Journal of Engineering Sciences Vol 14 Issue 02,2023

HandWritten Text Recognition Using Machine

Learning
K.Hemalatha, C. Lakshmi , N. Chinna Puli Raju , S . Ismail Zabiulla, D.Anil Kumar

Department of Computer Science and Engineering, Annamacharya Institute of Technology and Sciences(Autonomous), Rajampet,
Andhra Pradesh ,India-516126.

S. Ashok Kumar

Assistance Professor in Department of AI&DS , Annamacharya Institute of Technology and Sciences(Autonomous), Rajampet,
Andhra Pradesh ,India-516126.

Abstract: The study of handwritten character recognition has gained popularity. The handwritten characters that were scanned as
input in the proposed technique were identified using a variety of machine learning algorithms.. It segments each character in the
image and recognises the letters after receiving the handwritten document as input in the form of a high resolution image.
Additionally, it recognises the letters before going on to find the words in the image. Based on the training it received from the
training data, this is accomplished with the help of machine learning algorithms. The specified input image will be provided in word
document format as the intended output. Large data sets of images that display the various writing styles and shapes can be used to
train the system. When training the system with vast amounts of data, machine learning is crucial. This can also be used to businesses
and organisations that only keep critical records in writing form. Keywords: OCR algorithm, simple OCR, digit recognition, digit
OCR.

I.Introduction

The process of turning a handwritten text image into a text file that a computer can read and utilize for a variety of applications is
known as handwritten recognition. The goal of this method is to create software that can comprehend handwritten documents. The
technique extracts contour-based information to identify the character or word. One flaw in handwritten papers is that they are
challenging to read through. In recent years, one of the difficult study areas in the realm of image processing and pattern recognition
has been handwriting recognition. The task of handwriting recognition is difficult for a variety of reasons. The main explanation is
that different writers have distinctive writing styles. The abundance of characters, including capital letters, small letters, digits, and
special symbols, is a secondary factor. As a result, to train the system, a sizable dataset is needed. Since the system scans and detects
static images of the characters, optical character recognition (OCR) is frequently referred to as an off-line character recognition
method. We use the term "handwriting" to describe manuscript and cursive written writings. Because the characters are separated and
written in block letters, manuscript-style texts are simpler to recognise. Cursive handwriting, on the other hand, joins the characters as
they are written. To correctly perceive and recognise each individual character, handwriting recognition software is required. To
create various text recognition algorithms that can be converted from paper format to electronics. The writing style is not constrained
in a handwritten manuscript. The various human writing styles, variations in letter size and shape, and angles make it difficult to read
handwritten alphabets. A branch of OCR technology called handwriting recognition, often known as handwriting OCR or cursive
OCR, converts handwritten characters into corresponding digital text or commands in real-time. These systems use pattern matching
to recognise diverse handwriting styles to accomplish this goal. According to Wikipedia, handwriting recognition is: a computer's
capacity to read and comprehend legible handwritten input from sources like paper documents, photos, touch-screens, and other
devices.

II.Related Work:

The task of handwritten text recognition has been figured out during the past few decades with emphasis work. Various techniques
have been intended to be used in that effort. There have been difficult tasks that have required creative solutions for increased
precision and efficiency. Below is a brief summary of the various forms of recognition for work well done. Because of the numerous
areas of application, including the interpretation of ID numbers, processing of bank checks, recognition of ZIP codes, vehicle
registration numbers, and handwritten text, the topic of handwritten text recognition has received increased attention in pattern
recognition. As a result, handwritten text recognition is more widely used. around the globe. The feature extraction strategy that
produced the greatest results with the highest degree of precision after taking into account more traditional horizontal and vertical
approaches was concluded[1]. With this Neural network on better tested levels, an added feature with more noise tolerance and more
precise results are produced. The neural network used in this model is referred to as feed forward, and it was predominantly trained
using the back-propagation technique, which also provided training to a higher extent for character classification and recognition.
ISSN:0377-9254 jespublication.com Page 1
 Journal
Aside of Engineering
from that, characterSciences
recognition results are improved and are more precise when normalisation is used Vol in
14conjunction
Issue 02,2023
with
feature extraction. A design must be created for the most exact outcome. We must create a larger training data set for the optimal
neural network [1]. Because of the use of feature vectors, handwritten text recognition systems use artificial neural networks in
supervised learning methods. Bit map representations of input samples are used as feature vectors, which are first pre-processed and
then more effectively used by neural networks. This model provides better accuracy with unseen images that are inputs of the neural
networks for single images that consist of rotated images. Systems for optical text recognition are created using text that has been
scanned from documents and photos [3]. This includes text detection and recognition, where text is recognised and then classified .
Steps are categorised, and features are described in text using the template. A determined classifier is then applied to the described
characters.

