Revolutionizing Signature Recognition: A Contactless Method with Convolutional Recurrent Neural Networks

Conventional contact-based hand signature recognition methods are raising hygienic concerns due to shared acquisition devices among the public. Therefore, this research aimed to propose a contactless in-air hand gesture signature (iHGS) recognition method using convolutional recurrent neural networks (C-RNN). Experiments have been conducted to identify the most suitable CNN architecture for the integration of CNN and RNN. A total of four base architectures were adopted and evaluated, namely MS-CNN-A, MS-CNN-B, CNN-A, and CNN-B. Based on the results, CNN-A was selected as the convolutional layer for constructing the integration of C-RNN due to its superior performance, achieving an accuracy rate of 95.15%. Furthermore, three variants of C-RNN were proposed, and experimental results on the iHGS database showed that the ConvBiLSTM achieved the highest accuracy at 98.10%, followed by ConvGRU at 97.47% and ConvLSTM at 97.40%.

Convolutional-RNN; Gesture recognition; Hand gesture signatures; In-air signatures

Hand signature is a handwritten name or initials that are typically used for identity verification and authentication purposes in legal documents, banking and finance transactions, employment agreements, and government services (Hashim, Ahmed, and Alkhayyat, 2022). Due to its uniqueness, two individuals sharing the same name will show distinct hand signature. This is due to individual behavioral traits such as pressure applied to the writing surface, speed, and the angle of handling the writing instrument (Impedovo and Pirlo, 2007). These factors will influence the overall appearance and style of signature, which contribute to its uniqueness. The practicality and widespread use of hand signature has made it to become one of the most commonly applied forms of behavioral biometrics. However, the traditional method for recognition is through manual checking by humans, which is susceptible to error during the identification of the signature (Akram, Qasim, and Amin, 2012).

    The advancement of technology has led to the introduction of automatic hand signature recognition systems, which have mitigated the risk of human error in the identification of hand signatures (Faundez-Zanuy, 2005). Generally, there are two types of hand signature recognition systems, namely offline and online (Julian and Ortega-Garcia, 2008; Gilperez et al., 2008). The offline method only captures the appearance and shape by scanning the hand signature. Meanwhile, the online method captures other dynamic properties such as pressure applied to the writing surface and signature completion time. This system can be further classified into contact-based and contactless methods (Jain et al., 2020; Malik et al., 2018). The contact-based method captures the hand signature using a pen and paper, as well as the latest method, which uses a stylus and tablet technology. The contactless method does not require direct contact with acquisition devices by in-air signing in front of a camera sensor.

The contact-based method that uses pen and paper is vulnerable to forgery issues as individuals need to physically write their signature, exposing the shape and allowing imitation by others for unauthorized uses. Although the stylus and tablet method can address this issue by capturing dynamic properties, it is still a contact-based susceptible to germ and virus contamination as the acquisition devices are often shared among the public. Recently, the occurrence of the COVID-19 global pandemic has increased concerns about hygiene related to contact-based biometrics recognition systems, leading to high demands for contactless systems (Gan, 2022; Romadlon, Lestiana, and Putri, 2022; Supriatna et al., 2022; Yatmo, Harahap, and Atmodiwirjo, 2021; Carlaw, 2020). Contactless hand signature recognition methods are already in existence (Khoh et al., 2021; Fang et al., 2017; Sajid and Cheung, 2015), but their applications in currently in the experimental stage. According to previous research, commercially available systems still rely on contact-based methods for applications such as identity verification in financial institutions (Sudharshan and Vismaya, 2022; Karanjkar and Vasambekar, 2016; Cuceloglu and Ogul, 2014). Shao et al., (2020), Levy et al., (2018) proposed hand gesture-based signature recognition method using built-in sensors from mobile devices. Subjects perform in-air signatures when wearing a smartwatch or holding a smartphone to capture the gestures. However, this method only captures the device's coordinates over time, lacking spatial information on hand. Azlin et al. (2022) proposed a hand gesture-based recognition method using a camera sensor to capture signature. The method mainly focuses on compressing entire image sequences into a single image for recognition, without providing temporal information.

