Recognition of Handwritten Assamese Characters

Saurabh Sutradhar, Dr, Ridip Dev Choudhary

[email protected], [email protected]

Abstract - Handwriting recognition has earned a fair The Assamese language has not been much explored due to
amount of attention in the field of pattern recognition. A its less usage. The Assamese language is a part of the
lot of work has been done so far for the efficient Assamese-Bengali Script prominently used in the North-
recognition of handwritten characters. So far, various Eastern region of India. The Assamese script consists of 11
techniques have been put forward for handwritten vowels and 40 consonants [15]. It also has 122 conjunct
character recognition. However, not much work has characters [15]. Recognition of Assamese Characters is a
been done yet for the recognition of Assamese difficult task because of the complex features of the
handwritten characters because of their complex characters like convoluted edges (ক্ষ, শ, ঋ, etc.), similarity
structure and limited usage. The Assamese language in (অ, আ, য, য়) appearance presence of loops (ভ, ঞ,
spoken in some of the northeastern parts of India is etc.), etc. Unlike English script, the Assamese script does
lagging in this field. This paper aims to make human not have the concept of capitalization of first characters.
Assamese handwriting machine-readable. Successful Also, different people have different hand- writing styles
recognition will enable machines to read forms, postal therefore it becomes a very laborious task of determining
addresses, signboards written in Assamese. It will also the characters.
lead to advancements in other areas such as translations Nowadays, CNNs have been effectively applied to pattern
of handwritten Assamese scripts to other languages as recognition, image classification, and forecast studies to
well as enable a text to speech system, thus making the name a few. For the classification of Assamese digits and
Assamese language more acknowledgeable. In this pa- characters, a few versions of CNN have been used so far
per, we have attempted to recognize handwritten [13,14]. This research area of recognizing Assamese
Assamese characters using Con- volution Neural characters has not been explored much. The potential of this
Network (CNN) by collecting our dataset. We have field is immense considering the popularity of the language
judged the performance of our system by the in its native state.
performance metrics: Precision, Recall, and F1 score.
We have used the performance metrics of precision, recall
Index Terms - Assamese Characters, CNN, Classification, and F1 score for our calculation. The formulas of the
Dataset. respective performance metrics are given below:
INTRODUCTION
Precision = Tp/ (Tp + Fp) Recall = Tp/ (Tp+ Fn)
Handwritten character recognition is a technique of F1 = 2 X [(P X R) / (P + R)]
detecting, segmenting, and identifying characters from Where,
images. The main objective of handwritten character Tp = Number of true positives
recognition is to replicate the human reading capabilities so Fp = Number of false positives
that the computer can understand, read and work as humans Fn = Number of true positives (Tp) + Number of
do with text. It has been one of the most interesting and false positives (Fp)
laborious research areas in the field of image processing and
pattern recognition lately. Several research works have been Additionally, this paper is divided into five sections. The
focusing on new methods that would reduce the processing first is the Related Work section, where we have given a
time without compromising high recognition accuracy. detailed survey in this area. Next is the Methodology section
where we have explained the workflow of our system. We
A lot of work for handwritten character recognition for have also discussed our dataset and explained each of the
languages like European, Chinese, and Arabic has been steps we have thoroughly used in our system. It is followed
done. Genetic algorithm and Artificial Neural Network by the Results and Discussion section which focuses on the
(ANN) were used as a classifier to recognize English digits results obtained in our research. In the end, we have the
and alphabets in [6]. Deep- Learning was also used to conclusion and future scope section where we have
recognize English digits and alphabets in [10]. Even for underlined the future breadth of this research area.
domestic Indian languages like Hindi, Bangla [4], Oriya [2],
RELATED WORKS
Malayalam [5], Tamil [11], etc. a lot of research has been
done till now. Indic scripts like Oriya, Telugu, Bangla, Printed and handwritten Assamese digits were compared in
Roman have already been explored extensively [2,4]. [1]. For the feature ex- traction and classification, part feed-
forward neural networks and a tree-based classifier were
used. Digits were scanned from the document after pre- language. MATLAB is used to implement the proposed
processed images were cropped with a bounding box to OCR system. The proposed method of OCR is a hybrid
obtain individual digits. Features were extracted from each approach for feature extraction combining both structural
bound area of the image. The grayscale documents were and statistical features. In this paper, the curve features are
converted into a binary image and the background noises extracted using the water reservoir principle and a decision
were removed with the help of linear filtering, medial tree classifier is used for classification.
filtering, and adaptive filtering. After that skew detection
and correction were per- formed, following line, word, and An Artificial Neural Network (ANN) based approach was
character segmentation. used to segment hand- written text in Assamese [7]. After
the feature extraction part, the input was fed to an ANN
Word-level script identification for six handwritten Indic model, and the similarity measure was found. A text-to-
scripts (Bangla, Devanagari, Gurmukhi, Malayalam, Oriya speech synthesizer was used in [8] to facilitate English text
Telugu, and Roman) was proposed in [2]. This pa- per reading. The text was manually typed in the screen and the
proposed the elliptical and approximation approach to system was made using MATLAB. Handwritten Sindhi
design features. The original images are grayscale and numerals were simi- larly recognized in [9] using K-NN and
Otsu’s Global thresholding method is used to convert them SVM. They evaluated their system using the correlation
into binary images, removing the noisy pixels from the coefficient. The Sindhi numeral 0 achieved an accuracy of
binarized image using Gaussian Filter. For the classification 100% whereas the numeral 3 achieved 63%. Then for Tamil
part, the feature sets have been applied to 7 different script, features like character height, width, slope, etc. were
classifiers namely- Naïve Bayes, Bayes Net, MLP, SVM, considered in [11]. Zernike moments were used for feature
Random Forest, Bagging, and Multi-Class Classifiers. The ex- traction and then fed to a backpropagation model.
multi-layer perceptron (MLP) achieved the highest accuracy
of 94.35%. Similarly, for the classification of Assamese handwritten
digits, vowels, and consonants a zoning feature was used in
A computational model for Handwritten Assamese text [12]. A feed-forward neural network model was used with a
was proposed in [3] to per- form the preprocessing, text sigmoid function at every neuron to calculate the output. For
segmentation, and then extraction of different features from digits, vowels, and consonants a recognition accuracy of
individual characters. To segment a document image into 70.6%, 69.62%, and 71.23% was achieved respectively. In
various parts, a global projection profile approach of a word [14] Assamese handwritten digits were recognized by using
was used to identify the upper, lower, and middle areas of a CNN (Digi- Net model). This paper used six alternative
word. A combination of diagonal features using zoning Convolution and Max-Pooling layers which were later
concept and texture features via GLCM (Grey Level Co- followed by a Fully Connected layer and a Softmax
occurrence Matrix) were computed for extracting various classifier. CNN was further used in [13] by using LeNet-5,
features of individual characters. The mean and standard ResNet-50, Inception V3 and DenseNet-201 model.
median filters (SMF) were used to clean noises from an
input image. Random transform-based techniques were used METHODOLOGY AND DATASET
for skew detection and correction. An artificial neural The proposed system consists of the following phases:
network as a backend was used to perform classification and  Data Collection
recognition tasks.  Digitization and Normalization
 Preprocessing
A modified ResNet-18 architecture (a convolution neural  Convolution neural network / Classification
network architecture) was proposed in [4] to recognize
Bangla handwritten characters. This paper considered three The block diagram is given in Figure1:
main challenges: Recognizing convoluted edges to
distinguish between the repetitions of the same pattern in
different characters and different handwritten patterns for
the same characters. To have wider input for the generalized
performance of the network, input images are preprocessed
through the removal of noise with the median filter, edge
thickening filter, and the image is resized to a square shape
with appropriate paddings by default. The dataset used for
this paper is the two recently introduced datasets- Bangla
Lekha-Isolated dataset and CMATERdb dataset.

