Evaluating the Impact of Optimizer Hyperparameters on ResNet in Hanacaraka Character Recognition

1 Introduction

Character recognition is a crucial field in image processing and computer vision, with various applications such as Optical Character Recognition (OCR), handwriting recognition, and document digitization. In the context of Hanacaraka character recognition, the traditional Javanese script, the challenges are significant due to its complex and diverse characters.

ResNet (Residual Network) is a convolutional neural network architecture that has proven to be highly effective in various image recognition tasks (Han-wen et al. 2021; Singh and Schicker 2021). One of the main advantages of ResNet is its ability to overcome the vanishing gradient problem through the use of residual blocks (Nicholas et al. 2022; Sudewo, Biddinika, and Fadlil 2024b). This allows for the creation of very deep networks without losing critical information during training.

Recent studies demonstrate the capability of Convolutional Neural Networks (CNN) in solving diverse classification problems with high accuracy and efficiency. Murinto and Melany (2023) showed that MobileNetV2 combined with transfer learning improved classification accuracy for coffee beans up to 96 %, outperforming conventional CNN approaches (Murinto and Melany 2023). Similarly, Cahya and Murinto (2021) achieved remarkable results in classifying batik motifs, with a training accuracy of 100 % and testing accuracy of 99 %, underscoring the potential of CNN for intricate pattern recognition tasks (Cahya and Murinto 2021). In another study, Rosyda (2022) successfully applied the LogDIWPSO optimization algorithm to improve CNN performance on the CIFAR-10 dataset, demonstrating a significant accuracy increase from 28.07 % to 69.3 % within 10 epochs (Rosyda 2022). Lei, Pan, and Huang (2019) introduced the Hybrid Dilated CNN (HDC) model, reducing training time while enhancing accuracy by over 14 %, showcasing advancements in architectural innovations to improve CNN efficiency and precision (Lei, Pan, and Huang 2019).

This study aims to evaluate the impact of optimizer hyperparameters on the performance of ResNet in the task of Hanacaraka character recognition. By conducting systematic experiments and analyses on various optimizers and their settings, we aim to provide practical guidelines for developing more effective and efficient character recognition models. The integration of state-of-the-art optimization techniques and architectural insights is expected to address the challenges posed by the intricate nature of Hanacaraka script, paving the way for advancements in traditional script recognition technologies.

This study aims to evaluate the impact of optimizer hyperparameters on the performance of ResNet in the task of Hanacaraka character recognition. By conducting systematic experiments and analyses on various optimizers and their settings, we can gain a better understanding of how each optimizer affects the model’s convergence, accuracy, and generalization ability. The results of this research are expected to provide practical guidelines for developing more effective and efficient character recognition models.

2 Materials and Methods

This study uses a quantitative methodology, utilizing image data as the computational subject using the CNN method with ResNet-18 architecture and a learning rate of 0.001, with the optimizer as an additional comparison component. This research utilizes Google Colab and various libraries as supporting tools.

2.1 Research Steps

The flowchart of this research, shown in Figure 1, begins with a literature review to understand the basics and developments in Hanacaraka character recognition using CNN algorithms. Next, a dataset of Javanese script character images is collected and resized to 64 × 64 pixels. The preprocessing step is then performed by converting images to tensors and applying data augmentation if necessary. The ResNet-18 based CNN algorithm is implemented and the model is trained using various optimizers such as SGD, Adam, and others for 10 epochs. After model training, the model is tested with the validation dataset during the model testing phase to measure its performance. The testing results are evaluated based on accuracy, precision, recall, and F1-Score, determining the best optimizer. Finally, the output in the form of predicted images from the CNN model demonstrates the model’s performance in recognizing Javanese script characters.

Figure 1:

Research steps.

2.3 Datasets

This research uses a dataset of Hanacaraka Javanese script character images sourced from Kaggle, provided by Hanna Hunnafa, as shown in Figure 2. The dataset consists of 12,000 PNG images, divided into 8,400 training data images, 2,400 validation data images, and 1,200 testing data images.

Figure 2:

Javanese Hanacaraka script.

2.4 Preprocessing

Preprocessing is a stage to prepare the images before training the model using the dataset. This step is crucial to ensure that all images classified by the CNN method have uniform pixel sizes, as this can affect the accuracy of the results (Hasan et al. 2020). The preprocessing process includes converting image channels from RGB to grayscale and resizing the images so that all data have the same pixel resolution.

Figure 3(a) shows an image sized at 224 × 224 pixels. This size makes it difficult to train the model due to the high number of pixels that need to be extracted, necessitating the resizing technique during preprocessing. The resizing process involves reducing the image size from 224 × 224 pixels to 64 × 64 pixels as shown in Figure 3(b). Although this size differs significantly from the original size, it does not reduce the information in the 224 × 224 pixel image.

Figure 3:

Preprocessing Hanacaraka Script (a) before (b) after.