    The database and pre-processing methods were initially introduced in this section, followed by a detailed explanation of proposed CNN and integrated C-RNN.

2.1. In-air hand Gestures Signatures Database (iHGS)

        This database comprised a total of 2980 samples, including 2000 genuine and 980 forged samples. Hand gesture signature acquisition was carried out in a controlled environment, using a Microsoft Kinect sensor that captured both color and depth images. The sensor was pre-set to operate at a resolution of 640 x 480 and 30 frames per second. Furthermore, 100 individuals who participated in database collection were instructed to stand one meter away from the sensor and position their hand in front, activating the standby mode, before performing hand gesture signature. Each individual contributed 20 samples of genuine signature. For the collection of forged samples, participants were given a sufficient amount of time to learn and replicate the signature of other individuals. The acquisition of forged signature started after the participants showed readiness, with Figure 1 showing a sample of the depth image.

Figure 1 Depth image frame of a subject writing a signature in the air with their hand.

2.2. Pre-processing

        In this research, only the depth image sequences of in-air hand gestures were used to evaluate the performance of the proposed methods. After segmenting the palm area, 2D and 3D features were generated. Specifically, 2D features store only spatial information, which is used to train the base CNN. 3D features store both spatial and temporal information, which are designated for training the integrated C-RNN.

2.2.1. Palm Detection and Segmentation

        Although CNN is designed to learn and extract features without the need for handcrafted feature extraction algorithms, the process still requires a large number of training samples to obtain optimal performance. Due to the scarcity of hand gesture signature samples, the background noises were removed to prevent the models from misinterpretation. The entire palm area is preserved and considered as the region of interest (ROI) of a particular sample.

        The palm detection is initially performed using the thresholding method to remove the noise. A threshold value is set to 180, according to the outcome of the empirical tests to obtain the most optimum setting. Any pixel value in the image sequences below this threshold value is set to 0, corresponding to the pixel value of the black color. Although the palm region serves as the closest object to the sensor at the beginning of hand gesture signature motions, there are no restrictions on how the subjects can perform hand gesture. This phenomenon leads to the detection of both the palm and face regions. For instance, when subjects move their hand to the far left or right, both the face and palm regions become the closest objects to the sensor. Figure 2 shows instances where both regions are segmented together in the same frame. To address this issue, the predictive palm segmentation algorithm proposed by (Khoh, Pang, and Teoh, 2019) is applied, using the detected palm location in the first frame as a reference to predict the closest palm point in other frames. Figure 3 shows that only the palm region remains in the frame after applying the predictive palm segmentation.

Figure 2 Incorrect segmentation where both the face and palm region are captured

Figure 3 Correct segmentation where only the palm region is captured after applying the predictive palm segmentation algorithm

2.2.2. Feature Generation

        In the feature generation stage, both 2D and 3D imagery features are generated from the segmented image sequences. The Motion History Image (MHI), proposed by (Davis and Bobick, 1997) is used for generating features for both 2D and 3D hand gesture signature features. Moreover, the MHI is a static image that condenses spatio-temporal information from a sequence of frames into a single image (Sincan and Keles, 2022; Lakshmi and Devendran,  2021; Ahad et al., 2012). In the generation of 2D features, in-air hand gesture signature – Motion History Image (iHGS-MHI) is used as the designated notation. The silhouettes of palm motion from the segmented image sequences are compressed into a single static image. Figure 4 shows the process of generating iHGS-MHI features, where MHI is produced from the segmented image sequences.

The segmented image sequences are already in a 3D format which can be fed into the recurrent neural network (RNN). However, the method is not computationally efficient due to the need to identify the largest frame in the samples and apply zero-padding to standardize the number of depths for each hand gesture signature. This phenomenon is capable of causing an increase in computational resources as the model's input dimension needs to match the largest sample's frame numbers. The zero-padding method can also cause the model to consider the padded frames as part of the features due to the variation in the number of frames across the hand gesture signature samples. For instance, some samples can have fewer than 50 frames, while others are above 100 frames. To address this issue, a more computationally efficient method for 3D feature generation is introduced to reduce the number of padding frames.

