Scalable Facial Authentication System for a Distributed Environment Using Feature Dimensionality Reduction

doi:10.21203/rs.3.rs-2472355/v1

Download PDF

Research Article

Scalable Facial Authentication System for a Distributed Environment Using Feature Dimensionality Reduction

https://doi.org/10.21203/rs.3.rs-2472355/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

The user authentication of personal identification using a face image relies on biometric features known as facial features. A facial authentication system is critical for access control and intelligent human computer interaction. However, the challenges include numerous training data requirement and high processing power due to computational complexity. Therefore, this paper proposes a scalable face authentication system for a distributed environment using feature dimensionality reduction (FDR). This study contributes to a constructive comparison of classical and state-of-the-art feature extraction techniques to develop of distributed environment-based scalability using FDR. The scalability for the distributed environment, is proven to extend the user access capacity by overcoming the multiple training data requirements while the new end-user access is authenticated in system. The accuracy and processing time were chosen as the performance metrics evaluating the proposed method. The features extracted using deep convolutional neural network (DCNN) as the state-of-the-art algorithm reveal remarkable performance using FDR with the principal component analysis algorithm and support vector machine for classification, compared to classical feature vector extractor algorithms such as the Gabor filter, histogram of oriented gradient and local binary pattern. The experimental results on the FERET dataset, demonstrate that the feature extracted by the DCNN is a more robust representation of scalability with up to 50% for FDR using half the processing time compared to the facial feature extracted using classical methods.

Scalability

feature dimensionality reduction (FDR)

deep convolutional neural networks (DCNN)

facial features

principal component analysis (PCA)

Facial detection and recognition are preliminary steps to a broad range of applications including advanced human and computer interaction. A facial recognition (FR) system allows user authentication by simply verifying user accessibility using a camera. Various researches have been devoted to FR for years and have proposed algorithms to be as accurate as possible for the normal human state. An FR system is expected to automatically identify faces present in images or videos and can operate in either or both of two modes: facial authentication and recognition

Facial feature-based authentication algorithms have been developed, and their accuracy continues improving. They have become the preferred method over other biometric modalities. Identifying a person’s face accurately and effectively has always been an interesting research idea, both in research and industry [2]. It also has a high potential for used in public facilities, e-commerce, e-education, security, and other area. The primary factor that makes facial feature authentication algorithms a more interesting topic than others, is their non-intrusive nature. Compared to other biometric modalities such as voice recognition which requires the user to speak loudly [8] and fingerprint recognition which requires the user to hold a finger on a sensor [11], the use of facial features for FR systems make the applications simple and more user friendly. However, the challenges include the considerable training data requirements and high processing power due to computational complexity. In addition, compared to state-of-the-art facial extraction algorithms, classical facial feature extraction algorithms may lead to a lack of extracted feature information [3] while reducing the feature dimensions. Thus, the classification accuracy decreases due to matching features that are not well categorized and the limited user-access capacity of the system.

For this purpose, we exploit a combination of methods for facial detection, facial feature extraction and feature dimensionality reduction (FDR). In this paper, we introduce the novel approach of scalability in the distributed environment for authentication systems based on the convolutional neural network (CNN) model as a state-of-the-art method for facial feature extraction that increases the robustness of the method to extend user accessibility overcoming extensive training data requirements, high processing power and the computational complexity challenges. Furthermore, the main contribution focuses on facial feature extraction methods that can match the scalability technique to extend user accessibility while preserving recognition accuracy. Embedded hardware platforms such as the Raspberry Pi and a server serve distributed environment purposes and contribute to the system, offering costless, lightweight, low power that can meet the authentication requirements.

The rest of this paper is organized as follows: Section 2 reviews the theory of the related work on detection and feature extraction methods that can be considered for the scalability technique and widely used face features for FR purposes. Next, Section 3, outlines and details FDR algorithms and Section 4, presents the scalable face authentication structure as the proposed approach. Then, Section 5 discusses the extensive and comparative experimental results that demonstrate the efficiency and effectiveness of the proposed feature scalability algorithm. Finally, Section 6 presents the conclusion.

This section introduces two main tasks: selecting the facial area and extracting features from that area. The process is coherent, from detecting and cropping the facial areas to reducing the data size. The cropped facial images are resized for consistency to the defined size. In the proposed method, we use frontal facial detectors based on MTCNN that applies a cascade of neural networks. Primarily, FR systems use the Haar cascade in OpenCV as a facial image detector. However, Haar cascades have many issues such as low accuracy when applied to real-time facial image detection with varied lighting [32]. In this paper, MTCNN fits the proposed method because it can run in real-time. The MTCNN consists of three neural networks connected by stages in a cascade as illustrated in Fig. 1; P-Net, R-Net and O-Net, each have a particular task to improve detection accuracy [32] [33]. The P-Net produces candidate windows a shallow convolutional network, whereas R-Net rejects as many false-face (non-facial) windows as possible and O-Net refines the R-Net output.