They propose to use the unified character descriptor (UCD) to illustrate characters who fit that description. Then, categorization is
used to ensure that the matching. In consistently scanned text files, that recognition approach operates with excellent precision [4].
That does not indicate characters with more font variation or deformation. A neural network with a classifier is used to improve
recognition. Convolutional neural networks are very popular today because they perform better. Comparability between this paper's
work and the multilayer perceptron (MLP) plus CNN. In CNN, a convolutional network (Lenet-5) that supports 62 classes is used for
handwritten and machine-printed character recognition. Adding letters and the number [5] as well. We would have a database
specifically for the recognition task. Because this data base serves as the foundation for several handwriting recognition projects [6].
Amazing work has been done on the recognition in several languages, including Arabic [7] and Malayalam [6]. English, Bangla [8,
etc. Convolutional Neural Networks (CNN), a type of deep neural network, in particular, have provided promising results in a variety
of fields in recent decades. greatest accuracy With 96% accuracy, the CNN models used in this work produce superior outcomes for
recognising quantitative data, and 96% accuracy can continue to produce better results even when noisy conditions make it more
challenging.

III.THE PROPOSED ARCHITECTURE FOR HANDWRITTEN TEXT RECOGNITION USING MACHINE

LEARNING

In this machine learning project, we will use tensor flow to build a convolutional neural network (CNN) model that will
recognise handwritten characters, or the English alphabets from A to Z. A significant portion of the dataset on which we
will train our model includes pictures of the English alphabet.

An active topic of study in pattern recognition, artificial intelligence, and computer vision is handwriting recognition
(HR). The offline handwritten text recognition (HTR) process converts the scanned image into digital text as seen in Fig.
1. The neural network is created using an IAM dataset that has been trained and handwritten text recognition (HTR)
process converts the scanned image into digital text as seen in Fig. 1. The neural network is created using an IAM dataset
that has been trained and comprises word images. Although neural networks can be best executed on the GPU and CPU,
all layers are only slightly combined with word pictures. Handwritten Tensor flow text recognition requires
straightforward execution.

Fig. 1. Scanned text image to digital text conversion.

IV.MACHINE LEARING FOR DIGITAL TEXT CONVERSION

A contemporary area of study that includes pattern recognition, artificial neural networks, and computer vision is handwritten text
recognition. This use of artificial neural networks, which provides the ability to autonomously learn and improve from experience
without being explicitly designed, uses machine learning techniques. The creation of computer programmes that acquire information
and utilise it to learn for themselves is primarily what machine learning refers to. By using algorithms from touch screen devices and
pictures, handwritten text is recognised and its features are obtained. It is then translated into machine-readable form. Similar to this, a
variety of sectors and industries, including image processing, classification, regression, medical diagnosis, prediction, and learning
association, are adopting machine learning.

