Ethiopic Characters Recognition Model using Deep Learning

doi:10.21203/rs.3.rs-1751009/v1

Download PDF

Research Article

Ethiopic Characters Recognition Model using Deep Learning

https://doi.org/10.21203/rs.3.rs-1751009/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this older preprint version

Read the latest preprint version →

Purpose: - Many approaches have been developed to recognize different writing scripts, most of which focus on Latin scripts. However, in the Ethiopic script unavailability of readymade datasets and the attempts of a few studies show that the script requires further investigation by applying different recognition techniques. In addition, image quality degradation and numerous characters in the alphabet are other major problems with Ethiopic-script documents. In this study, a convolutional neural network-based character recognition model for Ethiopic character recognition is proposed. The contribution of this work is the development of a character recognition model using a convolutional neural network.

Method: - The model is designed using seven convolutional layers, four max-pooling layers, and two fully connected layers. Automatic features are extracted from each cropped 28 by 28 images using convolutional and max-pooling layers to prepare for recognition.

Results: - The proposed model is trained and tested with customized datasets and achievement of accuracy 97%, which is a promising recognition result of the script.

Conclusion: - In future works, the dataset should include all the letters, punctuation marks, and symbols found in Ethiopic scripts for both printed and handwritten text and increase the dataset size for different deep learning algorithms to get better results.

Artificial Intelligence and Machine Learning

Convolutional Neural Networks

Ethiopic Scripts

Optical Character Recognition

Deep Learning

Currently, many devices such as mobile, smart pads, and computers are commonly used in daily life to facilitate various tasks. However, applications of these devices often require data from different sources of documents for various government or commercial services [1]. It is obvious that governmental organizations can generate huge numbers of documents while they perform their daily activities. In fact, the huge number of documents are difficult to manage using human labor. Even it is hard to prevent documents from natural disasters such as fire, wind, animals, and moisture. Unless, using automatic recognition techniques and storing them on a remote server such as a cloud, it is difficult to preserve the resource from risks.

Optical character recognition (OCR) is a solution for extracting and recognizing characters from scanned an image [2]. It also plays a role in converting them into editable text. OCR is one of the most important applications of image pattern recognition [3]. It is useful for several applications such as post offices, banking, and vehicle numbers recognition from vehicle plates number [4].

Prior to this study, many character recognition techniques have been developed for recognizing different writing scripts. However, in this area, most of the applied techniques focus on Latin scripts. In the Ethiopic script, few studies are done. Ethiopic script recognition is a challenging task owing to a large number of characters, cursive letters, and shape similarities in the alphabet. In addition to this, image quality degradation is another main problem in document analysis. When documents are not handled properly, a recognition system difficult. Due to these problems, the development of an accurate Ethiopic-script recognition system is required.

The purpose of this work is to recognize characters from image documents that consist of texts. The recognition follows a series of operations such as image pre-processing, segmenting images into subparts (e.g., lines, words, and characters), and identifying the text in subparts that include characters.

The contribution of our work is the development of a character recognition model using a Deep Convolutional Neural Network (DCNN) for Ethiopic script.

The remainder of this paper is organized as follows. Section 2 describes the literature review; Section 3 introduces the methodology of the study; Section 4 shows the results, and a conclusion of the work is given in Section 5.

In this section, the concept of Ethiopic script and related studies in the area of optical character recognition are discussed. Here, we describe different techniques previously applied to Ethiopic and other language scripts. The results and data sources used are also explained.

The Ethiopic Scripts

The Ethiopic script has been used to write ancient histories, sciences, and the arts of countries such as Ethiopia and Eritrea [5]. The first language that uses Ethiopic script is Geez. It has 26 base characters. Other Ethiopian languages also start with this script and added additional characters. For example, the Amharic language adds seven other base characters. In the Ethiopic script, the reading process is from left to right horizontally and each letter generates different sounds depending on the base character accent. Fig. 1 displays the alphabet arrangement of the Ethiopic script for Ge'ez language.

Ethiopic Script Studies

Some researchers have highlighted their contribution to the field of Ethiopic-character recognition. Using statistical and traditional machine learning methods have been proposed to recognize Ethiopic text written by hand or printed. Table I shows some list of Ethiopic script studies on both handwritten and printed text.

