The Handwritten Character Recognition (HWCR) system is a tool that recognizes handwritten characters in papers, photographs, and other sources and converts them into editable computer characters for future use. A significant barrier still exists in the accurate recognition of complex formed compound handwritten characters. By learning discriminatory qualities from vast volumes of raw data, convolutional neural network (CNN) technology has recently made significant advancements in HWCR systems. In identification of the Vernacular (Twi) language characters from off-line training and testing datasets, convolutional neural network was applied in this study. The primary goal of this research is to examine the convolutional neural network's ability to identify Twi characters from the supplied image datasets and the identification accuracy after training and testing. In order to differentiate among characters, convolutional neural network evaluates their differences in styles, shapes and topographies. To determine the accuracy of handwritten characters, the convolutional neural network implementation is tested using the novel Twi language dataset. On the basis of the test findings, a dataset set of 13, 200 images from the Twi dataset achieved an accuracy of 88.15% during training and a test accuracy of 79.31%. This study aids in knowing the suitable architecture that can be fused into an optical character recognition device to recognize Asante Twi handwritten letters found in documents of this particular language and translate them into editable formats.
Research Article
Handwritten Character Recognition Using Convolutional Neural Network for Asante Twi Language Alphabets
https://doi.org/10.21203/rs.3.rs-3638464/v1
This work is licensed under a CC BY 4.0 License
Version 1
posted
You are reading this latest preprint version
Convolutional Neural Network
Handwritten Character Recognition
Asante Twi Handwritten Character Recognition
Deep Learning
Simply click print, and you will obtain an actual sheet of a paper with the characters or words you were looking at on your personal computer’s screen in just some few seconds. However, going to the reverse way, scanning documents into your personal computer is usually much difficult or better say challenging because contemporary scanners tend to take basically a picture of the document and save them as a Joint Photographic Experts Group popularly known as jpeg or jpg or Portable Network Graphic also known as .png format type unto your computer, here not only will you probably not have a little dust or noise due to the data compression but a fresh replica of your manuscript cannot be edited in your preferred word processing software because the scanners are unable to read every character singularly. Over the years, the popular MNIST Cohen et al. (2017) developed a dataset and trained lots of handwritten characters in the English language and other researchers also trained in different languages such as Arabia, India scripts etc. According to Wibowo et al. (2017), with the advanced-tech Convolutional Neural Network (CNN) which has shown beyond doubt to be the most advanced machine learning technique when it comes to pattern recognition and computer vision to recognize hand drawn characters of humans more accurately than other methods. Consequently, in a nation like Ghana where several different languages are spoken and having the Twi language which is popularly known to be a dialect in the Akan ethnic group and being widely spoken in the country than all other languages according to Azunre et al. (2021) because of the population of the natives and how widely they have spread across the nation there is a need to store their historic background information and other documented information, and converting all these information into a digitize data which by hand will cause lots of complications so there is a need to get an algorithm to be implemented in an Handwritten Character Recognition (HWCR) system that would be trained on the Twi language characters to foster the conversion of handwritten characters into digitized formats. The Twi characters are twenty-two (22) letters and they are hail from the Latin scripts, fifteen (15) of these characters are consonants thus b, d, f, g, h, k, l, m, n, p, r, s, t, w, y and seven (7) of them are vowels, thus a, e, ɛ, i, o, ɔ, u and this is shown in Fig. 1 (Awiba, 2016).
The keyboard is the typical method of entering data into a computer. However, this isn't necessarily the most effective or efficient method and thereby employing OCRs. The Optical Character Recognition (OCR) is a method for converting optical patterns in digital images which could have been printed or handwritten that match to alphanumeric or other characters (Anyline, 2017). Segmentation, feature extraction, and classification are all critical facets in character recognition (Arindam et al., 2017; MathWorks, 2022). Machine reading or computer vision has evolved from a fantasy to a reality in recent decades OCR (Optical Character Recognition) technology due to its advancement to a sophisticated and dependable level. These systems have proven to be among the most prosperous uses of computing technology in the domains of pattern recognition and artificial intelligence (Caspi, 2022).