A. Training the model in machine learning

The IAM data source, which includes a comprehensive English sentence database for off-line handwriting recognition, was utilised in
the paper by Marti and H. Bunke [11]. the database of sentences written in different handwriting styles in English. This database
contains complete sentences in English. We would have a database specifically for the recognition task because this data base is used
as a base for several handwriting recognition projects [12, 13]. This database was developed using a substantial amount of word
images. To prepare the batch elements' average loss values are input into an optimizer like RMSProp using a neural network. All of
the words in this IAM data neural network are aligned to the left. So the system never learns how to handle the graphics in the white
area on the left. It is certainly possible to learn to overlook
ISSN:0377-9254 white space.
jespublication.com Page 2
Journal of Engineering
B. Framework use inSciences
machine learning Vol 14 Issue 02,2023

Different frameworks are used in machine learning to recognise handwritten text in a very accurate manner. Some of these topics
were covered, including Tensor flow, NumPy, and others.

1) Tensor Flow: Tensor flow is an open source library for large-scale machine learning and numerical computation. In deep learning,
tensors are the de facto standard for data representation. Although machine learning is a complicated field of study, using tensor flow
makes it much easier to put the models into practise. Since data acquisition, model training, shopping prediction, and feature
preservation are processes, machine learning frameworks like Tensor Flow have made this possible.

The health care sector of GE Healthcare is training a neural network to identify and specific anatomy during the MRI brain exam in
order to help speed and reliability [14]. Tensor flow is used by many companies, including Airbnb company, which is applying
machine learning using tensor flow to classify the emergence and detect objects at a scale helping to improve test experience. Tensor
flow is being used by PayPal to stay on the cutting edge of fraud detection. Adopting Keras for high-level APIs that make machine
learning relatively simple and for additional flexibility, as well as Eager execution that allows for fast iteration and debugging, are
features of Tensor Flow that set it apart from the competition. Tensor flow kernels run instantly when eager execution is enabled
rather than building graphs that will execute later.

Numpy: NumPy is a software that can carry out a number of operations and is utilised in scientific computing. Due to the operation of
multidimensional array tools and objects and the usage of the same datatype for storage, it provides exceptional performance in
arrays. This array's indexing follows numbers starting at zero. NumPy is a powerful N-dimensional object of array that is a general-
purpose array handing out package.

Edit distance: This is defined as a change between two strings that involves a minimum amount of insertions, deletions, and
replacements. Sys and argparse are two other frameworks that are utilised. There are system-specific parameters and functions in sys,
and this access is used for some variables. According to the needs of scripting, argparse offers a recommended command line parsing
module in the Python standard library.

IV.STUDY OF HANDWRITTEN TEXT RECOGNITION

It uses Optical character recognition algorithm to recognize the handwritten Text. By 2025, the market for optical character
recognition (OCR) is anticipated to be worth USD 13.38 billion, growing 13.7% annually. Rapid business process digitalization
employing OCR to cut labour expenses and conserve valuable man hours is what is driving this growth. Although handwriting
recognition (Handwriting OCR) or handwritten text recognition (HTR) is a crucial part of optical character recognition, it is still
regarded as a difficult problem statement. Converting handwritten text into machine readable text is extremely difficult due to the
wide variation in handwriting styles across individuals and the inferior quality of handwritten text compared to printed text. However,
it's an issue that several sectors, including banking, insurance, and healthcare, need to solve.

Today, typed characters in high-quality photographs can be read with greater than 99% accuracy thanks to OCR technology.
However, as you can see in the accompanying image, the variety of human writing styles, spacing variations, and inconsistencies of
handwriting lead to less precise character recognition. As a result, technologies that read handwriting cannot be as accurate as OCR
systems on typed text.

However, because of ongoing study in this area, handwriting recognition now employs extremely complex algorithms to address this
issue, increasing its accuracy.

