JAVANESE SCRIPT HANACARAKA CHARACTER PREDICTION WITH RESNET-18 ARCHITECTURE

: This study aims to train computers to recognize Javanese script characters known as Hanacaraka. The evaluation was conducted on the use of Convolutional Neural Network (CNN) with the ResNet-18 architecture in recognizing these characters. The research objective is to overcome traditional character recognition barriers and improve accuracy. The method employed includes building a CNN model with the ResNet-18 architecture and using diverse datasets. The results show a training accuracy of 100%, validation accuracy of 98.01%, and accuracy, precision, recall, and F1-score each at 100%. This study concludes that the developed model successfully achieves a high level of accuracy and contributes positively to the development of Javanese Hanacaraka character recognition technology.


INTRODUCTION
The development of computerbased learning media has entered a rapid phase, offering a variety of applications to enhance interactivity and the effec tive ness of the learning process.In this dyna mic movement, digital image processing is a field that has seen significant advan cements, and the implementation of Arti ficial Neural Network (ANN) technology has become a primary choice.Digital ima ge processing offers various possibilities, including character recognition[1], alpha numeric characters[2], hand writing[3], and special characters such as Javanese script Hanacaraka characters.These appli cations leverage technological advancements to develop character reco gnition systems that are increasingly accurate and adaptable to various character formats.
Artificial Neural Networks, as information processing models inspired by the human neural network, prove to be an effective approach for understanding and analyzing image data [4].The use of neural network models to solve specific problems requires sophisticated processes.The main advantage of ANN is its ability to recognize patterns by learning from the trained image data [5].
CNN are neural networks tailored for visual pattern recognition, inspired by how the human brain processes visual data [6].CNN uses convolutional layers to extract local features from input data [7].CNN with ResNet-18 architecture has proven to be an efficient method for pattern recognition tasks in image data [8] [9].The ResNet-18 architecture, a type of CNN, excels in overcoming the problem of vanishing gradients, enabling the training of deeper models [10].However, there are still many challenges regarding the recognition of Javanese Hanacaraka script characters, which hold significant historical and cultural value, as shown in Figure 1.The continuous preservation and development of Java nese script require a character recognition system that is not only effective but also efficient.
Some differences between this res earch and previous studies using CN Ns include the study by A. Budiman et al., who identified Javanese script patterns using the Ngaglena script method and Histogram Chain code.However, the current research focuses on the Javanese scri pt Hanacaraka, known as the classical Jav anese script, which presents a unique cha llenge [11].Muhdalifah et al., who compared pooling methods in the CNN met hod for Javanese script characters [12], and Rasyidi et al. investigated Javanese script characters using the random forest algorithm with 6000 data points, while the current study employs 12000 data points [13].
Therefore, the implementation and development of CNN, especially Res Net-18, are crucial to improving the quality and accuracy of recognizing Javanese Hanacaraka script characters.This research aims to evaluate the effectiveness of the CNN method with ResNet-18 architecture in recognizing Javanese Hanacaraka script characters and addressing variations in writing and classic character recognition challenges.
The dataset comprises 20 labels of Javanese script characters, totaling 12,000 data points as shown in Table 1

Convolutional Neural Network
CNN is a feed-forward neural network that specializes in image recognition tasks inspired by biological visual recognition.It has deep learning capabilities [14].As seen in Figure 2, CNN is composed of input, convolutional, pooling, fully connected, and output layers, in contrast to classic neural networks.In order to create a deeper network, CNN usually consists of multiple convolutional and pooling layers.Fully connected layers may also have a multi-layered structure [15].

Image 2. Convolutional Neural Network adapted
The core component of CNN is the convolutional layer.Multiple feature maps, each made up of many neurons, are present in each convolutional layer.In order to extract picture features, the convolutional layer uses convolutional kernels to scan the image and completely utilizes information from nearby regions [16].
After entering the data, the CNN method with the resnet-18 architecture is used to train the data, then the test data is entered to carry out classification or detection.To display the resulting image, you must enter validation data, this process is as shown in Figure 3 recall, and F1-Score [24].Below are the formulas used to calculate each metric: (2) (3) True Negative (TN) indicates accurate negative forecasts, whereas True Positive (TP) indicates right positive predictions.False positives, also called false alarms, are denoted by the acronym FP (False Positive).Mistakes, often known as misses, are represented by the symbol FN (False Negative).

RESULTS AND DISCUSSION
In the results phase, the use of Convolutional Neural Network (CNN) techniques and ResNet-18 architecture in recognizing Javanese Hanacaraka script characters shows impressive outcomes.

Image 5. Training Results
As seen in Image 5 with 100 epochs, The model achieved 100% training accuracy after 100 epochs, with 98.01% validation accuracy and 0.0950 validation loss.In testing, it reached 100% accuracy.To determine the accuracy, precision, recall, and F1-score, plug equations (1)-(4) into the confusion matrix provided in Figure 8 . Some commonly used CNN architectures include ResNet-18[17], MobileNet[18], AlexNet[19], InceptionV3 [20], and other architectures[21].Image 3. Flowchart CNN Method Working System Residual Network (ResNet-18) ResNet-18 is a variant of the ResNet architecture with 18 layers, emphasizing the use of residual blocks to address training challenges in neural networks.With two 3x3 convolutions and a shortcut connection, ResNet-18 becomes efficient and performs well[22].Starting with the initial convolutional layer, this model concludes with global pooling and fully connected layers for classification tasks, as illustrated in Figure 4. Despite its simplicity, ResNet-18 remains effective and can train networks without significant performance degradation[23].Its main advantage lies in its good generalization ability, especially in image classification tasks.Image 4. ResNet-18 Architecture adapted At the end of training, an evaluation will be conducted using metrics to measure the detection and classification performance of each CNN model.The metrics used include accuracy, precision,