Alphabet Recognition in Sign Language Using Deep Learning Algorithm with Bayesian Optimization

Hand gesture recognition has evolved significantly over time, starting from basic image processing techniques in the 1960s to sophisticated deep learning models today. In the 1980s and 1990s, machine learning techniques like neural networks and Hidden Markov Models (HMMs) were explored, but were limited by computational resources. The early 2000s saw a leap with Convolutional Neural Networks (CNNs), which improved accuracy by learning features directly from pixel data. Large annotated datasets further fueled progress. Recent advancements include recurrent neural networks (RNNs), attention mechanisms, and graph neural networks (GNNs), enabling models to capture temporal and spatial dependencies for better accuracy. Integration of depth sensors like Kinect and RealSense has enhanced 3D gesture recognition in real-world scenarios.

Currently, state-of-the-art hand gesture recognition methods leverage of deep learning architectures, large-scale annotated datasets, and advanced sensor technologies to achieve high accuracy and robustness across a wide range of applications, including human-computer interaction, sign language recognition, and virtual reality interfaces.

A notable example of this is the fusion approach, which involves the amalgamation of Histograms of Oriented Gradients (HOG) for hand shape depiction and Local Binary Pattern (LBP) for hand texture description, coupled with a Support Vector Machine (SVM) employing a radial basis function as its kernel. This method has demonstrated significant promise in enhancing the classification performance for hand gesture recognition, as discussed in the study [12]. Empirical evaluations conducted on the NUS hand posture dataset-II illustrate that the proposed HOG-LBP features significantly enhance hand recognition accuracy, achieving an average recognition accuracy of 97.8% for Subset A and 95.07% for subset B. The method proposed showing the strength of achieving a better classification performance on hand gesture recognition. Nevertheless, there remains a weakness to augment the speed and efficiency of hand gesture recognition, with a specific focus on complex background scenarios.

In a study focusing on the recognition of Indonesian Sign Language (BISINDO), researchers employed deep learning techniques to analyze results from a dataset of 1100 images across 10 categories [11]. The dataset is split into subsets A (1000 images for training) and B (100 for validation), with preprocessing techniques applied before training. Results show CNN achieving 73% accuracy, LSTM 81%, and a fusion of CNN and LSTM reaching 96% accuracy with a 17% loss. This highlights the synergy between CNN and LSTM in improving sign language recognition systems, a key strength. However, limitations include the dataset's scope, lacking comprehensive coverage of all symbols. Future research should focus on expanding the dataset and exploring additional aspects like expression detection and body gesture recognition to enhance overall interpretation systems.

In a comprehensive investigation, researchers analyzed an American Sign Language dataset, which consisted of approximately 205 images per numerical class (0 to 9) at 100x100 resolution [13]. A sophisticated ten-layer Convolutional Neural Network (CNN) was deployed, showcasing a strong 87.5% accuracy rate. This demonstrates the effectiveness of the tailored approach. Additionally, preprocessing reduces computational complexity, aiding efficiency. Manual hyperparameter fine-tuning ensures optimal performance, suitable for mobile applications and embedded systems. However, scalability and generalizability to other datasets or tasks beyond ASL recognition require further investigation, marking a potential weakness.

In the realm of hand gesture recognition, an extensive study explored the use of convolutional neural networks (CNNs) in recognizing American Sign Language (ASL) gestures [14]. It involves a dataset of 2515 200x200 pixel images of ASL numerical digits and alphabet characters. The CNN architecture includes four layers: two convolutional layers, batch normalization, a pooling layer, dropout layers, and flattening operation, followed by two fully connected layers, each with dropout. This approach achieves an impressive accuracy of 94.34% in ASL gesture recognition. Its strength lies in its accessibility via a simple camera and scalability through continuous training. Moreover, it offers a cost-effective alternative to specialized cameras like Microsoft's Kinect. However, reliance on pre-made datasets may limit adaptability to diverse real-world scenarios. Further research could enhance robustness to lighting conditions and hand positions.

In a study conducted by researchers [15], the Multi-scale Cross Feature Aggregation Network (Cross-Feat) achieved remarkable accuracy rates of 98.33% on the NUS hand posture dataset-ii and 99.5% on the ASL fingerspelling dataset, showcasing its strength in hand posture recognition and ASL fingerspelling. CrossFeat’s efficient design and effective feature preservation enable precise gesture recognition by capturing both local and global information. With only 975K trainable parameters, it is lightweight and computationally efficient. However, its performance in more complex environments beyond benchmark datasets is a potential weakness, necessitating further testing in diverse real-world scenarios to assess its robustness and generalizability.

