be helpful in Scaning CorrectionssIn this undertaking, assorted webs were trained to acknowledge characters of the alphabet from a scanned image. The Linear Associator preformed the best and was besides the simplest to implement.After seeking assorted methods to better the public presentation, a character acknowledgment truth of 21 % was achieved when the paradigm information was generated from the same beginning the trial information was coming from.

An truth of 21 % means that out of every 100 letters, 21 will be right identified.This truth is still really low, so other methods need to be approached for this type of character acknowledgment, such as making a more complicated border sensing algorithm, or utilizing characteristic country ratios ( for illustration, of black pels to white pels ) of the characters to place them.

Appendix – Matlab Code Explanation

An account of provided matlab files:project.m – Graphical user interface for the character acknowledgment undertakinggetsamples.m – Gets prototype information in electronic image formathebb.m – Simulation of Hebbian larning utilizing Linear Associatorreadline.m – Reads and simulates web on a line of image informations( bitmap format )projectresult.txt – File where ensuing line of text is stored