A hybrid approach for recognizing Malayalam

handwritten characters was proposed in [5]. It takes into
account both the dependent and independent features of the Fig1: Workflow diagram of our system
 DATA COLLECTION

With the above-mentioned literature survey, it was observed

that there are quite a few algorithms that are commonly used
in the disease prediction process. Some of them are
discussed below:

The collection of datasets is the first and most necessary

part of the work. We went to a local school to ask small
children to write the Assamese characters in different font Fig. 4. After Preprocessing
sizes, for our research work. People from 7 years and above
were asked to fill in the boxes of 0.75 cm2 ,1.0 cm2, and
1.25 cm2. We wanted to include diversity and variation in  CNN / Classification
our dataset as much as possible.
We divided our dataset into two parts in an 80:20 ratio,
namely training and testing. The training dataset consists of
7,380 images and the testing dataset consists of 1,845
images. After dividing the datasets, the training images are
fed to Convolution Neural Network.

A CNN pipeline model usually is made of a feature

learning phase and a classification phase. The feature
learning phase consists of the Convolution and Pooling
layers. It learns the high-level features of our pre-processed
input image. The Convolution layer performs linear
operations in the input image for feature extraction. Then
we have used the relu and softmax functions as the
activation function to ensure non-linearity in our network.
Then for the pooling layer, we have used the max-pooling
function to reduce the dimensionality of the network and
detect the strong features. The convolution and pooling
Fig 2: A sample page of our collected dataset operations are performed on every pixel of the input image
independently. The output of the feature learning phase is a
 Digitalization and normalization flattened column vector.