A dynamic research field like pattern recognition, artificial intelligence, and others is called handwritten recognition (HR). Because
handwritten text contains so many varied styles of handwriting, it became challenging. Although word and character recognition rates
have increased, rates for offline handwritten text recognition have not increased as much. The field of text recognition has
experienced great success in a variety of real-world applications, including security applications, e-government systems, and other
areas [15]. The several methods of handwritten text recognition and how the process works are covered here. Here, each step for
handwritten text techniques and procedures is described in brief. Information that has been recognised primarily depends on the
recognition strategy systems that recognise text

ISSN:0377-9254 jespublication.com Page 3

Journal of Engineering Sciences Vol 14 Issue 02,2023

Fig. 2. Scanned text image to digital text conversion

VI. FLOW OF THE PROCESS

This section covered the fundamental workflow of a handwritten text recognition system. The system will take a handwritten text
image as input and produce the desired handwritten text as the output. The conventional text recognition system goes through a
number of steps, including image collecting, pre-processing, segmentation, and feature extraction, with classification as the last stage.
Some of these processes are occasionally combined by researchers. The basic process flow of the HTR system is shown in Fig. 3.

Online handwritten text recognition allows for real-time composition and recognition of handwritten text [16]. To create such type,
special digitizer tablets and pens are utilised. A digitizer has an electromagnetic tablet that continuously sends the coordinates of the
pen position to the computer. The primary examples that generate the text for online text recognition are tablets and Personal Digital
Assistants (PDAs). Due to the dynamic information that is typically present in online handwritten text, online handwritten text
recognition is easier than offline handwritten text recognition.

Text recognitions such as the number of strokes, the speed at which each stroke is written, and the direction of each stroke. When
compared to offline systems, this document recognition process typically results in systems that perform well.

Fig. 3. Basic process flow of HTR.

In the HTR system, the handwritten text data is first converted into a digital format with the fewest possible deteriorations. Using a
camera and scanner with a purchase, a grayscale or colour image is captured in offline handwritten word recognition mode.

Pre-processing, which comes after the image capture stage, is the main step that must be completed in a recognition method. Pre-
processing generally does not stop with handwritten text recognition. Pre-processing is used in picture processing. By eliminating
noise reduction, edge identification, picture normalisation, and thresholding, the pre-processing converts the image path to the
following stage of analysis. The acquisition circumstance in this case, such as poor document or file formatting, introduces noise.
Therefore, it is necessary to maintain relevant information while reducing noise as much as possible, which makes this a crucial pre-
processing step. When scanning a writing instrument or an image, noise can also be present. This can result in line segments that are
not connected and gaps between the lines, so it is first priority to remove this noise. Various techniques, including high pass filters,
low pass filters, median filters, and bandpass filters, can be used to accomplish this. In pre-processing, these filters are employed to
reduce noise.

PREPROCESSING OF THE DATA

 Before starting text recognition an image with text needs to be analyzed. For that purpose we need to provide preliminary

image preprocessing .
 The preprocessing algorithm includes following steps:
1.Skew Correction
2.Binarization
ISSN:0377-9254 jespublication.com Page 4
Journal of Engineering Sciences
3.Noise Removal Vol 14 Issue 02,2023
4.Thinning and Skeletonization

1.Skew Correction : We need to fix the text skew in an image with a block of text rotated at an unknown angle by:

1. Detecting the text in the image.

2. Finding the angle of rotated text.
3. Rotating the image to correct for the skew.

We typically apply text skew correction algorithms in the field of automatic document analysis, but the process itself can be applied to
other domains as well.

2.Binarization : It is the process of turning a coloured image into one with only black and white pixels (black pixels have a value of 0
and white pixels have a value of 255). As a general rule, this can be accomplished by setting a threshold, which is often 127 because it
corresponds to exactly half of the pixel range 0-255. The pixel value is regarded as a white pixel if it above the threshold; else, it is
seen as a black pixel.

3.Noise Removal : The noise removal stage's primary goal is to smooth the image by removing tiny dots or patches that are more
intense than the rest of the image. Both coloured and binary images can have their noise removed. Using the
fastNlMeansDenoisingColored function in OpenCV is one method of accomplishing noise removal.

4.Thinning and Skeletonization :

Depending on the setting in which OCR is being utilised, this preprocessing step is optional.