The depth for each sample is standardized to 10 blocks, containing 10 frames. In the initial stage of 3D feature generation, hand gesture samples of fewer than 100 frames are categorized as small, while those exceeding 100 frames are categorized as large. The segmented image sequences are partitioned into 10 blocks and MHI is generated from each block, compressing 10 frames into a single image, as shown in Figure 5. In comparison, the 10th block of small samples is an empty block filled using the zero-padding method, while large samples contain more than 10 frames, ranging from frame number 91 to the end of the sequence. Figure 6 shows a scenario featuring one sample with 43 frames and another with 183 frames. The small samples have sufficient frames only to the 4^th block and the remaining are padded with zeros, while the large samples exceed 100 frames. The 10^th block stores frame numbers 91 to 183, and the MHI appears larger compared to the previous 9 blocks due to the compression of more frames. Compared to the previous blocks, the 10^th block in a large sample captures more than 10 frames.

Figure 4 Feature generation of iHGS-MHI

Figure 5 Feature generation of iHGS-MHI-BLOCKS

Figure 6 Illustration of HGS-MHI-BLOCKS on small and large samples

2.3. Proposed architecture

        In this section, the architecture of both CNN and RNN are presented, along with the integration method to build C-RNN.

2.3.1. Convolutional Neural Networks (CNN)

In this research, two variants of CNN are proposed, namely multiscale Convolutional Neural Networks (MS-CNN) which extract and learn features at multiple scales simultaneously, and the base CNN operating at a single scale. Each variant has the same number of layers. The first layer of the models is a convolution layer with a kernel size of 3x3 and 32 filters. The second layer also has a kernel size of 3x3 but with 64 filters. Meanwhile, the main difference between both models is at the third and fourth layers. In MS-CNN, these layers consist of parallel convolution layers with 3x3 and 5x5 kernels, each having 64 filters to extract features at different scales. The feature maps from both convolutional layers are concatenated, resulting in a total of 128 filters. The third and fourth layers of the base CNN consist of convolutional layers with 128 filters each. Figure 7 shows the difference between parallel and typical single-scale convolution layers.

Figure 7 Comparison of parallel and typical convolution layers

The outputs of each convolutional layer are activated using the Rectified Linear Unit (ReLu) function. Max pooling is applied at the end before the feature maps are passed to other layers for further feature extraction and dimension reduction. The final layers of these models consist of fully connected (FC) layers, which learn and map the features extracted by the previous convolutional layers. Moreover, the first FC layer consists of 1024 neurons, where 50% are dropped before passing to the final layer using the dropout regularisation method. This layer contains 100 neurons, corresponding to the total number of classes. Subsequently, the Softmax function is applied for classification purposes, which computes the output of the final FC layer into a probability distribution with values ranging from 0 to 1, and the sum of all classes equal to 1.

Both MS-CNN and CNN have two variants, which are with and without batch normalization. Specifically, batch normalization is a regularization method that is used to accelerate training time and improve model accuracy. This improvement is achieved by reducing the internal covariate shift, which is the change in the distribution of inputs to each layer during the training process (Bjorck et al., 2018; Kohler et al., 2018; Santurkar et al., 2018; Ioffe and Szegedy, 2015). Table 1 presents the detailed information for each architecture.

Table 1 Proposed CNN Architectures

2.3.2. Recurrent Neural Networks (RNN)

        RNN is mainly designed for learning time series or sequential data. It can store and learn from sequential data due to the presence of recurrent connections. However, the base RNN often encounters vanishing gradient issues when dealing with long-sequence data. This limitation occurs due to the lack of a gating mechanism that can selectively store, update, or remove information. Hochreiter and Schmidhuber (1997) proposed the Long Short-Term Memory (LSTM) architecture to address the limitation of the base RNN.