OCR technology allows us to transcribe a variety of documents, including printed paper files, pdf files, and digital photography photos, into modifiable and searchable data. There is an automated identification, this refers to the capability of the recognition system to recognize items automatically, congregate information about them, and input that information into the computer systems without requiring the intervention of humans (Techtarget, 2022).
In the field of Optical Character Recognition system, identification or recognizing various printed text is easy since there is a collective database to match various characters but in the advent of handwritten or handwriting characters, these recognition systems find it difficult to recognize these characters since there is a vast range of hand writing styles from different human beings (Chomba, 2018). So, this has brought to the attention of some systems known as Intelligent Character Recognition systems, which is basically an enhanced form of an OCR, which is able to identify handwriting characters due to it possessing sophisticated electronic neural system that allows the program to automatically update the database and grasp any different breed of font or handwriting styles, improving accuracy and character identification (Taylor, 2022).
In this very review, Ram et al. (2018) following the overall architecture of convolutional neural network, the researchers employed a Devanagari character dataset from UCI machine learning store, they preprocessed the data and they did what they called a one-hot coding before serving it to the CNN. Laying aside the convolutional layers, sub-sampling (thus, the max-pooling was used) and dropout layers were employed to the network’s architecture to improve upon the efficacy of the model. Computing the model, dropout layers of two in number were used to preclude over-fitting; with each layer followed by an activation function. And these were continued by two (2) fully connected layers.
Roy (2020) also proposed an architecture for a Bengali handwritten character recognition system using a convolutional neural network called the AKHCRNet. The researcher obtained a dataset that has been made readily available online with a total of 166,000 Bengali characters, which contained compound characters, numerical digits as well as alphabets that produced 96.8% accuracy out of eleven (11) epochs after it went through the AKHCRNet model. After obtaining the dataset, a preprocessing was made, by using gray-scale image as input shape since no one really cares about the color of the character, he further resized the image to 32x32 pixels for both width and height respectively using the bi-linear interpolation approach. Therefore, the data was presented as (32, 32) gray-scale image to the network, then included (32, 32, 1) layer as input, he used (5, 5) convolutions and 32 filters. He then employed the ReLU activation function. Next, he used the batch normalization and then employed the max-pooling layer. Following these, he patched up the network with four (4) layers which are fully connected and with an output layer of 84 classes. The final output layer uses the softmax activation to yield desired result and L2 Regularization method to ward off any out-fitting issues.
Due to the magnificent ability of the novel convolutional neural network for handwritten character recognition, (Demilew and Sekeroglu, 2019) did not spare to use the model in identification and recognition of ancient Geez characters of Ethiopia. Researchers of this article developed a model which houses the following steps and stages in the character identification; this included the preprocessing stage, the digitized image segmentation and a good feature extraction model within the convolutional network. As the first step, pictures of the characters were gathered, and applied the techniques to get the best and quality images for the next stage. At this level, grayscale conversion was applied to decrease the intensity dimensions of every image, due to the characters taken being of old, the authors applied de-noising for better recognition. Following this step is the vital binarization and skew detection stage to easily segment the documented images of which the Otsu approach was used. When this was done the Hough transform was applied to the binarized characters and this method is skewing, skewing helps to redefine the irregular angle of images during the characters’ elicitation. The scanned pictures went through another crucial step, thus segmentation. This employed a contour analysis and text line segmentation with each image being clipped and stockpiled in a sorted style. When the segmentation is done, the isolated characters are channeled to the feature extraction state of convolution. Here, the authors of the model used three (3) convolutional layers for feature extraction and two (2) layers for a fully connected layer also known as the dense layer for the purpose of classification. The network used filter kernels, some activation functions (thus ReLu) and a softmax layer, a dropout of 0.2, the max-pooling technique and a dense layer specification. The output results of the second convolutional network are passed and then they are being flattened.
Muppalaneni (2020) also pounced on the usage of the convolutional neural network to train and predict rightfully the handwritten characters of the Telugu characters (thus a language spoken by Indians). Here, the suggested model’s architecture bares five (5) convo-layers and five (5) maximum-pooling layers. The designed architecture used a 50 by 50 by 256 for the first layer, with a max-pooling which employed an activation function using the ReLU, the same operations of the first layer were made in the other four remaining convolutional layers. The output of one layer was transferred unto the other. From here, a flatten layer was used which also used a dropout technique to prevent overfitting and finally the fully linked or connected layer was used for the classification process.