→ There is no need to complete this operation if we are using an OCR system for printed text because the printed text always has a
consistent stroke width. This operation must be completed since various writers have varied writing styles and, consequently, different
ISSN:0377-9254 jespublication.com Page 5
 Journal
stroke of Engineering
widths Sciences
when utilising Vol 14 Issue 02,2023
the OCR system for handwritten text. Therefore, we must execute Thinning and Skeletonization in order
to uniformly reduce the width of the strokes.

SEGMENTATION :

Any recognition system must include segmentation as a vital and essential phase. This step breaks apart the operation of the image for
the handwritten text. This results in the concealment of any image, including handwritten words, characters, or subwords. However,
this is not always achievable. Generally speaking, this procedure detracts from the system's credibility by In handwriting recognition,
there are two categories of segmentation methods: • Explicit segmentation: using this technique, characters and sub-words are
converted into characters or sub-characters. • Implicit Segmentation: This technique separates the lower, or so-called "grapheme,"
portions of the handwritten text, which are typically the words and characters. The words and characters are made up of these
grapheme-based elements.

Segmentation of the image done in three levels :

1. Line Segmentation
2. Word Segmentation
3. Character Segmentation
Line Segmentation : At this level of segmentation, a skew-corrected image with text that is written in lines is presented to us. To
segment the image into lines is the goal of line level segmentation.

If we horizontally project the binary image ,

 High numbers of foreground pixels in rows that indicate text in a line translate into higher peaks in the histogram.
 High numbers of background pixels are found in the rows that depict the spaces between the lines, which correspond to
lower histogram peaks.
 Rows that correspond to lower peaks in the histogram can be selected as the segmenting lines to separate the lines.
Word Segmentation : At this stage of segmentation, we are given an image with a single line that was previously segmented and
contains a list of words. Word level segmentation aims to separate the image into individual words. The idea is similar to the
previous step, but the only change is, here we have to project the image vertically (Sum is taken along columns) because we have to
segment words vertically.

If we vertically project the binary image,

 High numbers of foreground pixels in columns that indicate text translate into higher peaks in the histogram.
 High foreground pixel counts in text-indicating columns correspond to greater histogram peaks.

For segmenting words, lower peaks should be selected in such a way that they should span through a certain width (threshold). This is
because we’ll find lower peaks which correspond to the gaps between disconnected characters within a word, which we are not
ISSN:0377-9254 jespublication.com Page 6
 Journal of
interested in.Engineering
As we know,Sciences Vola 14
the gaps between the words are greater than the gaps between the characters within Issue
word, the 02,2023
threshold
should be chosen in such a way that neglects the thin gaps between the characters within the words.

Character Level Segmentation : At this stage of segmentation, we are given an image containing a single word that was previously
segmented and is composed of a string of characters. Character level segmentation aims to divide the image into distinct characters.
This level of segmentation is optional, which depends upon the context where the OCR is being used.

 Character Level Segmentation is not necessary if the OCR technology is being used on text when words have independent
characters. We can segment the characters in the previous step since a uniform space is maintained (even though it is small)
between the characters within a word (by setting a very low threshold).
 Character Level Segmentation must be used to separate the characters if the OCR system is being applied to text that contains
connected characters within words (cursive handwriting).


A few precautions should be followed while performing character recognition. The below figure shows one such problem, the
segments shown in the below figure are not accurate ,as h is extracted as l and i. Such errors are undesirable. Another precaution is of
ligatures . If the text image contains cursive type font then while segmenting the ligature should be separated for better efficiency.

FEATURE EXTRACTION :
This is the very important stage of handwritten character recognition system. The main objective of feature extraction is to extract all
the essential features of the scanned image.
In feature extraction stage each character is represented as a feature vector, which becomes its identity. The major goal of feature
extraction is to extract a set of features, which maximizes the recognition rate with the least amount of elements. Due to the nature of
handwriting with its high degree of variability and imprecision obtaining these features, is a difficult task. Feature extraction methods
are based on 3 types of features:
 Statistical
 Structural
 Global transformations and moments
Statistical Features :
 Representation of a character image by statistical distribution of points takes care of style variations to some extent.
 The major statistical features used for character representation are:
• Zoning
• Projections and profiles
• Crossings and distance

CLASSIFICATION :

Extracted features in the feature extraction steps are used to classify the images by assigning labels to the features.