The LSTM architecture is comprised of four main components, namely memory cells, input, output, and forget gates. The core component is the memory cell, which is responsible for maintaining the internal state of the unit. The input gate manages and decides the amount of new information from the current input to be stored in the memory cell. The output gate is responsible for deciding which information should be passed to subsequent units and stored in the hidden state. Meanwhile, forget gate decides which information should be stored and removed. Another variant of LSTM, known as bidirectional LSTM, has also been proposed by Graves and Schmidhuber (2005). This variant is a stacked LSTM that receives inputs from both the forward and backward directions, with a similar internal structure as the base LSTM.

A more recent variant of RNN is the Gated Recurrent Unit (GRU), proposed by Cho et al. (2014), which has fewer parameters and uses only two gating mechanisms. In this variant, the update gate decides which information should be kept or removed from the candidate state, serving as a temporary memory space where new information is stored. Meanwhile, the reset gate determines the amount of past information to be removed from the hidden state. Table 2 summarises the information on all RNN architecture variants.

Table 2 RNN Architectures Information

2.3.3. Convolutional Recurrent Neural Networks (C-RNN)

An integrated architecture is introduced that combines convolutional layers for spatial feature extraction with RNN for temporal feature learning. The sequence folding and unfolding layers developed by MathWorks (2022) are used to construct this integrated architecture. A sequence folding layer is added after the input layer, followed by convolutional layers, which are used to extract features from each frame from the image sequences. Subsequently, a sequence unfolding layer is added and convolutional features extracted from the image sequences are flattened and fed into the RNN layers for temporal feature learning. The outputs from the RNN layers are passed to a fully connected layer, where the Softmax function is applied for classification. Figure 8 shows the complete workflow of C-RNN, while the detailed information regarding the architectures is presented in Table 3. The three proposed variants of C-RNN include Convolutional Long Short-Term Memory (ConvLSTM), Convolutional Bidirectional Long Short-Term Memory (ConvBiLSTM), and Convolutional Gated Recurrent Unit (ConvGRU).

Figure 8 Convolutional Recurrent Neural Network workflow

Table 3 Proposed C-RNN Architectures

2.4. Experimental Setup

        In this section, the implementation environments, the initial training of CNN variants, and the experiment with C-RNN are explained in detail.

2.4.1.   Hardware and Software Configuration

        The experiments were conducted in Matlab R2021a on a desktop computer running Windows 11 64-bit operating system. The computer is equipped with an Intel i5-12400f CPU, a base clock speed of 2.5 GHz, 16GB of RAM, and an Nvidia RTX 3060 GPU with 12GB of VRAM and 3584 CUDA cores.

2.4.2.  Training Hyperparameters

        All the models are trained with a learning rate of 0.001, 20 epochs, and a mini-batch size of 32 using the Stochastic Gradient Descent with momentum (SGDM) optimizer. Moreover, the training process started from scratch five times, with the dataset divided into an 80:20 ratio. During each trial, 80% of the samples were randomly selected for training and 20% for testing purposes. Table 4 summarizes the training hyperparameters used in this research.

Table 4 Training Hyperparameter

2.5. Experiment Setting

2.5.1.  Experiment on CNN

        The four proposed variants of CNN architectures, including MS-CNN-A, MS-CNN-B, CNN-A, and CNN-B, are initially trained on the 2D features, namely HGS-MHI. The architecture with the highest accuracy and the least amount of training time is selected as the convolutional layer foundation for constructing the C-RNN.

2.5.2.  Experiment on C-RNN

        The CNN architecture that obtained the optimal performance from the first experiment is selected to construct the integrated architecture of C-RNN. The number of hidden units in the RNN serves as the main hyperparameter value. Experiments are carried out with hidden units of 128, 256, and 512 to identify the optimal value that obtains the highest recognition accuracy with minimal computation time. Table 5 shows the experimental setting for both CNN and C-RNN.