Another group of researchers, thus (Husnain et al., 2019) also brought their network’s architecture on the handwritten character recognition of the Urdu language. This language is more spoken by the people living in the South-East regions of Asia. Husnain et al, also presented their model in the handwritten script or character recognition utilizing the innovative feature of the ConvNet. From their proposed technique, after acquiring the image dataset, the preprocessing facet is done and this includes the removal of noise and then translates the various images into a gray-scale and clipped the image into a 28x28 pixels. The feature extraction facet is the next phase of the model, to extract some characteristics of the character such as the height, width, aspect ratio of the text image etc. And again, images with size of 28x28 pixels, are being fetched into the initial layer which is a 2-dimensional convolution layer accompanied with a 5x5 kernel size. The end result of this layer will be implanted with a feature map of 28x28 map. Every convolutional layer is propagated using the rectified linear unit activation function. As a final point, the dense-layer at the end part of the model classifies the specified input.
According to Amin et al. (2020) proposed model, they first generated their own dataset since there was none, by making three hundred natives of Pashto language to write the various characters and named the dataset called Poha, meaning learning. The dataset bared 26,400 photos of which it contains both letters and numbers. Each character was translated into a 28x28 pixel of size. This character-set was furthered into preprocessing to enhance the images quality for easy identification, noise removal from images was done using Gaussian blur and were smoothened based on the Gaussian function technique. Finally, the images were gray-scaled to now move into the next facet of the model. The images are then sent as inputs to the 1st convolutional layer which is 2-dimenstional with a kernel of 3 by 3, a ReLU was utilized as the activation function at this level and the input data were fed into the 1st maximum-pooling layer, the output of this layer submitted to the 2nd convolutional layer (2-dimentional) also with a kernel of 3x3 size. Again, the results of the layer are submitted into the 2nd max-pooling layer and then the output results are then submitted into the 3rd convolutional layer (2-dimensional) and then the output results are passed into the flattening layer for a vector of 1-dimensional. At the end, the two (2) fully-connected layers were employed; there contained a softmax layer in the final dense layer for the purpose of classification and prediction.
3.1.1 Dataset Creation
The creation of a Twi dataset is paramount in this research as it is the beginning of the entire process, datasets creation are the basis for every machine learning project (Lu et al., 2021) This dataset as described will harbor the twenty-two characters (22) within the Twi language in which these characters have uppercase and lowercase making the overall characters to bear forty-four (44) distinctive characters thus – uppercase letters: A, B, D, E, Ɛ, F. G, H, I, K, L, M, N, O, Ɔ, P, R, S, T, U, W, Y and the lowercase letters: a, b, d, e, ɛ, f, g, h, i, k, l, m, n, o, ɔ, p, r, s, t, u, w, y. this is done to enable the handwritten recognition system distinctively identify each character when presented to in the dataset. Most people tend to use secondary dataset by going to various websites to collect them, but in this project, a dataset by far has not been created, so this research will tend to use a primary data; some of the character’s elicitation will be done in the various educational levels since age and literacy ranks because these sometimes influence handwriting skills. These category of persons will be made to write the forty-four (44) characters of the Twi alphabets on a clean sheet of paper, to balance handwritten skills, a sample size of three hundred (300) people will be considered and these will yield a total of thirteen-thousand and two-hundred (13, 200) characters to be available, this size will be considered since deep learning projects are good at making good results when a dataset contains enough amount of data (Christin et al., 2019). These characters will then be digitized through the use of quality cameras and scanners to obtain quality digital images of these distinct alphabets and grouping the like terms, hence creating an offline dataset to train and test the suggested model.