Classification, arguably the most well-liked problem in machine learning, involves categorising images based on their properties and
in conjunction with a training dataset. For handwriting recognition, the essential elements of an input image are extracted and fed into
classifiers like logistic regression, kernel neural networks, and ANNs. These ML classifiers then classify the image more precisely by
comparing the featured image to the training dataset.

ISSN:0377-9254 jespublication.com Page 7

 Journal
POST of Engineering
PROCESSING : Sciences Vol 14 Issue 02,2023

Post-processing is in charge of correcting photos that were previously misclassified and changing detection at this step. It is
impossible to omit this procedure because it ensures the most accurate outcomes. Post-processing can be broken down into a number
of processes depending on the desired result.

Due to the significant issue of inter-class unevenness, feature extraction is one of the crucial processes in the recognition of
handwritten text. Characters have different forms depending on their positions. Useful data was needed by the handwritten text
recognition system to distinguish between one thing and other things. Therefore, this feature extraction stage is applied. Fourier
features, structural features, global transformations, and statistical features are a few examples of feature extraction techniques. Any
recognition system must include segmentation as a vital and essential phase. This step breaks apart the operation of the image for the
handwritten text. This results in the concealment of any image, including handwritten words, characters, or subwords. However, this
is not always achievable. Generally speaking, this procedure detracts from the system's credibility by In handwriting recognition,
there are two categories of segmentation methods: • Explicit segmentation: using this technique, characters and sub-words are
converted into characters or sub-characters. • Implicit Segmentation: This technique separates the lower, or so-called "grapheme,"
portions of the handwritten text, which are typically the words and characters. The words and characters are made up of these
grapheme-based elements.

VII. OVERVIEW OF WORK

As shown in Fig. 4, the model is created utilising recurrent neural networks, convolutional neural networks, and lastly, a connec-
tionist temporal classification layer. Equation (1), which transfers a picture (i.e., a matrix) M of size WH to a character sequence of
c1, c2, that has length from 0 and L, is utilised here as a neural network as a mathematical function. Due to the fact that text is
recognised at the character level, messages or words that are not included in the training data can also be recognised. Characters
can be accurately determined if they are defined as unique individuals.

Fig. 4. Overview and data process of the neural network.

A. Function of Model

CNN: convolutional neural network used to process scanned image input (CNN). The trained layers extract the picture's crucial key
characteristic. Initially, the layers apply a convolution function. The first two layers receive a kernel filter (size 55), while the final
three levels (33) receive a kernel filter. A non-linear RELU operation was then applied after this. Finally, the pooling layers condense
image regions. However, layers' image height is reduced by 2 and the channels are also added (maps). As a result, an output with a
feature sequence (or map) of 32256 size is shown in Fig. 6.

RNN: A time step of a feature sequence of 256 features is used by the recurrent neural network to convey pertinent information. This
RNN uses a broad LSTM execution and provides a considerably better training representation than a vanilla RNN. The output
sequence of the RNN records a matrix with a size of 3280. For the CTC blank label or CTC function, a 79-character IAM data base is
needed, thus every 32 time steps, a total of 80 accesses are made, as illustrated in Fig. 7.

CTC: Despite neural network training. The CTC output for RNN is supplied (Connectionist temporal classification). The ground truth
text is also supplied by CTC. The CTC determines the value of the loss. As a last step, CTC provides a matrix, which the final text
then uses to decode. Both the recognised text and the ground truth text have characters that are 32 characters long.
ISSN:0377-9254 jespublication.com Page 8
Journal
VIII. THEofPROPOSED
Engineering SciencesFLOW
ARCHITECTURAL Vol 14 Issue 02,2023

For the handwritten text recognition from input to output of handwritten text that convert to a scanned image to a digital text, the
major details of the model are briefly given below.