Table I. List of Ethiopic studies on both handwritten and printed text

Title	References	Techniques
“Amharic Handwritten Character Recognition Using Combined Features and Support Vector Machine”	[6]	HOG+LBP+SVM
“Handwritten Amharic Bank Check Recognition Using Hidden Markov Random Field”	[7]	HMRF
“HMM-Based Handwritten Amharic Word Recognition with Feature Concatenation”	[8]	HMM
“Amharic Character Image Recognition”	[9]	CNN
“Printed Ethiopic Script Recognition by Using LSTM Networks”	[10]	LSTM
“Ethiopic Character Recognition Using Direction Field Tensor”	[11]	Direction Field Tensor
“Multifont size-resilient recognition system for Ethiopic script”	[12]	Direction Field Tensor
“Deep learning for Ethiopian Ge'ez script optical character recognision”	[13]	CNN
“A Hybrid System for Robust Recognition of Ethiopic Script”	[14]	ANN and Template matching
“Amharic Text Image Recognition: Database, Algorithm, and Analysis”	[15]	CTC
“Optical Character Recognition for Tigrigna Printed Documents Using HOG and SVM”	[16]	HOG+SVM

Reta and Rana[6] conducted a study on handwritten Amharic character recognition. They used combined feature extraction techniques such as Histogram of Gradient (HOG), Local Binary Pattern (LBP), and geometrical features. In their study Local Discriminant Analysis (LDA) used to reduce the dimensions of combined features. A multiclass support vector machine was used as the classification algorithm and achieved the best accuracy of 92.8%. Alemu and Fuchs [7] studied handwritten Amharic Bank check recognition. They applied Hidden Markov Random Field (HMRF) features and pseudo-marginal probability classification method to achieve 89.02% and 99.44% accuracy without context and with context, respectively. Assabie and Bigun [17] conducted a study on Amharic words in unconstrained handwritten text. They used lexicon-based offline recognition and achieved a recognition rate of 73% for good quality texts and 36% for poor-quality texts. They stated that error rates mainly generated by sword segmentation and interconnected characters. They also attempted another study [8] using a Hidden Markov Model (HMM) with feature concatenation.

In the year of GPU, different studies have been conducted based on deep learning in recent years for Ethiopic scripts, such as [9], [13], [5]. Belay et al. conducted a study of Amharic printed character image recognition applying CNN with synthetic data [9]. In their approach, five convolutional layers, three max-pooling layers, and two fully connected layers considered to achieve an accuracy of 92.71%. However, applying deep learning on Ethiopic OCRs needs more attempts. That is why we proposed this study to get better accuracy.

Other script studies

Ahamed et al. proposed a study on handwritten Arabic numerals using a convolutional neural network [4]. They achieved 99.68% accuracy, which is an impressive result.

Alizadehashraf and Roohi studied handwritten Persian character recognition using CNN [3]. They attempted by using a single CNN and ensemble CNN techniques in their experiments. The results achieved in their study are 97.1% and 96.3% for the ECNN and SCNN techniques, respectively.

Historical Arabic manuscript text recognition was performed by Alsaedi et al. using convolutional neural networks [18]. They adjusted their collected dataset into three parts and conducted experiments. The results achieved using this technique ranges between 74.29% to 88.20%.

Aly and Mohammed have done a study using the cascade of PCA-SVMNet classifiers to recognize handwritten digits [19]. They achieved state-of-the-art recognition accuracy compared to other segmentation-free techniques.

Asha and Krishnappa [20] proposed a document recognition for the Kannada language. To implement the recognition system, they used a convolutional neural network. By considering non-overlapping lines of characters in the document, their proposed method achieved an accuracy of 98% [21].

Purkaystha et al. attempted the handwritten character recognition of Bengali characters using CNN [22]. The authors achieved results ranging from 89.93% to 98.66% for various data ranges, such as numerals, vowels, and all Bengali characters. Other authors[23] also evaluated Urdu handwritten text by integrating it with the MNIST dataset and achieved an accuracy of 93.0% by applying transfer learning.

Nair et al. conducted a study on Malayalam handwritten character recognition using CNN as the recognition algorithm [24]. Lui et al. has been attempted a Chinese license-plate recognition system [25]. An average recognition rate of 98.09% was achieved. Kim and Xie proposed a study using a deep convolutional neural network for handwritten Hangul recognition, and they achieved a 99.71% recognition rate on the MNIST database [26].

In this section we describe the proposed CNN based Ethiopic character recognition model. The major components of the proposed model and the data source used are described. The concepts about CNN which is one of the deep learning algorithms is explained in detail.

3.1 Data Sources

Binary ground-truth data are used in this study, which is collected from scanned images from a different document source. Synthetic data provide the necessary font and background diversity without time-consuming and challenging data preparation [1]. Our data source in found in [15] and customize them in to three parts and resize the character image by cropping. Each input image has been cropped to 28 × 28 dimensions to make it the same size because the selected algorithm accepts the same size.

Figure 2 shows chosen sample character images from dataset used and their corresponding syllables.