The researchers conducted a series of experiments utilizing Moore's algorithm for edge detection and Fourier descriptor transform for feature extraction. These experiments, outlined in the study [16], achieved a notable 96.92% accuracy in recognizing Indonesian sign language alphabet, BISINDO, showcasing its potential for broader applications in image processing and recognition. The strength of Fourier descriptors lies in their effective feature extraction, but a weakness arises in their performance with translated and rotated images. Future work should focus on improving accuracy under such transformations, possibly through refining feature extraction or employing more robust distance metrics.

In a pioneering study conducted by researchers [17], an innovative approach to SIBI sign language recognition was introduced, using a cutting-edge 3D-CNN methodology. This study employed advanced camera technology to analyze a self-collected dataset, leading to a significant breakthrough. The strength lies in its ability to internally gather a large and varied dataset, a departure from conventional techniques. The researchers achieved an impressive accuracy rate of 97.5% with their model. This underscores the potential of 3D-CNN technology and innovative data collection strategies. The strength of transfer learning here lies in its effectiveness with small datasets, as evidenced by the significant accuracy improvement. However, a potential weakness is its sensitivity to variations in data distribution or domain shifts, which could affect generalization across different datasets or real-world scenarios. Thus, careful consideration and further evaluation are needed to ensure the method's robustness and adaptability beyond the study's specific context.

In their comprehensive research, the authors conducted extensive experiments on the BISINDO dataset, employing various methodologies to explore sign language recognition [8]. They used deep learning models like AlexNet and VGG-16, achieving high accuracy rates of 98.6% and revealing dataset characteristics. They also developed a custom CNN model with a commendable accuracy score of 98.3%. These findings offer valuable insights into sign language recognition. The strength of the simplified CNN model lies in its effective recognition of BISINDO hand gestures, showing robustness across different conditions. However, its scalability and performance in complex environments like varied backgrounds remain unclear, affecting its real-world applicability. Additionally, concerns persist regarding its computational efficiency and speed for real-time implementations.

In their seminal study [18], researchers have introduced a novel approach using Convolutional Neural Networks (CNNs) to advance American Sign Language (ASL) recognition. The objective was to accurately recognize a diverse set of 24 ASL alphabet signs collected from real-life scenarios, showcasing the model's robustness. The CNN architecture, designed for intricate spatial feature capture, includes convolutional layers for pattern detection, max-pooling for information preservation, and a fully connected layer for feature fusion. Experimental results demonstrated an impressive 99.3% accuracy, highlighting the approach's efficacy. Strengths include efficient Deep Learning utilization, achieving remarkable accuracy, and fast training due to small elapsed time and low loss error. However, a potential weakness lies in its reliance on static images, limiting applicability to dynamic scenarios. Further exploration into dynamic gesture recognition methods could enhance versatility and robustness.

In the realm of empirical investigations using a curated dataset [19], a novel convolutional neural network (CNN) architecture was devised, incorporating two convolutional layers, max-pooling, dropout, and densely connected layers, totaling 4,073,540 parameters. This approach yields remarkable accuracy scores of 99.72% for color data and 99.9% for grayscale data. Strengths include comprehensive methodology and robust performance, with high accuracies across various conditions and datasets. However, scalability and adaptability to real-time processing, especially with dynamic signs and video-based datasets, remain weaknesses.

After an in-depth exploration of related research in the field, this study aims to contribute new insights through a comparative analysis of three prominent image classification models: ResNeXt101 [20], VGG19 [21], and ViT [22] which have a good track record in image classification. The approach goes beyond traditional evaluation by incorporating Bayesian optimization [23] for hyperparameter tuning where no previous related research has explicitly described the hyperparameter process. This advanced optimization technique is used to systematically fine-tune the configuration of each model. The expected outcome is a better understanding of how these models work in the specific context of this study, shedding light on the impact of hyperparameter tuning on their overall effectiveness. By conducting this comparative study, the authors aim to see real differences in performance, offering valuable insights that can inform future developments in the field of image classification and model optimization.

The experiments in this research will employ parameters outlined in Table 3, which serves as a reference for the baseline model settings. These baseline parameters are shared between ResNeXt101, VGG19, and Vision Transformer models. For optimization, Stochastic Gradient Descent (SGD) will be used, with a common momentum value of 0.9. Additionally, the table provides insights into the total number of parameters for each model, which are essential in understanding the model complexity.

Table 3. Baseline model parameters

Specifically, ResNeXt101 contains 20,101,194 parameters, while VGG19 and Vision Transformer are characterized by a more extensive architecture, with a total of 42,149,194 parameters, and 87,423,754 parameters. These baseline settings serve as a starting point for the forthcoming experiments, allowing for a comparative evaluation of model performance and efficiency against the results obtained using other hyperparameter configurations.