2.5 Convolutional Neural Network

Convolutional Neural Network (CNN) is one of the most effective artificial neural network architectures for processing grid-structured data such as images (see Figure 4). CNN consists of several key layers that work sequentially to extract and process important features from the input data (Liu, Pu, and Sun 2021; Muis, Sunardi, and Yudhana 2023). First, the convolutional layer uses filters to capture local patterns such as edges, textures, and other visual patterns from the images. This layer allows the network to learn hierarchical representations of the input where each filter is responsible for extracting specific features (Basha et al. 2020). After that, the pooling layer is used to reduce the dimensions of the feature maps produced by the convolutional layer, as shown in Figure 4. This not only reduces computational complexity but also helps prevent overfitting by simplifying the generated representations.

Figure 4:

Convolutional Neural Network (taken from: Researchgate.net).

Then, the ReLU (Rectified Linear Unit) layer introduces non-linearity into the network by converting negative values to zero. This is important because non-linearity allows CNN to learn more complex relationships between input features, enhancing the network’s ability to accurately classify and recognize objects. Finally, the fully connected layer at the end of the network combines the features extracted from the entire image to make the final decision such as classifying the objects in the image.

Some well-known CNN architectures, such as AlexNet (Madhulatha and Ramadevi 2020), VGGNet (Agrawal and Mittal 2020), GoogLeNet (Inception) (Shadin, Sanjana, and Lisa 2021), and DenseNet (Sudewo, Biddinika, and Fadlil 2024a), each offer innovations in the use of convolutional layers, managing multi-scale information, and addressing gradient issues. CNNs are not only used in image recognition and classification but are also widely applied in object detection, image segmentation, and even natural language processing. Their success in various image processing tasks makes CNN one of the most effective and popular tools in the world of pattern recognition and visual analysis.

2.6 ResNet-18

ResNet-18 is a variant of the Residual Network (ResNet) architecture known for balancing relatively shallow depth with its ability to address the vanishing gradient problem (Lu et al. 2023). By using residual blocks, ResNet-18 can build deeper networks without experiencing performance degradation (Chandu and Bharatha Devi 2023). Each block in ResNet-18 consists of several convolutional layers followed by skip connections, allowing direct information flow across multiple layers (Ahmed et al. 2024), as shown in Figure 5.

Figure 5:

Resnet-18 (source: Researchgate.net).

This capability makes ResNet-18 a popular choice for various image recognition tasks, such as image classification, object detection, and segmentation, due to its combination of good performance and relatively simple structure.

3 Results and Discussion

The results of comparing hyperparameter optimizers on the CNN model based on ResNet-18 were obtained by training the model using a dataset of Javanese Hanacaraka script images, divided into three sets: training, validation, and test. The images were resized to 64 × 64 pixels and converted into tensors. The learning rate used was 0.001, and the batch size was 64. The model was implemented using optimizers such as SGD, Adam, RMSprop, Adagrad, Adadelta, Nadam, and Adamax on the validation set and trained for 10 epochs.

In Figure 6, NAdam shows the best performance among all the optimizers tested, with accuracy, precision, recall, and F1-Score reaching 100 %. This indicates that NAdam can optimize the model maximally for the task of recognizing Javanese characters, providing perfect results without errors. The use of the Nesterov accelerated gradient (NAG) technique combined with moment adaptation from Adam appears to offer an advantage in faster and more efficient convergence.

Figure 6:

Training results.

Adamax also provides excellent results, with accuracy, precision, recall, and F1-Score of 97.92 % each. Adamax, as a variant of Adam that uses infinity norm for parameter updates, seems to offer very good numerical stability, allowing the model to handle large fluctuations in gradients.

Adam and Adagrad show high performance with metric values above 97 %, making both strong choices for the task of Hanacaraka character recognition. Adam, known for its combination of momentum and learning rate adaptation, and Adagrad, which adjusts the learning rate based on gradient frequency, both provide consistent and accurate results.

SGD (Stochastic Gradient Descent) performs quite well with 93.72 % accuracy and 93.57 % F1-Score. Although not as good as more advanced optimizers like Adam or NAdam, SGD remains a reliable and simple choice, especially for tasks that are not overly complex.

Adadelta shows adequate performance with 86.58 % accuracy and 86.15 % F1-Score. While not as effective as Adam or Adamax, Adadelta remains a solid choice, particularly when stability in parameter updates is a primary consideration.

Conversely, RMSprop yields the lowest performance with 81.74 % accuracy and 81.60 % F1-Score. RMSprop, designed to address the declining learning rate issue in Adagrad, appears less effective for the task of Hanacaraka character recognition in this study. This lower performance may be due to its inability to handle large variations in gradients as effectively as other optimizers.

The evaluation results show that the choice of optimizer significantly impacts the model’s performance in the task of recognizing Javanese Hanacaraka characters. NAdam and Adamax proved to be the best choices, providing highly accurate and consistent results in detecting Javanese characters. NAdam’s advantage is particularly evident from the combination of the NAG technique and moment adaptation, offering faster and more stable convergence.

Adam and Adagrad also demonstrate very good performance, reflecting the effectiveness of algorithms capable of adjusting the learning rate during training. Although not as effective as more advanced optimizers, SGD still provides solid and reliable results in many situations.