Based on the structural performance, MTCNN performs the following:

It has a considerable execution processing speed providing neighborhood dataset information that satisfies the task and power scaling to transfer data. Specifically, in the proposed method, after facial detection, the region of interest of facial image is cropped out and resized to 512x768 for convenience. Then a bounding box trims the facial parts for the next step of feature extraction.
It extracts relevant data from an extracted facial image area, where such data have a more significant representation of the related image in the classification process with an acceptable error rate.

For facial detection, the MTCNN algorithm improves detection accuracy and completes the affine transformation and trimming of the image. Thus, this algorithm increases the facial feature extraction performance gained from MTCNN compared to other facial detection algorithms.

The approach in this paper is to exploit various facial feature extraction algorithms such as Gabor filter, histogram orthogonal analysis (HOG), local binary pattern (LBP) and deep CNN (DCNN) methods that are useful for FR algorithms. The Gabor filter [4], HOG) [5] and LBP [6] are classical methods based on hand crafted facial feature extraction including texture descriptors and edges [7] and require more efforts in terms of computational cost for FDR [10]. However, using state-of-the-art facial feature extraction methods such as the DCNN, has an impressive gain in performance making these methods popular. Their novelty stems from their ability to extract meaningful presentations from image data known as deep features. The DCNN has the considerable advantages of high performance at an extremely low power consumption because the computationally intensive tasks of the DCNN can run on graphics processing unit (GPU) [10] [15].

2.1. Gabor Filter

For the Gabor filter, we use the Gabor wavelets technique to extract features for recognition by applying two dimensional (2D) Gabor kernels. A set of Gabor filters is used with five spatial frequencies and eight distinct orientations [21], resulting in 40 Gabor filters, as displayed in Fig. 2.

$$\left\{\begin{array}{c}W\left(x,y,\theta ,\delta , \phi ,\sigma ,\gamma \right)= {e\left(-\frac{{x}^{{\prime }2}+{\gamma }^{2}{y}^{{\prime }2}}{2{\sigma }^{2}}\right)\text{cos}2\pi \frac{{x}^{{\prime }}}{\delta }}^{+}\phi \\ {x}^{{\prime }} =x\text{cos}\left(\theta \right)+y\text{sin}(\theta )\\ {y}^{{\prime }} = -x\text{sin}(\theta )+y\text{cos}(\theta )\end{array}\right. \left(1\right)$$

where $(x, y)$ indicate the position of a light impulse in the visual field and $\mu , \phi , \sigma , \delta , \gamma$ are parameters of the wavelet [21] [22].

Referring to the feature extraction-based intelligent recognition systems, Ahsan, M., et al. [31] proposed an intelligent system for automatic fingerprint identification using feature fusion with Gabor filter and deep learning. The intelligent computational approach automatically authenticates fingerprints for personal identification and verification. The fusion of the Gabor filtering technique and CNN formed the feature vector and the PCA was performed to accurately reduce the overfitting problems to improve the classification results. By training the extracted features using a multiclass classifier, the performance on standard public databases demonstrated an accuracy of 99.87% compared to recent classifier, such as the support vector machine (SVM) at 97.86%. The Fusion of the CNN and Gabor filter features is one of the critical contributions to improving the accuracy of fingerprint images classification and recognition.

2.2. Local binary patterns

The LBP was initially used to extract texture features on the object surfaces, but later research applied it to facial feature extraction and the field of 2D FR achieving good results [26]. In this paper, the LBP in the comparative performance results was considered a feature extractor for the proposed scalable facial authentication system for FR, because of its strong advantage of robustness to light and resistive characteristics when the shooting and viewing angles change considerably [26] [27]. The basic concept of the LBP is to use the pixel value of the most central point as the threshold value make the neighboring pixels binary numbers, as presented below.