3.2 Deep Learning Algorithm and Description

Deep learning is a subset of machine learning in the field of artificial intelligence that mimics the neuron of the human brain or neural networks [27, 28]. It allows machines to solve complex problems in the field of pattern recognition and computer vision, even if the data is very diverse, unstructured, and interconnected. Artificial neural networks play a significant role in doing these. But regular artificial neural networks have not scalability for varied and large data sizes. To avoid such limitations, deep neural networks have the advantage of providing integrated feature extractors and classifiers [26].

In deep learning techniques, many algorithms exist. But in this study, we focused on convolutional neural networks since it accepts raw images as input and does not require separate feature extraction, except for size normalization [26].

Convolutional neural networks (CNNs or ConvNets) are a kind of neural network used for supervised learning [20] and one of the most promising deep learning techniques to deal with many pattern recognition tasks [29]. It is preferable for image or video data processing, especially for object detection, face recognition, pattern recognition, etc.

In regular artificial neural nets, the networks receive a single vector input and transform it through a series of hidden layers to reach the last fully-connected layer known the output layer. Each hidden layer receives the fully-connected neurons from the previous layer. For the classification applications, the output layer holds the class scores. Whereas the convolutional layers of CNNs have 3D arranged volumes of neurons. For example, if our input data is an image for the CNNs, the image should be arranged in 3D (width, height, depth) format. Here, width and height represent the dimensions of the input image, and the depth shows the channels (i.e., R, G, B) of the image. The following Fig. 3 and Fig. 4 clearly shows the diagrammatical arrangements of the neurons in regular networks and CNNs, respectively.

As the literature says in recent times, CNNs have been used for character and numeral recognition tasks extensively [4]. It has the advantage of sharing weight across CNN layers that reduces parameter numbers and improves performance [9].

3.2.1 Convolutional Neural Network Layers

Convolutional neural networks consist of multiple layers, as we set Fig. 4. The convolutional layers, pooling layers, and fully-connected layers are the main layers used to build CNNs architecture. In addition to containing layers, a CNNs architecture requires activation functions such as sigmoid, ReLU, tanh, Leaky ReLU, etc. to be functional. In general, it is possible to classify the CNNs in two parts i.e. feature extraction and classifications. On the feature extraction part, the convolution and pooling operations play a role. Whereas for the classification part, the fully-connected layers take part. The brief descriptions of CNNs layers are found in the following sub-section.

Feature extraction

Features are the pixel information of an image. Features are important components of character recognition. In image recognition relevant features must be extracted for recognition. To extract the relevant features CNN consists of input, convolutional, activation, padding and pooling layers.

Input Layer: - In this specific layer CNN should contain image data. The image data may be color or grayscale three-dimensional matrix filled by pixel values. we need to reshape the pixel values into a single row. The pixel values are normalized by dividing each value by 255 to make it in the range of 0 to 1, which has the advantage of making the learning faster and getting a better model performance.

Convolution Layer: - Convolution is one of the core building blocks of the CNNs of the convolutional layers because features of the image are get extracted within this layer. The operation of the convolution is computed by sliding the vital component called kernels across the patches of the input image [30]. The kernel has a role in creating a feature map which holds the values computed from element-wise matrix multiplication and the sums of it. Later, this feature map is fed to other layers to learn several other features of the input image.

In the following formula, the feature map F is calculated from the input image I and kernel K. The indexes of the width and height of the result matrix are marked with a and b, respectively.

$$F\left[a,b\right]=\left(I\text{*}K\right)\left[a,b\right]=\sum _{i}\sum _{j}I\left[a-i,b-j\right]K[i,j]$$

Here Fig. 5 shows how a 3x3-pixel filter convolved over the 6x6-pixel input image and generates a 4x4 feature map. Formula 2 specifies the size of the feature map.

$${n}_{new}=n-f+1$$

Where nnew represents feature map size, n represents the dimension of the input image, and f represents the filter or kernel dimension. In this study, we applied kernel size (3,3) at each convolutional layer.

Activation / Non-Linearity: - Any kind of neural network contains activation or non-linearity to be powerful. In hidden layers we may want to consider activation functions such as Rectified Linear Activation (ReLU), Logistic (Sigmoid), Hyperbolic Tangent (Tanh).

In CNN's, the ReLU is the most commonly used activation function to achieve its real potential on convolution results and its bias value. The ReLU activation helps to make all negative value to zero. The following Eq. 3 indicates the mathematical expression of the ReLU function.

$$y=max\left(0,\left(w*I\right)+b\right)$$

where w is weights and I is input value and b for bias.