3.1.2 Image Data Pre-processing
Pre-processing's purpose is to arrange or organize data so that the character identification system may explicitly recognize various characters easily and finish faster. Some people have speed in writing so they may skip some strokes or curves and others too may somehow hesitate in the process of writing therefore at this stage it is paramount to eliminate all undesired data, when this stage is ignored, the neural network may be trained to learn wrong patterns as well which won’t be appropriate. Once the data is being elicited, the next step is to refine our dataset through the use of following operations; checking up on the hand drawn characters to see if the actual feature of various characters are not compromised; this will be one of the operations. Normalization process will be applied to normalize the images. Here, in terms of size and pixel values, captured images are placed into a same statistical distribution hence resizing the images so that they will have a consistent size for training and testing, the proposed size will be 32 by 32 for the network. From this point, a binarization operation will be done. Binarization is explicitly the process of converting an input image into a grayscale image thus white and black (1 and 0). Due to background noise, the images need to be translated so that the main image will be focused on since colors in images do not matter in recognizing a letter in a language. From this facet, the images will be ready for training. After the preprocessing facet, the dataset will be split into two (2) datasets, these are, training dataset and testing dataset. Eighty (80) percent of the images in the dataset will be used for training whilst the remaining twenty (20) percent of images was used for analyzing and testing the neural network’s performance.
3.1.3 Convolutional Neural Network Model for Twi Handwritten Characters
These preprocessed images are then ushered in the neural network’s model for the training session to begin.
3.1.3.1 The Convolutional Layer
The convolutional layer uses a kernel for its feature extraction, the kernel is nothing but a filter in a matrix form that moves in strides or slides on the input image and performs a dot product with the sub-regions and produces an output dot product. From top to bottom, the filter is moved across the image from left to right, changing one pixel in each horizontal direction and one pixel in each vertical direction. The term "stride" refers to the amount of movement between applying a filter to the input image. Its height and width dimensions are nearly symmetric. The default stride is in one-dimension thus (1, 1), for the height and width. The output results may be sent to a number of convolutional layers for further processing. After a convolution is performed on an image, there will be a need for non-linearity for the feature extraction. In effect, an activation function is required to perform this operation, there are a number of activation functions and these are sigmoid, tanh and ReLU; by far, the Rectified Linear Unit (ReLU) has proven to be the best in image feature extractions. The purpose of using the ReLU in the neural network is to introduce non-linearity into the network which has been computing linear operations during convolutional operations. Since images are non-linear in nature, it is paramount to escalate the non-linearity within the neural network. The rectifier function removes all the dark parts from an image, leaving those with a positive value, thus gray and white colors. It would be faster to train a network when ReLU is employed, hence, it shall be the activation function to be used on the convolutional layer after each convolutional operation. For each convolutional layer there sometimes exists a pooling layer after it. Therefore, after a convolution the output results (which is a feature map) is sent to a pooling layer. Three (3) convolutional layers will be employed for this model and of course each convo-layer will use the ReLU as the activation function.
3.1.3.2 The Pooling layer
In CNN, a pooling is required to down sample the detection of features in a feature map. The number of pooling layers provides an algorithm to down sample (thus, decreasing the number of pixels from the output of the preceding convo-layer, which reduces the dimensionality of images) feature maps by summarizing the presence of feature in patches of the feature map. There exists two most commonly used pooling techniques, and these are the max-pooling (Maximum Pooling) and average pooling. As the average-pooling summarizes the average presence of a feature, the max-pooling summarizes the most activated presence of a feature. In doing so, in this study the maximum-pooling will be used to downsize the image and still maintain the important feature within the image. This pooling helps in reducing computational load of the network. Therefore, three (3) max-pooling layers shall be employed for this model.
3.1.3.3 Fully-Connected Layer or Dense Layer
After the various computations of the convo-layers and the maximum-pooling layers, the output results are then transited to the fully-connected layer also known as the dense layer. Within the proposed CNN architecture, the full-connected layers shall be employed to classify the images to a particular category. Before these images are finally transmitted, they shall be flattened into 1-dimensional vectors. The flattened layered array will be used as the input to the fully-linked layers. It is called fully-linked because every single neuron of that layered component is connected to all the neurons in the previous layer and to the next fully connected layer. In this research, three (3) fully-connected or dense layers to tighten the classification processes for the proposed architecture. The output of fully-connected layers will have the identical quantity of neurons as the number of categories used in this project. The final fully-connected layer shall employ the use of the softmax activation function to classify the respective images into a particular category and accurately predict the characters. It's mostly used to fit the overall output of neural network models between zero and one.