Schematic diagram of the LBP square neighboring conversion concept; In this case the binary number is 10101000 and the decimal number is 168 and the general concept of the LBP is expressed as;

$$LBP\left({x}_{c},{y}_{c}\right)= \sum _{p=0}^{p-1}{2}^{p}S\left({i}_{p}-{i}_{c}\right) \left(2\right)$$

$$S\left(x\right)= \left\{\begin{array}{c}1 x\ge 0\\ 0 x<0\end{array}\right. \left(3\right)$$

In Fig. 3, the summarized LBP calculation process is presented where each pixel in the facial image is considered the center point and takes a 3x3 size area around the pixel as the neighborhood. By comparing whether the pixel is greater than or equal to the center pixel value for the other eight-pixel values in the adjacent area, the pixel value is set to 1; otherwise, it is set to 0 [28]. Thus, an 8-bit sequence of 0s and 1s is generated to form an 8-bit unsigned integer binary value. The generated binary number from the calculation is the LBP eigenvalue in the center of the widow. The LBP is improved by making the LBP model operator to have rotation invariance, whereas the original limited area is expanded to any size. In this case, because the radius of the circular neighborhood is variable, there can be any number of pixels in it, and the circle has the characteristic of rotation invariance [26] [30]. Thus, the original square neighborhood is replaced by a circular neighborhood. For example, we consider the data in Fig. 3, and the coordinate of any point in the neighborhood can be calculated by

$${x}_{p}= {x}_{c}+ \text{R cos}\left(\frac{2\pi p}{N}\right) \left(4\right)$$

$${y}_{p}= {y}_{c}-R\text{sin}\left(\frac{2\pi p}{N}\right) \left(5\right)$$

From Eqs. 4 and 5, ${(x}_{c},{y}_{c})$ is the center point of the area and $({x}_{p},{y}_{p})$ is a sampling point in the neighborhood.

2.3. Deep convolutional neural network

The CNN has significantly improved state-of-the-art technology for many applications in computer vision. The classical feature extraction method can be replaced with CNN algorithms, specifically, the DCNN extracts complex features that express the image in much more details, can learn tasks specific features and is much more efficient. The DCNN uses local connections to effectively extract special information and shared weights to significantly reduce the number of parameters [34]. The DCNN employs several convolutional and pooling layers to extract features from the detected user access which are nonlinear, discriminant, invariant and more productive for image classification. The pre-learned features are used as filters and these filters convolves through the input image and produces features used by other network layers. The computationally intensive tasks of the DCNN can run on GPU, resulting in high performance at minimal power consumption. Thus, facial features extracted using DCNN frameworks, have been proposed for scalability in the distributed environment.

For this paper, among the different approaches of the CNN, the transfer learning approach with features from pre-trained CNN model can be considered in this implementation. The primary reason is that the goal is to extend user-access accessibility when dataset is large. Therefore, the CNN can be used via transfer learning to maintain the convolutional base in its original form where the output is input into the FDR stage and the results successfully realize scalability. Another considered approach is learning the model from scratch where the architecture of the pre-trained model is used and trained based on the dataset. The pre-trained model is considered a fixed feature extraction mechanism in this case. The pre-trained CNN model such as AlexNet or ResNet-50, is applied to analyze the performance of scalable facial authentication. The AlexNet model, uses a 277x277x3 image as an input and the rectified linear units activation function is applied to extract suitable features [29]. It includes five convolutional layers, three max-pooling layers, and three fully connected layers. In this case a 4096-dimensional feature vector represent the 277x277 image. The ResNet-50 model, also known a deep residual networks, is the latest popular pre-trained model, used for image recognition applications. The ResNet-50 architecture includes five convolutional layers and uses a 224x224x3 image as input.

The FDR process is an effective approach to downsizing features. In either case, the goal is to determine a low-dimensional representation of facial features while describing the features with sufficient accuracy. The FDR techniques overcomes the curse of dimensionality by transforming n-dimensional feature data into a k-dimensional subspace $(n>k)$. In this paper, the FDR technique is the primary point for scalability, and various algorithms for the FDR approach are compared, such as PCA, linear discriminant analysis (LDA) and singular value decomposition (SVD).

3.1. Linear Discriminant Analysis

The LDA is a supervised dimension reduction technique, that aims to determine components axes that maximize the class separation in the dataset. The LDA algorithm is realized by calculating three parameters: between-class variance, within-class variance, and the sub-space parameter. The between-class variance expresses the separability of different classes and is computed using the distances between the means of the different classes as indicated in Eq. (6). The within-class variance is calculated using the mean and samples of each class as expressed in Eq. (7):

$${S}_{b}=\sum _{i=1}^{j}{n}_{i}({\mu }_{i}-\mu )\left({{\mu }_{i}-\mu )}^{T} \right(6)$$

$${S}_{w}=\sum _{i}^{j}{n}_{i}({x}_{i}-{\mu }_{{k}_{i}})({x}_{i}-{\mu }_{{k}_{i}}{)}^{T} (7)$$

The ratio of the between-class variance and within-class variance is computed to determine the sub-space $W$. The optimal $W$value ${W}_{opt}$ is expressed in Eq. (8) as follows:

$${W}_{opt}= {argmax}_{w}\frac{\left|{W}^{T}{S}_{b}W\right|}{\left|{W}^{T}{S}_{w}W\right|} \left(8\right)$$

3.2. Principal Component Analysis

The PCA can express information in multiple datasets with fewer variables and minimal information loss. Thus, a dataset with numerous of interrelated variables is reduced a smaller size. After transformation, the output of the obtained variables comprises the principal components of the first variables.