In conjunction with these baseline settings, early stopping will be implemented with a criterion of patience set to 10 epochs and a minimum delta of 0.001. This approach aims to enhance training efficiency by halting the process if, for 10 consecutive epochs, there is no significant improvement in the loss function beyond the specified threshold. Such a strategy not only optimizes computational resources but also potentially prevents overfitting, thus positively influencing the generalization performance of the models. Consequently, integrating early stopping into the training process is anticipated to contribute to a more robust evaluation of model performance and efficiency, augmenting the findings derived from various hyperparameter configurations, bears the potential to influence the overall training dynamics and subsequently impact the results, warranting careful consideration and analysis throughout the experimentation phase.

The choice of the loss function plays a pivotal role in guiding the optimization process and ultimately determining the model's performance. The selected loss function for the task at hand is cross-entropy loss, which is commonly employed in classification tasks and particularly suited for scenarios where the model predicts probabilities across multiple classes. The appropriateness of cross-entropy loss stems from its ability to confidently penalize incorrect predictions more severely, thereby encouraging the model to learn more discriminative features and make accurate classifications.

4.1 Experiment on dataset subset A

Table 4 displays the key parameters obtained through a rigorous process of hyperparameter tuning. This tuning was carried out for three distinct models, namely ResNeXt101, VGG19 and ViT, which are used in the subsequent experiments. These parameters were determined using Bayesian optimization and are crucial in defining the performance of the models. In particular, the learning rates for ResNeXt101, VGG19, and ViT were found to be 0.0001987, 6.593e-05, and 0.0001308, respectively. The batch sizes chosen for the three models were 16 for ResNeXt101, 128 for VGG19, and 16 for ViT, while all models were trained for 200 epochs. It's worth noting that early stopping criteria were applied, resulting in the conclusion of training at the 143rd epoch for ResNeXt101, the 151st epoch for VGG19, and the 83rd epochs for ViT.

Table 4. Parameters obtained from hyperparameter tuning

Table 5. Training and validation results on accuracy and loss

The first experimental phase utilized the optimized parameters from Table 4 to conduct experiments on a subset of the data A, assessing the performance of pre-trained models ResNeXt101, VGG19, and ViT. Notably, all models showed significant improvements in accuracy and reductions in loss. Details of the loss and accuracy values for both training and validation phases are provided in Table 5. In summary, ResNeXt101 and ViT achieved remarkably low training losses of 0.0010 and 0.0012, with 100.00% training accuracy. VGG19 exhibited a slightly higher training loss of 1.4631 but maintained a commendable training accuracy of 99.92%. During validation, ViT achieved a loss of 0.0552 and an accuracy of 98.50%, showcasing its effectiveness in generalization. ResNeXt101 and VGG19 had validation losses of 0.0860 and 1.4692, with validation accuracies of 98.00% and 96.00%, respectively, demonstrating robust performance. These results underscore the effectiveness of the chosen hyperparameters and the capabilities of the three models in both training and validation phases.

Table 6 presents the experimental results for accuracy and F1-score for three prominent models, ResNeXt101, VGG19, and ViT. These metrics are essential for evaluating the models' overall performance. ResNeXt101 achieved a remarkable accuracy of 99.03%, indicating its high precision in correctly classifying data points. Additionally, it recorded an impressive F1-score of 0.99, underscoring its effectiveness in both precision and recall. ViT and VGG19, while slightly lower in accuracy at 99.01% and 97.01%, still demonstrated strong classification capabilities. It achieved an F1-score of 0.99 and 0.97, suggesting its proficiency in achieving a balance between precision and recall. These results illustrate the excellent performance of all models, with both model ResNeXt101 and ViT exhibiting a slight edge in accuracy and F1-score, making it a potentially preferable choice for tasks that prioritize precision and recall in classification.

Table 6. Testing result on dataset subset A

4.2 Experiment on dataset subset B

In Table 7, the authors present the parameters obtained through a meticulous hyperparameter tuning process for three distinct models, ResNeXt101, VGG19. and ViT. These parameters are critical in shaping the training and performance of the models. For all models, the author determined the optimal learning rate, batch size, and the number of training epochs. The learning rate for ResNeXt101 was set at 0.0018, while VGG19 and ViT used a slightly higher value of 0.0022 and 0.0112. All models utilized a batch size of 16 during training, and the author conducted training for a total of 200 epochs, implementing an early stopping mechanism at 32 epochs for ResNeXt101, 42 epochs for VGG19 and 34 epochs for ViT.