Table 5 Experimental Setting

2.6.  Evaluation Measurement

        The evaluation metrics of accuracy, precision, specificity, recall, and F1-score are used to evaluate the performance of the proposed architectures. Equation 1 describes the calculation of accuracy, which is obtained by dividing the sum of True Positives (TP) and True Negatives (TN) by the total number of test data instances. This serves as the sum of True Positives (TP), True Negatives (TN), False Negatives (FN), and False Positives (FP), measuring the ratio of correctly predicted observations to the total observations. Equation 2 defines precision as the ratio of TP to the sum of TP and FP, measuring the proportion of TP predictions out of all positive predictions made. Meanwhile, Equation 3 defines specificity as the ratio of TN to the sum of TN and FP, which measures the proportion of actual negatives that are correctly identified. Equation 4 defines recall as the ratio of TP to the sum of TP and FN, measuring the proportion of actual positives that are correctly identified. Equation 5 defines the F1-score, which is a weighted average of precision and recall. The measurements are described as follows:

    In this section, the experimental results of averaged classification accuracy and model training computation time are analyzed and discussed in detail.

3.1. CNN Experimental Results

        This section discusses the experimental results of four base variants of CNN to identify optimal convolutional layers for constructing the integrated architectures of C-RNN. As shown in Table 6, the experimental results for both variants of CNN and MS-CNN. MS-CNN-A achieved an accuracy of 95.05%, requiring 1 minute and 56 seconds for computation. MS-CNN-B recorded an accuracy of 93.90%, showing the need for only 1 minute and 35 seconds to train. Although batch normalization is generally used as a regularization method to enhance model accuracy and minimize training time, the first MS-CNN variant with batch normalization consumed more computation time by 21 seconds. Meanwhile, batch normalization also helped to increase the model accuracy by 1.15%.

Table 6 Performance Comparison of MS-CNN and CNN

        The MS-CNN variants are expected to have better performance compared to CNN, due to the presence of parallel convolution layers extracting features at different scales simultaneously. However, both variants of CNN outperform the MS-CNN, achieving an accuracy of 95.15%, which is 0.10% higher than MS-CNN-A. CNN-B obtained an accuracy of 94.50%, which is 0.65% lower than the first variant, requiring a training time of 1 minute and 27 seconds. This significant difference is attributed to the lack of batch normalization, as observed in the MS-CNN results.

        Figures 9 and 10 show the classification accuracy and computation time for all variants of MS-CNN and CNN. These initial experiments are carried out to identify the model with the highest recognition accuracy and the least amount of computational time, resulting in the selection of CNN-A as the convolutional layer for constructing C-RNN. Although CNN-B is the fastest model, requiring the least amount of computation time to train, it is underperformed in terms of accuracy. Both variants of MS-CNN also obtained lower accuracy, leading to the selection of CNN-A as the best architecture, considering that all proposed architectures had training times within the range of 1 minute.

Figure 9 Accuracy comparison of MS-CNN and CNN

Figure 10 Computation time comparison of MS-CNN and CNN

3.2.    C-RNN Experimental Results

        CNN-A with the highest recognition accuracy from the preliminary experiments is selected as the base for convolutional layers in the integrated architecture of C-RNN. The experimental results for three C-RNN variants show a different number of hidden units of 128, 256, and 512, respectively. The optimal value for the hidden units is explored and analyzed to identify the architecture with the highest accuracy with minimal computation time.

        All three variants of the C-RNN achieved greater accuracy compared to CNN architectures, as shown in Table 7. The average computation time also increased from 1 minute to 3 minutes, which was expected as the model complexity rose with the integration of both CNN and RNN architectures. The ConvLSTM, ConvBiLSTM, and ConvGRU models achieved accuracy of 97.20%, 97.40%, and 97.10%, respectively. Regarding computation time, ConvLSTM lasted for 3 minutes and 21 seconds to train, while ConvBiLSTM which used stacked LSTM layers to receive inputs from both forward and backward directions, was expected to require significantly more time. According to experimental results, 3 minutes and 54 seconds were required for training, which was only 33 seconds more than the previous LSTM model. The fastest model among the three variants was ConvGRU, which requires 3 minutes and 10 seconds to train. This result is obtained due to the fewer parameters of the internal structure of GRU compared to LSTM, thereby requiring less computation time.