Stride and Padding: - Stride specifies the number of moves of kernel at each time. By default, the stride size is usually one, which means the kernel slides pixel by pixel. It is also possible to set a bigger size if less overlapping required between the receptive fields. Using a bigger kernel size makes the resulting feature map smaller, and skipping will happen over potential locations. In our model, we directly applied the default stride value.

In ConvNet, the feature map size is always smaller than the input size because of the convolution applied. Then dimension mismatch will be happening with input size, and shrinking exist on a feature map. To preserve the feature maps shrinking using padding is desirable techniques when padding applied to our computation of features mathematically computed as follows.

$${n}_{new}=n-f+2p+1$$

Where nnew represents feature map size, n represents the dimension of an input image, f presents the filter or kernel dimension, and p is the paddings applied to the input image.

The general convolutional layer is expressed in the following equations.

$${C}_{p}^{l}=\sigma (I\text{*}{K}_{l,p}^{l}+{b}_{p}^{l})$$

Where $\sigma$, is an activation function used, for example, if we are using a sigmoid activation function, the equation becomes the following.

$$\sigma \left(x\right)=\frac{1}{1+{e}^{-x}}$$

$${C}_{p}^{l}\left(a,b\right)=\sigma \left(\left(I\text{*}K\right)\left[a,b\right]+{b}_{p}^{l}\right)=\sigma \left(\sum _{i}\sum _{j}I\left[a-i,b-j\right]\text{*}K\left[i,j\right]+{b}_{p}^{l}\right)$$

Where C is a convolution layer, l indicates layer number, and i, j denotes the map indices of current and next layers, respectively. Indices a, b is a row and column size of the feature map, respectively. σ is the activation function, I is an image, K is a kernel, and b is the bias [31].

Pooling Layer: - The Pooling Layer usually serves as a bridge between the Convolutional Layer and the Fully Connected Layer. After convolution, activation, stride, and padding operations are done, performing pooling is the next step. Pooling is done at the pooling layer, which takes the output of the previous layer and down-samples by choosing the prominent features [30]. The Down-sampling technique reduces the dimensionality of each convolutional layer. So, a reduction of the dimensions makes computationally efficient. It may create an advantage for both minimizing the training time and overwhelmed overfitting.

Even many pooling techniques are found; max-pooling is the most common one to facilitate the pooling operation because it keeps the most important or most abundant features. It takes the maximum value by sliding "n x n" windows from the input. Figure 6. Max pooling: 2x2 windows from 4x4 input and 1 step stride

Figure 6 shows the computation process and its result of the max-pooling layer using a 2x2 window with stride 1.

In our study, also we applied the max pooling.

Classification

Fully-connected layer: - In convolutional neural networks, a fully connected layer is just like a hidden layer of a regular neural network that consists of a 1D dimensional vector value. Since the convolutional layer generates 3D data, it should be flattened to have 1D data. The flatten data passes to one or many fully connected layers to feed the SoftMax unit, which consists of output classes.

$$\widehat{o}={\sigma }(\text{w}\text{*}f+b)$$

Output layer: - The output layer unit assigns a probability value to all the possible classes [32]. The one which contains the largest amount must be taken as the predicted class label. One-hot matrix representation takes these probability values and keeps the largest probability value to be "1", and the remaining is set to "0". In ConvNet, the SoftMax unit is taken as an output layer, and it calculates a probability of a given image to be in a particular output class.

$$p(y=j|{\theta }^{\left(i\right)})=\frac{{e}^{{\theta }^{\left(i\right)}}}{\sum _{j=0}^{k}{e}^{{\theta }_{k}^{\left(i\right)}}}$$

Where, "${\theta }^{\left(i\right)}$" holds the value of the transpose of the weight's matrix $W$ multiplied by the feature vector X. Given as follows:

$$\theta ={W}_{0}{X}_{0}+{W}_{1}{X}_{1}+\dots +{W}_{k}{X}_{k}=\sum _{i=0}^{k}{W}_{i}{X}_{i}={W}^{T}X$$

Where $X$ and $W$ is the network input the weight vector, respectively. The "${W}_{0}{X}_{0}$" is the bias.

Loss function: - Since this work is a multi-class classification problem, we applied the multi-class cross-entropy loss function that used to optimize the parameter values in the model.

Assuming that the actual output label is $O$, the loss function is expressed by:

$$L = \frac{1}{2}\sum _{i=1}^{n}{(\widehat{O}\left(i\right)-O\left(i\right))}^{2}$$

3.3 The proposed model

In this study the proposed algorithm is convolutional neural network which consists of seven convolutions followed by ReLU activation, two max-pooling, two fully-connected, two dropout layers, and output with a SoftMax layer. Figure 7 depicts the block diagram of the proposed model. Furthermore, Table II described the detail specifications of all network parameters and size in the model.