4.1.1 Data Gathering and Dataset Creation
As proposed in the previous chapter, a sheet of A4 size which bore printed Twi characters with spaces provided to enable the writers to write the various corresponding characters in their own writing styles was presented as shown in Figure 2; and these writers mounted to three-hundred. Since, there are forty-four (44) characters thus both capital letters and small letters a total of 13200 characters were obtained. These 300 sheets were scanned in a grayscale format for onward processing. Various characters were then cropped out of each scanned paper to obtain images in a jpeg format for further processing. After successfully cropping out all the images, a software tool called Light Image Resizer v6.0.6.0 was used to resize all the images into a uniform size of 32 x 32 pixels. Other effects were applied to enhance these images before training. The colors of each image were inverted, added auto-enhancement and also adjusted the brightness of the image to a 100 percent.
After a successful effect on the image, this is the outcome of sampled images as shown is Figure 3.
This was done to generate a dataset to be ready for the neural network. Then the entire dataset was split into two (2) major folders, thus training and testing. All letters of the same caliber were placed in a particular folder, for example, all cropped handwritten image characters containing capital “A” were put into a common folder and this was the same for all other characters. This gave a total of 44 folders, hence giving us a total of 44 classes (thus, 22 upper case letters and 22 lower case classes). Now, each subfolder in the training folder contained two hundred and forty (240) images as well as each subfolder in the testing folder contained sixty (60) images. This made the training and testing folders to bear a total of ten thousand, five hundred and sixty (10,560) images and two thousand, six hundred and forty (2,640) images respectively. From here, the dataset is ready to undergo training and testing. The characters in the training dataset are different from that in the testing and this is done to help analyze the training percentages of trained images and tested images. The aids to see if neural can rightfully find similarities in classified folders based on percentages in accuracies and losses to make good predictions. Note, the higher the percentage in accuracy of the training and testing datasets the better the predictions.
4.2 System Simulation
During the implementation session the following tools were installed onto the machine to compute the model and the programming language used is the python to carry the operation. These are:
i. Anaconda Integrated Development Environment: this contains packages like the jupyter notebook which serves as a platform for training of datasets.
ii. Jupyter Notebook: this package contains lots of libraries for performing deep learning model implementation of which we made use of the Tensorflow, Numpy, PIL (Python Image Library) and OpenCV2 libraries.
Step 1: Importing the tensorflow libraries as shown in Figure 4.
Step 2: Creating CNN’s architecture
The neural network contained the following parameters as shown in Figure 5, it harbored three (3) convolutional layers:
First Convolutional Layer: the initial convolutional layer of 32 neurons and an image’s input shape of size 32x32 pixels and a 5x5 kernel for the feature extraction. It utilized the ReLU as its activation function and used the max-pooling layer of 2x2 and the results of this was transmitted unto the second convolutional layer.
Second Convolutional Layer: it has a layer of 64 neurons and a kernel size of 3x3, which also applied the ReLU activation function and the maximum-pooling layer of 2x2. The computational results of this were then sent to the third convolutional layer.
Third Convolutional layer: this also has 128 neurons, and a kernel of size 3x3 and still maintained the ReLU activation function and used a max-pooling layer of 2x2. Then the convolutional results are then sent to the fully-connected layer 1.
Fully-Connected Layer 1: the network flattened the output into a one-dimensional array and used a dense layer of 128 and also applied the ReLU activation function.
Fully-Connected Layer 2: this layer maintained the fully-dense layer of 128 neurons, and the ReLU activation function with a dropout 0.5.
Fully-Connected Layer 3: this is the final year for the prediction of the various characters which bore dense layer of 44 neurons due the 44 classes we generated from our dataset and it is a rule that the final dense layer size must match the number of classes in a multi-categorical problem. This layer used the softmax activation function for the prediction.