In this case, PCA serves as the critical point of scalability by seeking a parameterized model of the form $\text{X}=\text{m}\text{o}\text{d}\text{e}\text{l} \left({\beta }\right)$ where ${\beta }$ represents a vector of the model parameter, and that model can be used to generate new vectors X. Thus, if we can obtain a model with the distribution of parameters${p}_{\beta } \left({\beta }\right)$, it is possible to limit them, so that generated models Xs are similar to the same model as the training set. Thus, it should also be possible to estimate${p}_{X} \left(\text{X}\right)$ using the model. The PCA approach for dimensionality reduction is realized in the following steps.

Step 1: Compute the mean of the data-set

$${\phi }=\frac{1}{s}\sum _{i=1}^{s}{X}_{i} \left(9\right)$$

Step 2: Compute the covariance matrix of the data set

$$\text{S}= \frac{1}{s-1}\sum _{i=1}^{s}{({X}_{i}-\phi )({X}_{i}-\phi )}^{T} \left(10\right)$$

Step 3: Compute the eigenvectors, ${p}_{j},$ and corresponding eigenvalues, ${\sigma }_{j}$, of the dataset S from the covariance matrix

Step 4: Approximate X and, input the value from the training set to select the feature vector

$$\text{X}\approx {\phi }+ {P}_{s}{b}_{s} \left(11\right)$$

Step 5: Multiply the transpose of the feature vectors by the transpose of the adjusted data.

The main advantages of the PCA technique are the reduced memory and capacity requirements, low sensitivity to noise, and effectiveness of operations in low-dimensional spaces. Due to these advantages, The PCA is given as input to machine learning algorithms to obtain a reduced feature vector easily.

3.3. Singular Value Decomposition

The SVD is one of the most widely used algorithms for dimension reduction. It is also a powerful tool for matrix decomposition, widely used in many areas including image processing. Technically, SVD is a procedure that allows one to represent data in a new sub-feature space, capturing the majority of variations in the data. This procedure is achieved by rotating the axes of the original feature space to form new axes which are linear combinations of the original axes/features. The new axes are helpful for the FDR process because they systematically break down the variance in the data points based on each direction of the data variance. The technical details of SVD are defined as follows:

Representation of any matrix A as a product of three matrices; ${A}={U}{\epsilon }{{V}}^{{T}}$ where U and V are rotation matrices, and $\epsilon$ is a diagonal scaling matrix.
Computation of the eigenvectors ${A}{{A}}^{{T}}$. Which are the columns of U; thus, the square roots of the eigenvalues are singular values.
Computation of the eigenvectors of ${{A}}^{{T}}{A}$. which are columns of V or rows of${{A}}^{{T}}$

Based on the theoretical approach of the various dimensionality reduction methods, FDR is evaluated through the comparative results presented in Section 5.

3.4. Facial Recognition System

Facial recognition is one of the most effective and relevant applications of image processing and biometric systems. In Fig. 4 every biometric-based FR system has four main parts: facial detection and alignment, feature extraction, face classification, and FR [17].

In the experiment, the FERET dataset [20], contains 11,338 face images with a size of 512x768 pixels captured with 13 poses from 994 subjects. First, we evaluate the feature extraction time with feature dimension size of the feature extractors; the Gabor filter, LBP, HOG (classical feature extraction methods), and DCNN (the state-of-the-art feature extraction method). With credit-sized and full capabilities like a computer, the Raspberry Pi, was used at the user-access point to detect and extract facial features from captured face images and distribute them to the server for FDR and authentication processing.

This paper focuses on developing a distributed environment approach with scalability using a combination of existing conventional FR methods that contributes to reducing the training data requirements and processing power due to computational complexity. The critical issue is SFAS, because the problem can include the simultaneous FR of numerous user-access points. Thus, the need for the intelligent and dynamic rebuilding of the dataset models whenever a new user is added can compromise the system performance in terms of processing time flexible user-access extension. Additionally, the scalability in a distributed environment complementary reduces processing time and power, contributing to an authentication system for modern infrastructures.