Table 7 Average Classification Result (%) on 128 Hidden Units

        Table 8 shows a significant rise in average training time, as the number of hidden units increases from 128 to 512. Based on the results, the computation time for ConvLSTM and ConvGRU increases by 32 seconds and 27 seconds, respectively. ConvBiLSTM shows the most significant increase in training time, rising from 3 minutes and 10 seconds to 4 minutes and 56 seconds, due to the bidirectional nature. Although the number of hidden units is set to 256, a total of 512 hidden units is used, where 256 are allocated for the forward pass and 256 for the backward pass. As the computational time increases significantly, there is only a minor improvement in accuracy. The ConvLSTM, ConvBiLSTM, and ConvGRU also show accuracy increase of 0.20%, 0.07%, and 0.30%, respectively.

Table 8 Average Classification Result (%) on 256 Hidden Units

        In Table 9, as the number of hidden units increased from 256 to 512, ConvLSTM accuracy remained the same compared to the previous experiment with 256 units. However, the average training time increased from 3 minutes and 53 seconds to 5 minutes and 3 seconds, showing that a high number of hidden units does not benefit ConvLSTM. ConvGRU showed a modest accuracy increase of 0.05%, requiring an additional 1 minute and 4 seconds to train. The ConvBiLSTM showed the highest observed improvement in accuracy, with the number of hidden units increasing to 512, showing a 0.63% high as the average accuracy rises from 97.47% to 98.10%. This improvement was attributed to the increased training time, which facilitated the slowest architecture among the C-RNN variants.

Table 9 Average Classification Result (%) on 512 Hidden Units

      
    Figure 11 shows an accuracy comparison of all C-RNN variants. Based on the results, ConvBiLSTM with 512 hidden units is the only architecture to achieve the highest accuracy of 98.10%, while other variants range from 97.10% to 97.47%. As shown in Figure 12,  average training time also rises with the increase in the number of hidden units. The architectures with 128, 256, and 512 hidden units fall within an accuracy range of 97%, except for ConvBiLSTM with 512 hidden units. Furthermore, all architectures with 128 hidden units are the most computationally efficient, with an accuracy within the range of 97% and the least amount of training time.

Figure 11 Accuracy comparison of C-RNN

Figure 12 Computation time comparison of C-RNN

In conclusion, this research successfully proposed the use of C-RNN for in-air hand gesture signature (iHGS) recognition, introducing four base CNN architectures. The results showed that MS-CNN-A and MS-CNN-B captured and extracted features at different scales parallelly while the CNN-A and CNN-B were operated at a single scale. CNN-A was selected as the convolutional layer foundation for the integrated architecture, with the highest accuracy of 95.15%. Subsequently, three additional architectures were proposed, namely ConvLSTM, ConvBiLSTM, and ConvGRU. The experimental results showed a significant increase in recognition accuracy by minimizing the loss of information, preserving in a 3D representation, and feeding into the C-RNN models. However, this improvement was obtained due to increased computation time caused by the architectural integration, with the ConvBiLSTM obtaining the highest accuracy of 98.10%, while ConvLSTM and ConvGRU achieved 97.40% and 97.47%, respectively. The average training duration increased from 1 minute to 3 minutes because of the architectural integration in the initial C-RNN experiments with 128 hidden units due to the increased complexity. The training time further increased to 4, 5, and 7 minutes, as the hidden units rose to 256 and 512. The optimal hyperparameter value was 128 hidden units for all variants of C-RNN, with an accuracy of 97% and an average training time of 3 minutes. However, only minor accuracy improvement was observed from 128 to 256 and 512, along with higher computational time.

    The authors are grateful to the Ministry of Higher Education (MOHE) for funding under the Fundamental Research Grant Scheme (FRGS) (FRGS/1/2021/ICT02/MMU/03/3).