Step 3: Compile the CNN model
After defining the model and stacking the layers, there is a need to configure it. During the compilation step, it is necessary to perform this configuration process. Before training, the model must be built, with the loss function, optimizers, and prediction metrics defined. CNN's architecture used the Adam optimizer, the loss argument used the categorical cross entropy and the metrics used to measure the accuracy, since we are interested in the accuracy levels of both training and testing datasets during training.
Step 4: Compile the CNN model
Step 5: Training dataset and testing dataset performances
The Figure 8 shows the accuracies and losses for both datasets, the first two (2) columns indicate the losses (loss) and accuracies (accuracy) for the training dataset respectively and the last two (2) shows the losses (val_loss) and accuracies (val_accuracy) for the testing dataset respectively with just 15 epochs. The CNN model achieved an average of 88.15% for the training and obtained 79.31% for the testing. This shows that our model is doing well with predictions due to high percentages for both training and testing datasets.
Step 6: Saving the Model
The trained model and its performance is saved for further testing of new characters. These characters were not a part of the trained and tested datasets. So, a json file was created to harbor the saved model as shown in the Figure 9.
Step 7: Creating a graphical user interface to draw a character for prediction
The Figure 10 shows a code for creating an interface that serves as a platform to draw any preferred letter to test the CNN model to see if it can rightfully predict new hand drawn characters which are not a part of neither the trained dataset nor the tested dataset. Setting the parameters for the interface and calling the paint interface to allow drawing of the characters.
Figure 11 shows a hand drawn letters in Twi, ready to be learned by the CNN model to predict the letters respectively. This figure shows the target size of the drawn image which is a jpeg format and it’s being saved into a folder called “Singleprediction” and then the model will pick the image saved, analyze it based on learnt or trained patterns and then predicts the character.
Figure 12 shows a successful prediction of the ɔ and Ɛ letters respectively.
4.3 Discussion of Results
The convolutional neural network architecture achieved an average accuracy of 88.15% on a dataset with few gray-scaled images, which is 13,200. These images were of size 32 by 32, height and width respectively. The network architecture harbored three (3) convo-layers for feature extraction alongside with the ReLU activation function and three (3) fully-connected layers in which one of them is a flatten layer and final fully connected layer had a softmax activation function for prediction. As shown in figure 8 there was a steady increment from one epoch to the other with just 15 epochs during training. The training accuracies and validation accuracies have a close gap in each step of the training periods. In two separate datasets, tested the accuracy of the proposed model (see Figure 5, the neural network used to classify Twi handwritten letters). To classify the Twi handwritten numbers, 80 percent of the 13,200 data points were used for training and 20 percent for testing. It took into account the rate of learning, quantity of hidden neurons, and a batch size of 32 when training the convolutional neural network with three (3) convolutional layers for the tests and three (3) fully connected layers. The results showed that raising the network scale increased the efficiency of the convolutional neural network, but there was one big drawback: over-fitting due to the prolonged training period and large gaps between training and testing accuracies. On the other hand, by adjusting the batch size, it is feasible to achieve the model's ideal state. When the batch size was raised to a specific amount, such as 50, the model cannot be trained due to the rule of thumb. Moreover, the batch size was determined by the amount of memory available. The impact of batch size is depicted in Figure 7. To prevent the problems highlighted above, the trained model was fine-tuned to produce this good outcome.