4.1. Distributed environment

The concept of an SFAS for a distributed environment using the FDR introduced in this paper is designed based on embedded artificial intelligence (AI). The first step in the system design is to perform the distributed environment and decompose the goals dividing them into functional modules according to the purpose of the design as illustrated in Fig. 5. The purpose of the scalability-based authentication system is to extend the user-accessibility by reducing the time processing preserving authentication accuracy.

The characteristics of scalability to the distributed environment are evaluated by considering the computational analysis of the processing time for new end-user access and the storage cost with respect to the database size. The facial feature extraction for a distributed environment using the FDR approach is evaluated, where the concept of FDR was considered for a scalable face image. As the number of face image in the dataset increases, the complexity of representing the dataset increases. By adjusting the FDR scales proportional to the increase in end-user accessibility referring to the authenticated user’s face dataset, the facial feature extraction descriptor plays a critical role in realizing scalability in a distributed environment.

4.2. Scalable face feature

A distributed environment connection transforms the facial data between two embedded platforms due to the integrity of the facial feature extraction and dimensionality reduction by scales. Referring to a comparative analysis of the experimental results from various algorithms, as stated in Section 5, the primary algorithms used for facial detection and the feature extraction module are the MTCNN and DCNN, respectively. For FDR and classification, the matching algorithms are PCA and SVM. For the embedded platforms Raspberry Pi and standard computer are used as the hardware acceleration platform of the system due to the low power consumption and the efficient operating cost. By putting them together, the proposed SFAS for distributed environment using FDR workflow is presented in Fig. 6.

The concepts of vector space and subspace to scalability are applied. The set of authenticated user faces with MxN pixels forms a subspace called the "scalable face space" of the user face space (dataset) containing all images with MxN pixels. This scalable face space defines the variation of the authenticated user faces. The projection of a new user face (non-authenticated) onto the face space is compared to the available projections of the scalable face space to authenticate the projected new user access. The dimension of the scalable face space is much lower than the entire image, thus, it is much easier to compare projections than in the original images, pixel by pixel.

4.3. Application scenarios

This paper aims to contribute to the functionality of FR performance by introducing scalability into a distributed environment to extend user accessibility and improve the processing time by preserving the recognition accuracy. For the proposed SFAS in this paper, we proceed with two approaches, as displayed in Fig. 6. The first approach is applying the state-of-the-art pre-trained DCNN algorithm for extracting facial features from the end user-access. The second approach is reducing the extracted facial feature dimension using PCA from the server and then retraining the model.

In this paper, we followed the following stages. First, facial feature extraction, in wherein we resized each image into a suitable size for each DCNN model and employed a pre-trained CNN to represent the face for user-access. The ResNet-50 was used as a pre-trained CNN. The DCNN were used to extract suitable image features and use them in the following classification stage. In the second stage, the extracted facial features were processed for FDR before training the model from the server. The experimental tests were conducted using the FERET dataset. Then, we examined the results of various algorithms analyzed the effectiveness of each approach and compared the results. Finally, we confirmed using the DCNN algorithm as a facial feature extractor, the PCA algorithm for FDR and the SVM as a classifier to recognize faces because of its observable classification result on nonlinear data. The SVM has many advantages in solving pattern recognition and machine learning problems such as FR and function overfitting [37]. In this paper, dividing tasks by introducing the distributed environment has improved the processing speed and training data requirements when the system gains a new user-access point. Instead, the server can simultaneously retrain the model without affecting the end-user accessibility which is a different case for the existing end-to-end FR systems.

this section, analyzes the combination of existing conventional methods of FR for SFAS based on various comparative analysis results. In actual scenarios, the scalability structure in a distributed environment for an authentication system is achieved based on facial feature extraction and FDR methods, which are often accompanied by substantial calculation [13] [14]. Two embedded platforms (Raspberry Pi on the end-user side and a standard computer on the server side) serve as distributed environment platforms to improve the easy process of facial feature extraction and FDR tasks. In the experiments, the FERET dataset [20], that contains 11,338 face images with an image size of 512x768 pixels captured with 13 poses. The specifications of the embedded platform hardware are as follows:

Raspberry Pi on the side of the end-user-access with Broadcom BCM2711, Quad-core Cortex-A72 (ARM v8) 64-bit SoC @ 1.5GHz, 8GB LPDDR4-3200 SDRAM, and 2.4 and 5.0 GHz IEEE 802.11ac wireless, Bluetooth 5.0, BLE gigabit Ethernet.
A standard computer used as a server: AMD Ryzen 5 2600, 6-core processor with NVIDIA GeForce GTX 1650 GPU with VS code, WSL, and TensorFlow-GPU = = 1.15 installed.