Table 7. Parameters obtained from hyperparameter tuning

In the second experiment, the authors focused on evaluating a subset of data labeled as 'B' using hyperparameter settings derived from Bayesian optimization results, as outlined in Table 7. Employing pretrained ResNeXt101, VGG19, and ViT models, significant improvements in accuracy and loss reduction were observed. Table 8 provides specific numerical values for training and validation loss and accuracy, offering a comprehensive overview of model performance. ViT and ResNeXt101 exhibit remarkably low training loss of 0.0001 and 0.0206, respectively, highlighting their capacity to fit the training data well. Similarly, VGG19 achieves a respectable training loss of 1.4612. Regarding validation loss, ResNeXt101 achieves a value of 0.0890, while VGG19 and ViT record validation losses of 1.5081 and 0.4611, respectively. These results underscore the robust performance of the models in generalizing to unseen data, further validating the efficacy of the chosen hyperparameter settings and the overall success of the experiment.

Table 8. Training and validation results on accuracy and loss

Table 9. Testing result on dataset subset B

Table 9 presents the experimental results for accuracy and F1-score, focusing on three prominent models, ResNeXt101, VGG19, and ViT. In these experiments, ResNeXt101 achieved a commendable accuracy of 95.31%, showcasing its ability to correctly classify data instances. Additionally, it obtained an F1-score of 0.95, indicating its proficiency in striking a balance between precision and recall, which is crucial in classification tasks. VGG19 and ViT, while slightly lower in accuracy at 93.97% and 89.97%, still demonstrated strong performance in data classification. It obtained an F1-score of 0.93 and 0.89, highlighting its reliability in handling classification tasks with a favorable trade-off between precision and recall. These results underscore the effectiveness of all models ResNeXt101, VGG19, and ViT in the context of accuracy and F1-score, offering valuable insights into their capabilities for specific tasks.

4.3 Experiment on combination dataset subset A and B

Table 10 provides an overview of the hyperparameters obtained through a thorough tuning process for the third experiment. The three models used, ResNeXt101, VGG19, and ViT, have different parameter settings. For ResNeXt101, the learning rate was set at 0.0159, while a batch size of 128 was utilized, and training continued for 200 epochs, with early stopping occurring at the 22nd epoch. Meanwhile, VGG19 and ViT were configured with a learning rate of 0.0043 and 0.008679, with a batch size of 32 and 16, and it also trained for 200 epochs, but early stopping took place at the 19th epoch and 30th epochs.

Table 10. Parameters obtained from hyperparameter tuning

In the third experiment, a combination of data from subsets A and B was tested using the hyperparameters obtained from Bayesian optimization, as detailed in Table 10. All models demonstrated impressive performance improvements, exhibiting increased accuracy and reduced loss within each epoch. Table 11 presents a comprehensive view of the experiment's results, showcasing the training and validation metrics. Both the ResNeXt101 and ViT models achieved remarkable 100% accuracy in training, with a negligible training loss of 0.0000. Moreover, they maintained strong accuracies of 98.36% and 97.82% on the validation set, with low validation losses of 0.021 and 0.0559, respectively. Conversely, the VGG19 model also achieved high training accuracy of 100%, albeit with a slightly higher training loss of 1.4612. In the validation phase, VGG19 demonstrated an accuracy of 96.73% with a validation loss of 1.4910. These results in Table 11 highlight the exceptional performance of both models, emphasizing their suitability for the task at hand.

Table 11. Training and validation results on accuracy and loss

In Table 12, the authors present the experimental results of accuracy and F1-score for three well-established models, ResNeXt101, VGG19, and ViT. These results provide valuable insights into the performance of these models in the context of the task at hand. The ResNeXt101 model achieved an impressive accuracy of 99.02% and an F1-score of 0.99, indicating its remarkable capability to correctly classify and provide a balanced trade-off between precision and recall. Similarly, the VGG19 and ViT models displayed a high level of accuracy, registering at 99.01% and 988.42, and while its F1-score was slightly lower at 0.97amd 0.98, it still showcased strong overall performance. These results underscore the effectiveness of all models ResNeXt101, VGG19, and ViT in accurately and reliably classifying the data, making them valuable choices for this task.

Table 12. Testing result on combination dataset subset A and B

In comparing sign language recognition models ResNeXt101, VGG19, and ViT to the previous HOG-LBP + SVM approach on the NUS Hand Posture Dataset-II, notable advancements and limitations arise. In Subset A, VGG19 falls short while ResNeXt101 and ViT succeed. In Subset B, only ResNeXt101 succeeds, with VGG19 and ViT showing shortcomings. Combined experiments yield high accuracy, indicating overall success. ResNeXt101 consistently outperforms other models, though its computational complexity and resource-intensive training are notable. ViT may struggle with fine-grained spatial information, and VGG19's limitations underscore the need for further research on model scalability and effectiveness in sign language recognition tasks.