Table II. Convolutional Neural Networks configurations

Layer Type	Network parameters
Input Layer	Input Image: 28281
Convolution2D	Filters: 32, activation: ReLU, kernel: 3*3, strides: 1, padding: same
Convolution2D	Filters: 32, activation: ReLU, kernel: 3*3, strides: 1, padding: same
Max-pooling	Pool_size: 2*2
Convolution2D	Filters: 64, activation: ReLU, kernel: 3*3, strides: 1, padding: same
Convolution2D	Filters: 64, activation: ReLU, kernel: 3*3, strides: 1, padding: same
Max-pooling	Fool_size: 2*2
Convolution2D	Filters: 96, activation: ReLU, kernel: 3*3, strides: 1, padding: same
Convolution2D	Filters: 96, activation: ReLU, kernel: 3*3, strides: 1, padding: same
Max-pooling	Pool_size: 2*2
Dropout	0.25
Convolution2D	Filters: 128, activation: ReLU, kernel: 3*3, strides: 1, padding: same
Max-pooling	Pool_size: 2*2
Dropout	0.25
Flatten	---
Fully-connected	Filters: 256, activation: ReLU
Dropout	0.25
Fully-connected	Filters: 128, activation: ReLU
Output	Classes Labels: 231, softmax
Model Compilation	Optimizer: adam, metrics: accuracy, loss function: categorical_crossentropy

The proposed model is depicted pictorially in Fig. 7, which consists of CNN component symbols and labels.

3.4 Training CNN Model

The evaluation of the proposed model is done with a local dataset. The dataset first divided into two, i.e., training and testing. Here 80% of the dataset is used to train while the remaining is used to test the model. This model has been trained using the stochastic gradient descent optimization algorithm. The number of epochs in each experiment is 100.

3.5 Experimental Tools

In this study we have done extensive experiments using google Colab tool with python version 3.

To evaluate the proposed model, we used the customized Ethiopic character image dataset, which is classified into three parts, as depicted in Table III. Dataset_1 consists of Geez base characters; Dataset_2 holds Ethiopic base character; Dataset_3 presents all Ethiopic script data.

Table III. Total number of classes and samples in each dataset

Dataset	Number of classes (output labels)	Number of examples
Dataset_1	26	8911
Dataset_2	33	11191
Dataset_2	231	75995

The proposed system is configured and executed on the Google Colab platform with Graphical Processing Unit (GPU) service. During the execution of the experiments, hardware reserved by Google Colab is RAM 12.72 GB and 107.77 GB hard drive. This hardware requirement is enough to experiment appropriately.

Apart from specific libraries, all algorithms have been developed and implemented with essential software tools and packages, as mentioned in Table IV.

Table IV. Software tools and packages for experimentation

Tools and packages	Version	Description
Python	3.6.	Programing Language to write code
Colab	Latest	Editor with Google compute engine back end
OpenCV	4.3.0	Supports image processing
TensorFlow	2.0.0	Symbolic math library
Pillow	Latest	Supports image file management
Keras	2.3.0	Wrapper for TensorFlow

Table V clearly shows the model summary output while we are using the Dataset_1 as the training dataset. Here there are 383, 418 total trainable parameters, and not untrainable parameters existed.

Table V. Model summary for Dataset_1

Layer (type)	Output shape	Parameters
cov2d (Conv2d)	(None, 28, 28, 32)	32
cov2d_1 (Conv2d)	(None, 28, 28, 32)	9248
max_pooling2d(MaxPooling2D)	(None, 14, 14, 32)	0
cov2d_2 (Conv2d)	(None, 14, 14, 64)	18496
cov2d_3 (Conv2d)	(None, 14, 14, 64)	36928
max_pooling2d_1(MaxPooling2D)	(None, 7, 7, 64)	0
dropout (Dropout)	(None, 7, 7, 64)	0
cov2d_4 (Conv2d)	(None, 7, 7, 96)	55392
cov2d_5 (Conv2d)	(None, 7, 7, 96)	83040
max_pooling2d_2 (MaxPooling2D)	(None, 3, 3, 96)	0
Dropout_1 (Dropout)	(None, 3, 3, 96)	0
cov2d_6 (Conv2d)	(None, 3, 3, 128)	110720
max_pooling2d_3 (MaxPooling2D)	(None, 1, 1, 128)	0
Flatten (flatten)	(None, 128)	0
dense (Dense)	(None, 256)	33024
Dropout_2 (Dropout)	(None, 256)	0
Dense_1 (Dense)	(None, 128)	32896
Dropout_3 (Dropout)	(None, 128)	0
Dense_2 (Dense)	(None, 26)	3354
Total params: Trainable params: Non-trainable params:		383,418 383,418 0

In this study, we conducted three experiments. As we have seen in Fig. 8 and Fig. 9, the convolutional neural network model is validated using dataset_1 and generates a 27% loss and 95.5% accuracy at the iteration of 100.