Facial feature extraction is crucial stage in this paper, because we address FDR to realize scalability in the distributed environment. Choosing a feature descriptor that helps determine the robustness and performance of scalability for authenticated systems is critical. We consider feature descriptors, such as the Gabor filter, LBP, HOG and DCNN. To evaluate the effectiveness and efficiency of the feature descriptors, we also considered the FR accuracy by comparing the results of various classifiers such as the SVM, k nearest neighbor and artificial neural network. From the classification stage, the authentication is validated for the end-user access through the adopted classifier that trains and tests extracted features to validate each detected end-user face in the corresponding dataset. The performance of feature descriptors in terms of accuracy and processing time with various classifiers is presented in Tables 1 and 2.

Table 1

Recognition-accuracy feature descriptors performance of classifiers by authentication system user access
Feature descriptors	Classifier-Processing time (ms)
Feature descriptors	SVM	k-NN	ANN
Gabor filter	302	410	290
HOG	431	503	378
LBP	387	479	281
DCNN	178	237	238

Table 2

Processing-time feature descriptors performance of classifiers by authentication system user access.
Feature descriptors	Classifier-Processing time (ms)
Feature descriptors	SVM	k-NN	ANN
Gabor filter	302	410	290
HOG	431	503	378
LBP	387	479	281
DCNN	178	237	238

We evaluate the feature descriptor performance based on recognition accuracy and processing time. By analyzing the comparative results of the presented feature descriptors with various classifiers on the FERET dataset, the combination of the DCNN and SVM algorithm display promising results with the highest authentication accuracy compared to other feature descriptors and classification algorithms.

Furthermore, we compared the results of various FDR methods such as PCA, LDA and SVD. To gain more insights into the FDR performance of the proposed method, we conducted experiments by applying various dimension reduction algorithms to the extracted facial features with dimension reduction scales of 10%, 20%, 30%, 40% and 50%, the performance results are provided in Fig. 7.

Based on the presented experimental results in Fig. 10, the combination of the DCNN with PCA can be adopted for a scalability approach in the distributed environment. Thus, this method inherits properties of the scalability approach. For example, features are extracted from the original face image, and the of the FDR schemes performance is significantly influenced, revealing that the accuracy of scalable faces is preserved for features extracted using the DCNN.

In this paper, the purpose of FDR and the distributed environment approach is to improve the processing time to recognize end-user access, which is evaluated in Table 3. The recognition accuracy is preserved for the features extracted using the DCNN by up to 50% of FDR compared to other feature descriptors. The processing time reduces sharply with the preserved recognition accuracy.

Table 3

Performance results of SFAS for a distributed environment using feature dimension reduction
Feature dimension reduction (%)	Recognition accuracy (%)	Recognition time per user (ms)
0	98.9	378
10	98.4	321
20	97.5	256
30	96.1	182
40	95.3	134
50	91.6	114

As describe in Fig. 6, the SFAS framework based on this scalability and distributed environment, consists of feature extraction for user-access faces, facial FDR, and classification for authentication. By comparing the experimental results in Tables 1, 2 and 3 for feature extraction, we adopted a DCNN-based feature extraction approach. The concept behind this adoption is its descriptive resistance to dimensionality reduction using PCA because it helps preserve the classification accuracy using the SVM, with a low computational cost. Thus, the combination of the DCNN, PCA and SVM is connected to the proposed scalability algorithm in a distributed environment.

Table 4

Response of the proposed method (SFAS) compared with other related research on FERET dataset
Research	Schemes	Accuracy (%)	Processing Time (ms)
Xue Lv et al. 2021 [34]	LCNN based on R, C-E, and MFGP modules	83.5	415
S.Sharma et al.2020 [35]	PCA + SVM	96.4	398
Sahan. J.M et al. 2021 [36]	1D-CNN + LDA	89.7	276
Proposed method	DCNN + PCA + SVM with FDR of 50%	91.6	114

As described in Table 4, the experimental results from other related research demonstrate that the proposed approach using the existing combined FR algorithms can achieve better or comparable recognition performance. Furthermore, the SFAS for a distributed environment using FDR was designed to demonstrate that scalability is useful for reducing the FR operation time and preserving the flexibility of the user-access capacity, while maintaining competitive FR performance compared to other methods. Our findings reveal that the Deep Convolutional Neural Network–extracted feature is a stronger illustration of scalability for FDR by using half of the processing time when compared to the facial feature extracted by classical methods.