4.3.1 Model Summary
Each layer in traditional neural networks is made up of a group of neurons. The input to these networks is transformed through a series of hidden layers coupled to previous and subsequent layers by neurons. It was chosen to apply the CNN on the proposed Asante Twi dataset because of its notable and increased performance feature. Below is the model’s summary:
The CNN model was given input images with a size of 32x32 pixels each from the Twi dataset. The model's 1st layer was a 2D convolutional layer with a 5x5 kernel size. Pixels of every input image is used by this convo-layer. A 28x28 feature map was used to implant the result of this layer. To create a feature vector, this feature map was integrated with geometrical features. Every output of the convolutional layer was activated using a rectified linear unit as an activation function. ReLU was chosen because of its capacity to solve the gradient vanishing problem. For stimulation, the ReLU utilizes the intrinsic threshold invariant, which is analogous to the brain of human’s routine. The output results of the initial convo-layer were sent to the maximum-pooling layer bearing a stride of 2 for nonlinear down-sampling. A 14x14 feature vectors were then embedded onto the 2nd convo-layer as a result of this layer's output. On the input, a 3x3 mask of the kernel was applied. This layer's output was routed back into a max-pooling layer for non-linear down-sampling with a stride of 2. The output results of this layer were 6x6 pixels, which was sent into the third 2D convolutional layer. Flattening was used to create a 1D feature vector from the output of these layers. After the convo-layer and maximum-pooling layer, the flattened layer was required to utilize a fully-linked layer. The two completely connected (dense) layers are employed towards the end of the model. With a dense of 44, the final fully-connected layer utilizes a softmax activation function to classify output based on the distribution of probabilities for each class. The Adam optimizer is utilized to optimize adaptive learning as shown in figure 6. The Adam optimizer was chosen for our model because it scales the learning rate with square gradients. During this implementation, the optimizer's default value was used to produce the best results. Various layers, learning rates, optimizers, and datasets of size 32 × 32 were examined throughout the experimentation phase. The finest outcomes, however, were attained using the model provided in Table 1. With fewer parameters, the model performed better, computation and resource usage were minimal.
In a nutshell, the research has been a successful one. With the novel dataset obtained and the architecture of the neural network developed showed a very good performing accuracy for handwritten characters in Twi using the convolutional neural network for both training and testing datasets, obtaining 88.15% and 79.31% respectively. The model’s architecture provides the basis that can help train more images which can be fused into optical character recognition devices for high level recognition of handwritten Twi characters in documents into editable formats.
This research can be improved by gathering more images of handwritten Twi letters, including punctuations and numbers in this same language to have a complete dataset to train and improve the performances of accuracies obtained. Enough time should be allocated during thesis writing so that more data can be collected for experiments. Noting, this will give a researcher a good standing to create word character recognition system for an optical character recognition (OCR) device for translating handwritten Twi words to editable formats with ease. To also provide the basis for the design of a powerful convolutional architecture for an image feature extraction and classification a research can be made to know the right number of parameters (the right tuning) to be used for any given size of dataset for this language.
1. Originality
We certify that the work being submitted is original and that all information sources used in the study have been properly cited. This work has not been published before and is not currently being considered for publication in any other publication.
2. Authorships’ Contribution
We hereby declare that each of the mentioned authors has significantly contributed to the work's conception, design, acquisition of data, analysis, and interpretation. After reading and approving the final draft of the work, all authors have decided to submit it for publication. Kingsley Drah and Ernest Okoampah aided in the conceptualization, designed and collected data for the project. Robert Sey did the analysis and interpretation whilst Dr. Eric Opoku Osei supervised the entire work.
3. Ethical Approvals
We hereby declare that this study conforms to all applicable institutional rules and ethical standards, including. The relevant review board has given the study ethical approval, and when necessary, participants' informed consent was acquired.
4. Conflicts of Interest
We declare any potential conflicts of interest that may be perceived to influence the research presented in the manuscript. This includes financial, personal, or professional relationships that could be construed as having an impact on the research and its interpretation.
5. Funding
We thus declare that we have received no funding or other forms of support for this research or its publication. The funding sources have been made public, and they have no bearing on the study's conduct, design, analysis, or reporting.
6. Code and Data Availability
We affirm that, upon reasonable request, data corroborating the results of this investigation will be made accessible. If not included in the publication, the raw data and analysis codes can be sent to the journal or other interested parties for replication or additional study.
7. Consent to participate
Not Applicable
8. Consent for Publication
By submitting this work, we acknowledge that, should it be accepted for publication, we consent to transfer the copyright to the Machine Learning Journal. But we still reserve the right to use our work into other projects as long as we give due credit.
We affirm that all information provided in this declaration is accurate and complete.
All authors
November 10, 2023
- Amin, M. S., Yasir, S. M. and Ahn, H. (2020). Recognition of Pashto Handwritten Characters Based on Deep Learning. Sensors, 20, 5884.