Ahad, M.A.R., Tan, J.K., Kim, H., Ishikawa, S., 2012. Motion History Image: Its Variants and Applications. Machine Vision and Applications, Volume 23, pp. 255–281

Akram, M., Qasim, R., Amin, M.A., 2012. A Comparative Study of Signature Recognition Problem Using Statistical Features and Artificial Neural Networks. In: 2012 International Conference on Informatics, Electronics & Vision (ICIEV), pp. 925–929

Azlin, A.I.A.b., Han, P.Y., How, K.W., Yin, O.S., 2022. Hand Gesture Signature Recognition with Machine Learning Algorithms. In: International Conference on Computational Science and Technology, pp. 389-398

Bjorck, N., Gomes, C.P., Selman, B., Weinberger, K.Q., 2018. Understanding Batch Normalization. Advances in Neural Information Processing Systems, Volume 31, pp. 1–12

Carlaw, S., 2020. Impact on Biometrics of Covid-19. Biometric Technology Today. Volume 2020(4), pp. 8–9

Cho, K., Merrienboer, B.V., Gulcehre, C., Bahdanau, D., Bougares, F., Schwenk, H., Bengio, Y., 2014. Learning Phrase Representations using RNN Encoder–Decoder for Statistical Machine Translation. In: Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 1724–1734

Cuceloglu, ?., Ogul, H., 2014. Detecting Handwritten Signatures in Scanned Documents Detecting Handwritten Signatures in Scanned Documents. In: Proceedings of the 19^th Computer Vision Winter Workshop, Volume 2014, pp. 89–94

Davis, J.C., Bobick, A.F., 1997. The Representation and Recognition of Human Movement Using Temporal Templates. In: Proceedings of IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pp. 928–934

Fang, Y., Kang, W., Wu, Q., Tang, L., 2017. A Novel Video-Based System for in-Air Signature Verification.  Computers & Electrical Engineering. Volume 57, pp. 1–14

Faundez-Zanuy, M., 2005. Signature Recognition State-of-The-Art. IEEE Aerospace and Electronic Systems Magazine, Volume 20 (7), pp. 28–32

Gan, K.B., 2022. FACE SHIELD@UKM: An Initiative by UKM To Protect Our Frontliner During Covid-19 Pandemic. International Journal on Robotics, Automation and Sciences, Volume 4, pp. 30–34

Gilperez, A., Alonso-Fernandez, F., Pecharroman, S., Fierrez, J., Ortega-Garcia, J., 2008. Off-line Signature Verification Using Contour Features. In: 11^th International Conference on Frontiers in Handwriting Recognition, pp. 1–6

Graves, A., Schmidhuber, J., 2005. Framewise Phoneme Classification with Bidirectional LSTM Networks. In: Proceedings 2005 IEEE International Joint Conference on Neural Networks, Volume 4, pp. 2047–2052

Hashim, Z., Ahmed, H.M., Alkhayyat, A.H., 2022. A Comparative Study among Handwritten Signature Verification Methods Using Machine Learning Techniques. Scientific Programming. Volume 1, p. 8170424

Hochreiter, S., Schmidhuber, J., 1997. Long Short-Term Memory. Neural Computation, Volume 9(8), pp. 1735–1780

Impedovo, S., Pirlo, G., 2007. Verification of Handwritten Signatures: an Overview. In:14^th International Conference on Image Analysis and Processing (ICIAP), pp. 191–196

Ioffe, S., Szegedy, C., 2015. Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. International Conference on Machine Learning, pp. 448–456

Jain, A., Singh, S.K., Singh, K.P., 2020. Handwritten Signature Verification Using Shallow Convolutional Neural Network. Multimedia Tools and Applications, Volume 79, pp. 19993–20018

Julian, J., Ortega-Garcia, J., 2008. On-Line Signature Verification. Handbook of Biometrics, pp. 189–209

Karanjkar, S.L., Vasambekar, P.N., 2016. Signature Recognition on Bank Cheques Using ANN. In: IEEE International WIE Conference on Electrical and Computer Engineering (WIECON-ECE), pp. 44–47