In contrast, RMSprop’s performance shows that although some optimizers are designed to address specific weaknesses in other algorithms, they may not always be suitable for all types of tasks. In this case, RMSprop was less effective in handling large gradient variations for Hanacaraka character recognition.

Overall, this study emphasizes the importance of selecting the appropriate optimizer to achieve optimal performance in Hanacaraka character classification. Optimizers such as NAdam and Adamax show great potential for practical applications in character recognition while results obtained from other optimizers provide valuable insights into their effectiveness under various training conditions.

Table 1 presents the evaluation results of various optimizers in recognizing Javanese Hanacaraka characters, covering precision, recall, and F1-Score metrics for each character. From the table, it is evident that the NAdam and Adamax optimizers consistently deliver high performance across almost all characters, with precision, recall, and F1-Score values close to or reaching 1.00. For instance, for the character “ba,” these two optimizers achieve perfect scores in all three metrics. This indicates that these optimizers are highly effective in accurately and consistently identifying Javanese script characters. Conversely, optimizers such as RMSprop, Adagrad, and Adadelta exhibit greater performance variation. For example, RMSprop has low precision values for some characters like “da” (0.85) and “ba” (0.73) while Adagrad and Adadelta show good results for some characters but lack overall consistency. SGD and Adam consistently deliver good results, though still falling short of NAdam and Adamax in some cases. From these results, it can be concluded that NAdam and Adamax are the best choices for the task of recognizing Javanese Hanacaraka characters, providing higher accuracy and consistency compared to other optimizers.

Figure 7 displays a series of graphs comparing the performance of various optimizers used in training the neural network. Each graph shows loss and accuracy for training and validation data over several epochs. The optimizers compared include Adam, SGD, RMSprop, AdaGrad, Adadelta, and NAdam. The four lines on each graph represent Training Loss (blue), Validation Loss (orange), Training Accuracy (green), and Validation Accuracy (red).

Figure 7:

Accuracy and loss graph.

The Adam optimizer demonstrates stable performance with a consistent decrease in Training Loss and a gradual increase in Training Accuracy. However, slight overfitting is visible from the difference between Training Accuracy and Validation Accuracy. The Adadelta optimizer shows a similar pattern, with a consistently decreasing training loss and increasing accuracy for both training and validation data, indicating better generalization compared to other optimizers.

On the other hand, RMSprop and NAdam optimizers show significant instability, evidenced by large fluctuations in Validation Loss and Validation Accuracy. Although Training Loss for these optimizers decreases well, the instability in validation data suggests that the model may have generalization issues. This makes these optimizers less reliable than Adam and Adadelta in terms of overall performance.

The SGD optimizer shows a rapid decrease in Training Loss but reaches a plateau, indicating that adjustments to the learning rate or more epochs might be needed for better results. Meanwhile, the AdaGrad optimizer seems to experience overfitting, with Training Accuracy increasing but Validation Accuracy remaining flat. This indicates that while the model learns well on training data its performance does not improve on validation data. Overall, Adam and Adadelta appear to provide the best balance between training and validation performance while other optimizers exhibit issues with overfitting or instability. There are some differences with the literature review results, as shown in Table 2.

In the comparison of hyperparameter optimization results for the ResNet-18 based CNN model in recognizing Javanese Hanacaraka characters, NAdam proved to be the best optimizer with accuracy, precision, recall, and F1-Score reaching 100 %. This demonstrates NAdam’s effectiveness in optimizing the model for this task, outperforming other optimizers such as Adamax, Adam, and SGD which also showed good performance, albeit with slightly lower values.

Comparisons with other studies show variations in CNN application and results depending on the object. For example, a study on coffee bean classification using MobileNetV2 achieved 96 % accuracy while batik motif classification reached 100 % accuracy on training data and 93.3 % on test data. This indicates that each model and dataset have specific characteristics and results, although the focus on using CNN for classification tasks varies.

4 Conclusion and Recommendation

The conclusion of this study is that the choice of optimizer significantly impacts the model’s performance in recognizing Javanese Hanacaraka characters. Among the tested optimizers, NAdam proved to be the best, achieving 100 % accuracy, precision, recall, and F1-Score. This highlights NAdam’s effectiveness in optimizing the model through the combination of Nesterov Accelerated Gradient (NAG) and momentum adaptation, leading to faster and more stable convergence. Adamax also demonstrated excellent performance, followed by Adam and Adagrad, with all metrics scoring above 97 %.

SGD, although simple, performed reasonably well with an accuracy of 93.72 % but was outperformed by more advanced optimizers like NAdam and Adamax. Adadelta provided adequate performance with stable parameter updates, while RMSprop showed the lowest performance, likely due to its inability to handle large gradient variations effectively for Hanacaraka character recognition.

Overall, this study emphasizes the importance of selecting the appropriate optimizer to achieve optimal performance in Hanacaraka character classification. NAdam and Adamax proved to be the most effective choices while the results from other optimizers provide additional insights into their effectiveness under different training conditions.