- Anyline (2017). What is OCR? An Introduction to Optical Character Recognition. Available: https://anyline.com/news/what-is-ocr [Accessed 29-03-2022].
- Arindam, C., Krupa, M., Pratixa, B. and Ghosh, S. K. (2017). Optical Character Recognition Systems for Different Languages with Soft Computing, Studies in Fuzziness and Soft Computing. 352, 9-41.
- Awiba, S. (2016). The Akan (Twi) Alphabet. Available: https://learnakan.com/akan-alphabet/ [Accessed 13-04-2022]
- Azunre, P., Osei, S., Addo, S., Adu-Gyamfi, L. A., Moore, S., Adabankah, B., Opoku, B., Asare-Nyarko, C., Nyarko, S. and Amoaba, C. (2021). Contextual Text Embeddings for Twi. Ghana, 1-15, Available: https://arxiv.org/pdf/2103.15963.pdf.
- Caspi, G. (2022). The evolution of deep learning and machine learning. Available: https://www.itproportal.com/features/the-evolution-of-deep-learning-and-machine-learning/ [Accessed 29-03-2022]
- Chomba, B. (2018). What is the difference between optical character recognition (OCR) and handwritten character recognition (HWCR)? Available: https://www.quora.com/What-is-the-difference-between-optical-character-recognition-OCR-and-handwritten-character-recognition-HWCR [Accessed 30-03-2022].
- Christin, S., Hervet, É. and Lecomte, N. (2019). Applications for deep learning in ecology. Methods in Ecology and Evolution, 10, 1632-1644.
- Cohen, G., Afshar, S., Tapson, J. and Van S. A. (2017). EMNIST: Extending MNIST to handwritten letters. International joint conference on neural networks (IJCNN), IEEE, 2921-2926.
- Demilew, F. A. and Sekeroglu, B. (2019). Ancient Geez script recognition using deep learning. SN Applied Sciences, 1, 1-7.
- Husnain, M., Saad, M. M., Mumtaz, S., Jhanidr, M. Z., Coustaty, M., Muzzamil Luqman, M., Ogier, J. M. and Sang C. G. (2019). Recognition of Urdu handwritten characters using convolutional neural network. Applied Sciences, 9, 2758.
- Lu, J., Xia, S., Lu, J. and Zhang, Y. (2021). Dataset Construction to Explore Chemical Space with 3D Geometry and Deep Learning. Journal of Chemical Information and Modeling, 61, 1095-1104.
- MathWorks (2022). Text Detection and Recognition. Available: https://in.mathworks .com/help/vision/text-detection-and-recognition.html;jsessionid=3ef258eaf7afabc78495fea6eb39 [Accessed 29-03-2022].
- Muppalaneni, N. B. (2020). Handwritten Telugu compound character prediction using convolutional neural network. International Conference on Emerging Trends in Information Technology and Engineering (ic-ETITE), 2020. IEEE, 1-4.
- Ram, S., Gupta, S. and Agarwal, B. (2018). Devanagri character recognition model using deep convolution neural network. Journal of Statistics and Management Systems, 21, 593-599.
- Roy, A. (2020). AKHCRNet: Bengali handwritten character recognition using deep learning. Bangladesh, 1-12. Available: https://arxiv.org/pdf/2008.12995.pdf. [Accessed 25-05-2022]
- Taylor, K. (2022). OCR vs ICR: What Differentiates the two Character Recognition Software? Available: https://www.hitechnectar.com/blogs/ocr-vs-icr/#:~:text=OCR%20has %20a%20fixed%20and,improves%20accuracy%20and%20character%20recognition [Accessed 31-03-2022].
- Techtarget. (2022). OCR (optical character recognition). Available: https:// www.techtarget. com/searchcontentmanagement/definition/OCR-optical-character-recognition [Accessed 30-03-2022].
- Wibowo, M., Soleh, M., Pradani, W., Nizar, A. H and Arymurthy, A. M. (2017). Handwritten javanese character recognition using descriminative deep learning technique. 2nd International Conferences on Information Technology, Information Systems and Electrical Engineering (ICITISEE), pp. 325-330.
No competing interests reported.