From using dataset_3, we validated the recognition model and generated the result of a 16% loss and approximately 97% accuracy. The generated results are represented in Fig. 10 and Fig. 11.

In the proposed recognition model, some confusion has happened to some characters. Here we tried to extract mostly confused characters from predicted and actual values. For example, as we recorded in

Table VI, Fidel's get confused more than two times are "ሀ", "ሐ", "ሠ", "በ", "ወ", "ፀ", and "ዐ". This confusion is happening due to the similarity of shapes with other characters.

Table VI. Confusion matrix for eleven selected base characters

	Predicted classes
Actual classes		ሀ	ሐ	ሠ	በ	ወ	ፀ	ፈ	ጠ	መ	ሰ	ዐ
	ሀ	69	0	0	0	2	0	2	0	0	0	1
	ሐ	0	67	0	0	1	0	0	4	1	1	0
	ሠ	1	0	66	0	0	0	0	0	2	0	0
	በ	0	0	0	78	0	0	0	0	0	2	0
	ወ	0	0	2	0	64	0	0	1	0	0	1
	ፀ	1	0	0	0	0	67	0	0	0	0	2
	ፈ	0	0	0	0	0	0	59	0	0	0	0
	ጠ	0	0	0	0	0	0	0	65	1	0	0
	መ	0	0	0	0	0	0	0	0	65	0	0
	ሰ	0	0	0	0	0	0	0	0	0	78	0
	ዐ	0	0	0	2	0	1	0	0	0	1	74

4.1 Result comparison with different methods

To further demonstrate the performance of this proposed model, we compared it with the other approaches for their effectiveness in the area. In general, our method outperforms in recognition accuracy. Table VII shows the summarized experimental results and methods of different studies on Ethiopic scripts.

Table VII. Performance comparison of related works

Techniques	Ref.	Results (Accuracy)
Similarity pattern matching, ANN	[14]	91%
HOG, LBP, SVM	[33]	92.80%
CNN	[34]	92.71%
ANN	[35]	89.88%
PCA, LDA, SVM	[36]	90%
LSTM	[37]	94.68%
The proposed method		97%

In this study, we developed the CNN based character recognition model for Ethiopic Script, and it is validated with three sample images of data, which is prepared by ourselves for the training and valuation process. The dataset is divided into 80% and 20% for training and validation, respectively. We achieved approximately 97% accuracy, which is a promising result as compared to the results of the studies done.

We have a plan to extend the efforts on handwritten documents to solve problems of handwritten variations and different pen colors in the future. We will also account for our dataset to include all the letters, punctuation marks, and symbols found in Ethiopic scripts for both printed and handwritten and increasing the data set size for different deep learning algorithms to get better results.

Acknowledgment

We thank our colleagues from Delhi Technological University and Debre Tabor University who provided insight and expertise that greatly assisted the research.

Competing Interests: Authors have no need for financial or non-financial interests that are directly or indirectly related to this paper.