After the collection of datasets, the image of every character Since we must classify 41 characters (consonants), our
was taken by our phone cameras. The images were then output layer consists of 41 nodes. Then for the classification
manually cropped. For every character, 225 images were phase, a feed-forward network must be established. It is
collected. So, our total dataset consists of 9,225 images. achieved by mapping the flattened column vector to the
Each image of the dataset is fixed to a size of 50 x 50. output layer with the aid of a fully connected layer. This
fully connected layer learns the logic behind the feature
 Preprocessing learning phase and performs the classification of Assamese
characters. We have used five layers in our CNN network,
The images were converted into a grayscale image. The the dropout and dense layers being alternatives. We went for
color of the images was then inverted. For noise removal, categorical cross entropy and RMSprop() for the loss
we used Gaussian blur and Otsu’s thresholding. Gaussian function as the optimizer.
blur was used for smoothening and Otsu’s thresholding was
used for filtering in our system. In the testing phase, the testing dataset (1845 images were
from the test folder) was compared with the training dataset.
We tested for 41 Assamese alphabets.
RESULTS AND DISCUSSIONS
We evaluated the performance of the classification by the
metrics accuracy, recall, precision, and F1 score. We went
for a batch size of 128 with epochs of 15. The result of our
system has been shown in Table 1.
Fig. 3. Before Preprocessing
 ACCUR PRECISI RECA F1 SCORE [11] Wahi, A., Sundaramurthy, S., Poovizhi, P.: Handwritten
ACY ON LL Tamil Character Recognition. Fifth International
Conference on Advanced Computing (ICoAC). 2013
70.02% 68.5% 69.23% 70.71% [12] Medhi, K., Kalita, S.: Assamese Character Recognition
Using Zoning Feature. Advances in Electronics,
Communication, and Computing. January 2018
CONCLUSION AND FUTURE SCOPE [13] Yadav, M., Mangal, D., Srinivasan, N., Ganzha, M.:
Assamese Character Recognition using Convolutional
We have discussed the progress in the field of Assamese Neural Networks. May 2021
handwritten character recognition in this paper. In our [14] Dutta, P., Muppalaneni, N.: DigiNet Prediction of
Assamese Handwritten Digits using Convolutional
research work, we have achieved an accuracy of 70.02%, Neural Network. Concurrency and Computation:
precision of 68.5 %, recall of 69.23%and F1 score of Practice and Experience. June 2021
70.71% by using a convolution neural network. The [15] Wikipedia,
accuracy can be further improved by increasing the size of https://en.wikipedia.org/wiki/Assamese_alphabet, last
accessed 2021/10/12.
the dataset, adding more pre-processing steps, or by using a
hybrid model.In the future, the Assamese handwritten
character recognition system can further be extended by AUTHOR INFORMATION
adding a voice feature in the post-processing part. The
recognition part could also be used to identify conjunct Saurabh Sutradhar, Research Scholar, Discipline of
characters shortly. After successful recognition of conjunct Computer Science, Chandra Bhuyan School of Science and
characters, words and sentences can be targeted next. Technology, KKHSOU

Dr. Ridip Dev Choudhury

REFERENCES Associate Professor, Discipline of Computer Science,
Hiranya Chandra Bhuyan School of Science and
[1] Medhi, K., Kalita, S.: Assamese Digit Recognition with
Feed Forward Neural Network. International Journal of Technology, KKHSOU
Computer Applications. Volume 109 - No. 1 (January
2015).
[2] Singh, P., Doermann, D.: Word-level Script Identification
for Handwritten Indic Scripts. 13th International
Conference on Document Analysis and Recognition
(ICDAR). IEEE, 2015
[3] Bania, R.: Handwritten Assamese Character
Recognition using Texture and Diagonal Ori- entation
features with Artificial Neural Network. International
Journal of Applied Engineering Research 13.10 (2018):
7797-7805.
[4] Alif, M., Ahmed, S., Hasan, M.: Isolated Bangla
Handwritten Character Recognition with Convolutional
Neural Network. 20th International Conference of
Computer and Information Technology (ICCIT). IEEE,
2017.
[5] Sujala, K., James, A.Saravanan, C.: A Hybrid Approach
for Feature Extraction in Malaya- lam Handwritten
Character Recognition. Second International
Conference on Electrical, Computer and
Communication Technologies (ICECCT). IEEE, 2017
[6] Agarwal, M., Kaushik, B.: Text Recognition from Image
using Artificial Neural Network and Genetic Algorithm.
IEEE, 2015
[7] Bhattacharya, K., Sarma, K.: ANN-based Innovative
Segmentation Method for Handwritten Text in
Assamese. IJCSI International Journal of Computer
Science Issues, Vol. 5, 2009
[8] Gopinath, J., Aravind, S., Chandran, P., Saranya, S.:
Text to Speech Conversion System using OCR.
International Journal of Emerging Technology and
Advanced Engineering, Volume 5, Issue 1, January 2015
[9] Sanjrani, A., Baber, J., Bakhtyar, M., Noor, W., Khalid,
M.: Handwritten Optical Character Recognition System
for Sindhi Numerals. IEEE 2015
[10] Vaidya, R., Trivedi, D., Satra, S.: Handwritten Character
Recognition Using Deep-Learning. 2nd International
Conference on Inventive Communication Technologies
(ICICCT). 2018