A. Input of the model

The image, or gray-value image, is standardised to make the duty of neural network simpler. Instead of arbitrarily scaling the image or
aligning it to the left, make a copy of the image and place it in random locations. Here, the gray-value image measures 128 by 32
pixels. This size is typically not exactly the same as what is needed for the data set. So, the image gets adjusted till it is either 32
inches tall or 128 inches wide. Then, this image is copied into a target image that is white and has a size of 128 by 32. An example of
this is shown in Figure 5, where the left image is a random size that has been mapped to fit a dataset target image with a size of 128
by 32, and the empty space is filled with the target level's white colour.

Fig. 5. Random image is converted into target image with white color

CNN Output: The output of CNN layers is a 32-word sequence, with each entry having 256 features. Additionally, these properties
are governed by RNN layers.

Fig. 6. Computation of 256 features per time-step by CNN layers, middle fig shows input image and below fig show 32nd feature that have character “e” image
with high correlation.

RNN Output: the top-most graph showing character scores, with the last (80th) row in the matrix and the CTC label left blank. For the
depiction of the RNN output, the image contains the text "small." Every entry in the matrix is also arranged from higher to lower.

IX. RESULTS AND DISCUSSION

The IAM off-line HTR dataset was used to train the tensor flow algorithm for handwritten text recognition. This neural network
model can recognise text that is contained within a word-segmented image. Complete text lines appear larger in the image than
individual words do. In this case, the neural network is kept small to make CPU training practical. Around 34 of the validation-words
set's are correctly predicted, and 10% of the words have character errors. The IAM dataset was used to pre-train this model, and
higher word accuracy was found when converting the input image to digital text. Here, the above-described word beam search
strategy is employed in this model with Python 3.6 and Tensor Flow 1.3 on Windows 10.

Figures 8 and 9 depict the output in relation to location, visible letters in the word, contrast of the image, writing style, and similar
factors of scanned input images of words. Accuracy can be increased by eliminating spelling errors, using writing styles, and in
relation to positions.

ISSN:0377-9254 jespublication.com Page 9

Journal of Engineering Sciences Vol 14 Issue 02,2023

Fig. 7. Output matrix, middle fig is input picture and below fig shows CTC blank label and probabilities of character like “l”, “i”, “t”, “e” of RNN layers.

Fig. 8. Output is “character” from input of “character” with probability 97.97 percent

Fig. 9. Output is “character” from input of “character” with probability 99.67 percent

X.CONCLUSION AND FUTURE WORK

When compared to the prior system, the new system achieves relatively good recognition rates for both the evaluation of text line
elements and the collection of horizontal, vertical, and skewed lines. This handwritten letter recognition neural network includes two
recurrent neural networks, five convolutional neural networks, and character probability matrices as outputs. Finally, this detection
and implementation using Tensor flow provides suggestions for improving accuracy. FASTag deployment in India and books are
currently the key application areas for this. digital translation from one language to another. By minimising several types of
handwriting, such as Data augmentation techniques include normalising contrast, skew, slant, and size of the handwritten text or
character, as well as other ways to increase the accuracy of neural networks.

REFERENCES:

[1] Paine, Scott W., and James R. Fienup. ”Machine learning for improved image-based wavefront sensing.” Optics letters
43.6 (2018): 1235-1238.
[2] Obaid, Ahmed & El-Bakry, Hazem & Eldosuky, M.A. & Shehab, Abdulaziz. (2016). Handwritten Text Recognition
System based on Neural Network. International Journal of Advanced Research in Computer Science & Technology. 4. 72-
77.
[3] Noor, Rouhan, Kazi Mejbaul Islam, and Md Jakaria Rahimi. ”Handwritten Bangla Numeral Recognition Using
Ensembling of Convolutional Neural Network.” 2018 21st International Conference of Computer and Information
Technology (ICCIT). IEEE, 2018.
[4] U. -. Marti and H. Bunke, ”A full English sentence database for offline handwriting recognition,” Proceedings of the Fifth
International Conference on Document Analysis and Recognition. ICDAR ’99 (Cat. No.PR00318), Bangalore, India, 1999,
pp. 705-708, doi: 10.1109/ICDAR. 1999.791885. (data base)
[5] S. Ben Driss, M. Soua, R. Kachouri, and M. Akil “A comparison study between MLP and convolutional neural network
models for character recognition”, Proc. SPIE 10223, Real-Time Image and Video Processing 2017, 1022306
[6] P. P. Nair, A. James and C. Saravanan, ”Malayalam handwritten character recognition using convolutional neural
network,” 2017 International Conference on Inventive Communication and Computational Technologies (ICICCT),
Coimbatore, 2017, pp. 278-281, doi: 10.1109/ICICCT. 2017.7975203.
[7] B. Alizadehashraf and S. Roohi, ”Persian handwritten character recognition using convolutional neural network,” 2017
10th Iranian Conference on Machine Vision andjespublication.com
ISSN:0377-9254 Image Processing (MVIP), Isfahan, 2017, pp. 247-251, doi: Page 10
Journal of Engineering Sciences
10.1109/IranianMVIP.2017.8342359. Vol 14 Issue 02,2023
[8] B. N. K. Sai and T. Sasikala, ”Object Detection and Count of Objects in Image using Tensor Flow Object Detection
API,” 2019 International Conference on Smart Systems and Inventive Technology (ICSSIT), Tirunelveli, India, 2019, PP
.)542-546, doi: 10.1109/ICSSIT46314.2019.8987942.
[9] D’souza, Lyzandra, and Maruska Mascarenhas. ”Offline handwritten mathematical expression recognition using
Convolutional Neural Network.” 2018 International Conference on Information, Communication, Engineering and
Technology (ICICET). IEEE, 2018.
[10] Q. Ye and D. Doermann, ”Text Detection and Recognition in Imagery: A Survey,” in IEEE Transactions on Pattern
Analysis and Machine Intelligence, vol. 37, no. 7, pp. 1480-1500, 1 July 2015, doi: 10.1109/TPAMI.2014.2366765.
[11] X. Yin, Z. Zuo, S. Tian and C. Liu, ”Text Detection, Tracking and Recognition in Video: A Comprehensive Survey,” in
IEEE Transactions on Image Processing, vol. 25, no. 6, pp. 2752-2773, June 2016, doi: 10.1109/TIP.2016.2554321.
[12] Pooja and R. Dhir, ”Video Text extraction and recognition: A survey,” 2016 International Conference on Wireless
Communications, Signal Processing and Networking (WiSPNET), Chennai, 2016, pp. 1366-1373, doi:
10.1109/WiSPNET.2016.7566360.
[13] E. A. Shamsan, O. O. Khalifa, A. Hassan and H. G. M. Hamdan, ”Off line Arabic handwritten character using neural
network,” 2017 IEEE 4th International Conference on Smart Instrumentation, Measurement and Application (ICSIMA),
Putrajaya, 2017, pp. 1-5, doi: 10.1109/ICSIMA. 2017.8312026.
[14] S. Espa˜na-Boquera, M. J. Castro-Bleda, J. Gorbe-Moya and F. Zamora- Martinez, ”Improving Offline Handwritten Text
Recognition with Hybrid HMM/ANN Models,” in IEEE Transactions on Pattern Analysis and Machine Intelligence, vol.
33, no. 4, pp. 767-779, April 2011, doi: 10.1109/TPAMI.2010.141.
[15] B. Purkaystha, T. Datta and M. S. Islam, ”Bengali handwritten character recognition using deep convolutional neural
network,” 2017 20th International Conference of Computer and Information Technology (ICCIT), Dhaka, 2017, pp. 1-5, doi:
10.1109/ICCITECHN.2017.8281853.

ISSN:0377-9254 jespublication.com Page 11