Using Raspberry Pi and a server as the embedded platform hardware for facial detection, feature extraction and FDR has provided a new orientation for research on image processing and computer vision in the field of embedded (AI). Facial feature extraction and FDR using combined conventional FR algorithms were analyzed to prove the scalability of authentication systems in a distributed environment. The detected user-access faces under the MTCNN algorithm, and facial feature extraction using DCNN with the PCA algorithm as FDR and the SVM as a classifier are interesting and efficient combined algorithms of scalability in a distributed environment proposed for FR applications. With the FERET dataset, the experimental results reveal that the accuracy of features extracted using DCNN algorithm can be preserved from 98.4–87.1% with FDR scales from 10–50% of the user-access extension capacity. For the processing time, the scalability in the distributed environment approach saves approximately half of the processing time for the authentication of each user-access point compared to other methods. Overall, the purpose was to extend the user accessibility capacity, reducing the processing time by preserving the recognition accuracy using costless, lightweight, and low power that can meet authentication requirements. Most research on FR applications has employed the fusion of features to improve recognition accuracy yet the processing time continues to increase significantly with the user-access capacity limitation.

Wang H, Jiani H, Deng W. (2009) Face feature extraction: a complete review. IEEE Access 6: 6001-6039.
Ali, M MH, Mahale VH, Yannawar P, Gaikwad AT. (2016) Overview of fingerprint recognition system. In: 2016 International Conference on Electrical, Electronics, and Optimization Techniques (ICEEOT), pp. 1334-1338. IEEE
Ahsan M, Based MA, Haider J, Kowalski M. (2021) An intelligent system for automatic fingerprint identification using feature fusion by Gabor filter and deep learning. Computers & Electrical Engineering. 95:107387.
Barbu, T. (2010) Gabor filter-based face recognition technique. Proceedings of the Romanian Academy 11: 277-283.
Abhishree, TM, Latha, J, Manikantan, K. Ramachandran, S, (2015). Face recognition using Gabor filter-based feature extraction with anisotropic diffusion as a pre-processing technique. Procedia Computer Science, 45: 312-321.
Ojala, T, Pietikainen, M, & Maenpaa T. (2002) Multiresolution gray-scale and rotation invariant texture classification with local binary patterns. IEEE Transactions on Pattern Analysis and Machine intelligence 24: 971-987.
Trigueros DS, Meng L, Hartnett M, (2018) Face recognition: From traditional to deep learning methods. arXiv preprint arXiv:1811.00116.
Lei Y, Scheffer N, Ferrer L, McLaren M, (2014) A novel scheme for speaker recognition using a phonetically aware deep neural network. In 2014 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (pp. 1695--1699). IEEE.
Ma J, Yuan Y, (2019) Dimension reduction of image deep feature using PCA. Journal of Visual Communication and Image Representation, 63: 102578.
Mayya, V, Pai RM, Manohara Pai MM (2016) Automatic facial expression recognition using DCNN. Procedia Computer Science 93: 453-461.
Krizhevsky A, Sutskever I, Hinton GE, ImageNet classification with deep convolutional neural networks. In: Pereira F, Burges C, Bottou L, Weinberger K, (eds) Advances in Neural Information Processing Systems 25. Curran Associates, Inc. 1097–1105.
Li Y K, Wu X J, Kittler J. (2019) L1-2D2PCANet: a deep learning network for face recognition. Journal of Electronic Imaging 28: 1.
Tadikonda B, Prabu AV, Rajasoundaran S, Raghava Rao K, Aruna Sri SG (2018) Secured door lock system based on fingerprint authentication. Journal of Advanced Research in Dynamical and Control Systems. 10 (02): 473-480.
Sawhney S, Kacker K, Jain S, Singh SN, Garg R, (2019) Real time smart attendance system using face recognition techniques. IEEE 2019 9^th International Conference on Cloud Computing, Data Science & Engineering (Conference), pp. 522-525, doi: 10.1109/CONFLUENCE.2019. 8776934.
Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, et al. (2011) Scikit-learn: Machine learning in Python. Journal of Machine Learning Research 12:2825–2830.
Lu GF, Zou J, Wang Y (2012) Incremental complete LDA for face recognition. Pattern Recognition 45 (7): 2510-2521.