Khoh, W.H., Pang, Y.H. Yap, H.Y., 2022. In-air Hand Gesture Signature Recognition: An iHGS Database Acquisition Protocol. F1000Research, Volume 11, p. 283

Khoh, W.H., Pang, Y.H., Teoh, A.B.J., 2019. In-air Hand Gesture Signature Recognition System Based on 3-Dimensional Imagery. Multimedia Tools and Applications, Volume 78(6), pp. 6913–6937

Khoh, W.H., Pang, Y.H., Teoh, A.B.J., Ooi, S.Y., 2021. In-air Hand Gesture Signature Using Transfer Learning and Its Forgery Attack. Applied Soft Computing, Volume 113, p. 108033

Kohler, J., Daneshmand, H., Lucchi, A., Zhou, M., Neymeyr, K., Hofmann, T., 2018. Towards A Theoretical Understanding of Batch Normalization. Stat, Volume 2018, pp. 1–33

Lakshmi, K., Devendran, T., 2021. Human Fall Detection Using Motion History Image and SVM. In: Advanced Informatics for Computing Research: 4th International Conference, ICAICR 2020, Gurugram, India, December 26–27, 2020, Revised Selected Papers, Part I 4, pp. 466–476

Levy, A., Nassi, B., Elovici, Y., Shmueli, E., 2018. Handwritten Signature Verification Using Wrist-Worn Devices. In: Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, Volume 2(3), pp. 1–26

Malik, J., Elhayek, A., Ahmed, S., Shafait, F., Malik, M., Stricker, D., 2018. 3DAirSig: A Framework for Enabling In-Air Signatures Using a Multi-Modal Depth Sensor. Sensors, Volume 18(11), p. 3872

MathWorks, 2022. Classify Videos Using Deep Learning - MATLAB & Simulink. Available online at: https://www.mathworks.com/help/deeplearning/ug/classify-videos-using-deep-learning.html, Accessed on July 17, 2022

Romadlon, F., Lestiana, F., Putri, N.A., 2022. An Exploration of Personal Decision as Mediating Effect between Passenger Concern and Airport Service Information During COVID-19 Outbreak. International Journal of Technology, Volume 13(3), pp. 664–676

Sajid, H., Cheung, S., 2015. VSig: Hand-Gestured Signature Recognition and Authentication with Wearable Camera. International Workshop on Information Forensics and Security (WIFS), pp. 1–6

Santurkar, S., Tsipras, D., Ilyas, A., Madry, A., 2018. How Does Batch Normalization Help Optimization. In: Advances in Neural Information Processing Systems, Volume 31, pp. 1–11

Shao, Y., Yang, T., Wang, H., Ma, J., 2020. AirSign: Smartphone Authentication by Signing in the Air. Sensors, Volume 21(1), p. 104

Sincan, O.M., Keles, H.Y., 2022. Using Motion History Images with 3D Convolutional Networks in Isolated Sign Language Recognition. IEEE Access, Volume 10, pp. 18608–18618

Sudharshan, D.P., Vismaya, R.N., 2022. Handwritten Signature Verification System using Deep Learning. In: International Conference on Data Science and Information System (ICDSIS), pp. 1–5

Supriatna, Zulkarnain, F., Ardiansyah, Rizqihandari, N., Semedi, J.M., Indratmoko, S., Rahatiningtyas, N.S., Nurlambang, T., Dimyati, M., 2022. Communicating the High Susceptible Zone of COVID-19 and its Exposure to Population Number through a Web-GIS Dashboard for Indonesia Cases. International Journal of Technology, Volume 13(4), pp. 706–716

Yatmo, Y.A., Harahap, M.M.Y., Atmodiwirjo, P., 2021. Modular Isolation Units for Patients with Mild-to-Moderate Conditions in Response to Hospital Surges Resulting from the COVID-19 Pandemic. International Journal of Technology, Volume 12(1), pp. 43–53