Chernyshova, Y.S., Sheshkus, A.V., and Arlazarov, V.V.: ‘Two-Step CNN Framework for Text Line Recognition in Camera-Captured Images’, IEEE Access, 2020, 8, pp. 32587–32600
Hidayat, A., Nurtanio, I., and Tahir, Z.: ‘Segmentation and Recognition of Handwritten Lontara Characters Using Convolutional Neural Network’, in Editor (Ed.)^(Eds.): ‘Book Segmentation and Recognition of Handwritten Lontara Characters Using Convolutional Neural Network’ (2019, edn.), pp. 157–161
Alizadehashraf, B., and Roohi, S.: ‘Persian handwritten character recognition using convolutional neural network’, in Editor (Ed.)^(Eds.): ‘Book Persian handwritten character recognition using convolutional neural network’ (2017, edn.), pp. 247–251
Ahamed, P., Kundu, S., Khan, T., Bhateja, V., Sarkar, R., and Mollah, A.F.: ‘Handwritten Arabic numerals recognition using convolutional neural network’, J Amb Intel Hum Comp, 2020
Weldegebriel, H.T., Liu, H., Haq, A.U., Bugingo, E., and Zhang, D.: ‘A New Hybrid Convolutional Neural Network and eXtreme Gradient Boosting Classifier for Recognizing Handwritten Ethiopian Characters’, IEEE Access, 2020, 8, pp. 17804–17818
Reta, B.Y., Rana, D., and Bhalerao, G.V.: ‘Amharic Handwritten Character Recognition Using Combined Features and Support Vector Machine’, in Editor (Ed.)^(Eds.): ‘Book Amharic Handwritten Character Recognition Using Combined Features and Support Vector Machine’ (2018, edn.), pp. 265–270
Alemu, W., and Fuchs, S.: ‘Handwritten Amharic Bank Check Recognition Using Hidden Markov Random Field’, in Editor (Ed.)^(Eds.): ‘Book Handwritten Amharic Bank Check Recognition Using Hidden Markov Random Field’ (2003, edn.), pp. 28–28
Assabie, Y., and Bigun, J.: ‘HMM-Based Handwritten Amharic Word Recognition with Feature Concatenation’, in Editor (Ed.)^(Eds.): ‘Book HMM-Based Handwritten Amharic Word Recognition with Feature Concatenation’ (2009, edn.), pp. 961–965
Belay, B.H.H., T. A.; Stricker, D.: ‘Amharic Character Image Recognition’, in Editor (Ed.)^(Eds.): ‘Book Amharic Character Image Recognition’ (2018, edn.), pp. 1179–1182
Addis, D., Liu, C., and Ta, V.: ‘Printed Ethiopic Script Recognition by Using LSTM Networks’, in Editor (Ed.)^(Eds.): ‘Book Printed Ethiopic Script Recognition by Using LSTM Networks’ (2018, edn.), pp. 1–6
Assabie, Y., and Bigun, J.: ‘Ethiopic Character Recognition Using Direction Field Tensor’, in Editor (Ed.)^(Eds.): ‘Book Ethiopic Character Recognition Using Direction Field Tensor’ (2006, edn.), pp. 284–287
Assabie, Y., and Bigun, J.: ‘Multifont size-resilient recognition system for Ethiopic script’, Int J Doc Anal Recog, 2007, 10, (2), pp. 85–100
Weldegebriel, H.T., Chen, J., and Zhang, D.: ‘Deep learning for Ethiopian Ge'ez script optical character recognision’, in Editor (Ed.)^(Eds.): ‘Book Deep learning for Ethiopian Ge'ez script optical character recognision’ (2018, edn.), pp. 540–545
Assabie, Y., and Bigun, J.: ‘A Hybrid System for Robust Recognition of Ethiopic Script’, in Editor (Ed.)^(Eds.): ‘Book A Hybrid System for Robust Recognition of Ethiopic Script’ (2007, edn.), pp. 556–560
Belay, B.H., Habtegebirial, T., Liwicki, M., Belay, G., and Stricker, D.: ‘Amharic Text Image Recognition: Database, Algorithm, and Analysis’, in Editor (Ed.)^(Eds.): ‘Book Amharic Text Image Recognition: Database, Algorithm, and Analysis’ (2019, edn.), pp. 1268–1273
Fitsum, K.T., and Patel, Y.: ‘Optical Character Recogniation for Tigrigna Printed Documents Using HOG and SVM’, in Editor (Ed.)^(Eds.): ‘Book Optical Character Recogniation for Tigrigna Printed Documents Using HOG and SVM’ (2018, edn.), pp. 1489–1494
Assabie, Y., and Bigun, J.: ‘Lexicon-based offline recognition of Amharic words in unconstrained handwritten text’, in Editor (Ed.)^(Eds.): ‘Book Lexicon-based offline recognition of Amharic words in unconstrained handwritten text’ (2008, edn.), pp. 1–4
Alrehali, B., Alsaedi, N., Alahmadi, H., and Abid, N.: ‘Historical Arabic Manuscripts Text Recognition Using Convolutional Neural Network’, in Editor (Ed.)^(Eds.): ‘Book Historical Arabic Manuscripts Text Recognition Using Convolutional Neural Network’ (2020, edn.), pp. 37–42
Aly, S., and Mohamed, A.: ‘Unknown-Length Handwritten Numeral String Recognition Using Cascade of PCA-SVMNet Classifiers’, IEEE Access, 2019, 7, pp. 52024–52034