Mai G, Cao K, Yuen PC, Jain AK (2018) On the reconstruction of face images from deep face templates. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(5):1188–1202. doi: 10.1109/TPAMI.2018.2827389
Zeiler M, Fergus R, (2014) Visualizing and understanding convolutional networks. In: Proceedings of the European Conference on Computer Vision, pp. 818–833.
Kasar M, Bhattacharyya D, Kim TH (2016) Face recognition using neural network: a review. International Journal of Security and Its Applications 10: 81-–100.
http://www.itl.nist.gov/iad/humanid/feret/feret_master.html
Dora L, et al. (2017) An evolutionary single Gabor kernel-based filter approach to face recognition. Engineering Applications of Artificial Intelligence, 62: 286-–301.
Li C, Huang Y, Xue Y (2019) Dependence structure of Gabor wavelets based on copula for face recognition. Expert Systems with Applications 137:453-–470.
Jemaa YB, Khanfir S (2009) Automatic local Gabor features extraction for face recognition. arXiv preprint arXiv:0907.4984.
Shi L, Wang X, Shen Y (2020) Research on 3D face recognition method based on LBP and SVM. Optik 220: 165157.
Fontes CH, Embiruçu M (2021) An approach combining a new weight initialization method and constructive algorithm to configure a single Feedforward Neural Network for multi-class classification. Engineering Applications of Artificial Intelligence 106: 104495.
Shi L, Wang X, Shen Y (2020) Research on 3D face recognition method based on LBP and SVM. Optik 220: 165157.
Abbas HH, Ahmed BZ, Abbas AK (2019) 3D face factorisation for face recognition using pattern recognition algorithms. Cybernetics and Information Technologies, 19:28–37.
Hamouda M, Ettabaa KS, Bouhlel MS. (2020) Smart feature extraction and classification of hyperspectral images based on convolutional neural networks. IET Image Processing. 14:1999-–2005.
Almabdy S, Elrefaei L (2019) Deep convolutional neural network-based approaches for face recognition. Applied Sciences 9:4397.
Faudzi SAAM, Yahya N (2014) Evaluation of LBP-based face recognition techniques, 2014 5th International Conference on Intelligent and Advanced Systems (ICIAS) 1–6 doi:10.1109/ICIAS.2014. 6869522.
Ahsan M, Based MA, Haider J, Kowalski M (2021) An intelligent system for automatic fingerprint identification using feature fusion by Gabor filter and deep learning. Computers & Electrical Engineering 95:107387
Rahmad C, Asmara RA, Putra DRH, Dharma I, Darmono H, Muhiqqin I (2020) Comparison of Viola-Jones Haar Cascade classifier and histogram of oriented gradients (HOG) for face detection. In: IOP Conference Series: Materials Science and Engineering (vol. 732, no. 1, p. 012038). IOP Publishing.
Meivel S, Devi KI, Selvam TM, Maheswari SU (2021) Real time analysis of unmask face detection in human skin using tensor flow package and IoT algorithm. Materials Today: Proceedings.
Lv X, Su M, Wang Z. (2021) Application of face recognition method under deep learning algorithm in embedded systems. Microprocessors and Microsystems. https://doi.org/10.1016/j.micpro.2021.104034..
Kim HI, Choi JY, Lee SH, Ro YM. (2016) Feature scalability for a low complexity face recognition with unconstrained spatial resolution. Multimedia Tools and Applications. 75:6887-–6908.
Sahan JM, Abbas EI, Abood ZM. (2021) A facial recognition using a combination of a novel one-dimension deep CNN and LDA. Materials Today: Proceedings https://doi.org/10.1016/j.matpr.2021.07.325..
Almabdy S, Elrefaei L (2019) Deep convolutional neural network-based approaches for face recognition. Applied Sciences 9:4397.

No competing interests reported.

TheHighlights.docx

Download PDF

Reviewers invited by journal
01 Feb, 2023
Editor assigned by journal
28 Jan, 2023
Submission checks completed at journal
15 Jan, 2023
First submitted to journal
12 Jan, 2023

You are reading this latest preprint version

Scalable Facial Authentication System for a Distributed Environment Using Feature Dimensionality Reduction

Status:

Version 1

Abstract

Figures

1. Introduction

2. User Face Detection And Feature Extraction

2.1. Gabor Filter

2.2. Local binary patterns

2.3. Deep convolutional neural network

3. Feature Dimension Reduction In Distributed Environment

3.1. Linear Discriminant Analysis

3.2. Principal Component Analysis

3.3. Singular Value Decomposition

3.4. Facial Recognition System

4. Scalable Face Authentication

4.1. Distributed environment

4.2. Scalable face feature

4.3. Application scenarios

5. Experimental Results

6. Conclusion

References

Additional Declarations

Supplementary Files

Status:

Version 1