Asha, K., and Krishnappa, H.K.: ‘Kannada Handwritten Document Recognition using Convolutional Neural Network’, in Editor (Ed.)^(Eds.): ‘Book Kannada Handwritten Document Recognition using Convolutional Neural Network’ (2018, edn.), pp. 299–301
Aneja, N., and Aneja, S.: ‘Transfer Learning using CNN for Handwritten Devanagari Character Recognition’, in Editor (Ed.)^(Eds.): ‘Book Transfer Learning using CNN for Handwritten Devanagari Character Recognition’ (2019, edn.), pp. 293–296
Purkaystha, B., Datta, T., and Islam, M.S.: ‘Bengali handwritten character recognition using deep convolutional neural network’, in Editor (Ed.)^(Eds.): ‘Book Bengali handwritten character recognition using deep convolutional neural network’ (2017, edn.), pp. 1–5
Ahmed, S.B., Hameed, I.A., Naz, S., Razzak, M.I., and Yusof, R.: ‘Evaluation of Handwritten Urdu Text by Integration of MNIST Dataset Learning Experience’, IEEE Access, 2019, 7, pp. 153566–153578
Nair, P.P., James, A., and Saravanan, C.: ‘Malayalam handwritten character recognition using convolutional neural network’, in Editor (Ed.)^(Eds.): ‘Book Malayalam handwritten character recognition using convolutional neural network’ (2017, edn.), pp. 278–281
Liu, Y.J., Huang, H., Cao, J.D., and Huang, T.W.: ‘Convolutional neural networks-based intelligent recognition of Chinese license plates’, Soft Comput, 2018, 22, (7), pp. 2403–2419
Kim, I.J., and Xie, X.H.: ‘Handwritten Hangul recognition using deep convolutional neural networks’, Int J Doc Anal Recog, 2015, 18, (1), pp. 1–13
Gerber, C., and Chung, M.: ‘Number Plate Detection with a Multi-Convolutional Neural Network Approach with Optical Character Recognition for Mobile Devices’, J Inf Process Syst, 2016, 12, (1), pp. 100–108
Gurucharan, M.K.: ‘Basic CNN Architecture: Explaining 5 Layers of Convolutional Neural Network’, in Editor (Ed.)^(Eds.): ‘Book Basic CNN Architecture: Explaining 5 Layers of Convolutional Neural Network’ (upGrad, 2020, edn.), pp.
Banharnsakun, A.: ‘Towards improving the convolutional neural networks for deep learning using the distributed artificial bee colony method’, Int J Mach Learn Cyb, 2019, 10, (6), pp. 1301–1311
Jain, B., Singh, P., and Goel, R.: ‘Handwritten Marathi Character Image Recognition using Convolutional Neural Network’, in Editor (Ed.)^(Eds.): ‘Book Handwritten Marathi Character Image Recognition using Convolutional Neural Network’ (2019, edn.), pp. 326–329
Zhang, Z.: ‘Derivation of Backpropagation in Convolutional Neural Network (CNN)’, University of Tennessee, Knoxvill, TN, 2016
Saha, C., Faisal, R.H., and Rahman, M.M.: ‘Bangla Handwritten Basic Character Recognition Using Deep Convolutional Neural Network’, in Editor (Ed.)^(Eds.): ‘Book Bangla Handwritten Basic Character Recognition Using Deep Convolutional Neural Network’ (2019, edn.), pp. 190–195
Reta, Y.B., and Rana, D.: ‘Amharic Handwritten Character Recognition Using Combined Features and Support Vector Machine’, in Editor (Ed.)^(Eds.): ‘Book Amharic Handwritten Character Recognition Using Combined Features and Support Vector Machine’ (IEEE, 2018, edn.), pp. 265–270
Belay, B.H., Habtegebrial, T.A., and Stricker, D.: ‘Amharic Character Image Recognition’, 2018 Ieee 18th International Conference on Communication Technology (Icct), 2018, pp. 1179–1182
Beyene, E.G.: ‘Handwritten and Machine printed OCR for Geez Numbers Using Artificial Neural Network’2019 pp. Pages
Meshesha, M., and Jawahar, C.V.: ‘Optical Character Recognition of Amharic Documents’, African Journal of Information and Communication Technology, 2007, Volume.3, pp. 53–66
Malhotra, R., and Addis, M.T.: ‘Ethiopic Base Characters Image Recognition using LSTM’, in Editor (Ed.)^(Eds.): ‘Book Ethiopic Base Characters Image Recognition using LSTM’ (IEEE, 2021, edn.), pp. 94–98

Download PDF

Version 1

posted

You are reading this older preprint version

Read the latest preprint version →

Ethiopic Characters Recognition Model using Deep Learning

Status:

Version 1

Abstract

Figures

1 Introduction

2 Literature Review

Ethiopic Script Studies

Other script studies

3 Methods

3.1 Data Sources

3.2 Deep Learning Algorithm and Description

3.2.1 Convolutional Neural Network Layers

3.3 The proposed model

3.4 Training CNN Model

3.5 Experimental Tools

4 Results And Discussion

4.1 Result comparison with different methods

5 Conclusion And Future Works

Declarations

Acknowledgment

References

Status:

Version 1