Face recognition algorithm based on stack denoising and self-encoding LBP

Yanjing Lu; Mudassir Khan; Mohd Dilshad Ansari

doi:10.1515/jisys-2022-0011

Article Open Access

Face recognition algorithm based on stack denoising and self-encoding LBP

Yanjing Lu , Mudassir Khan and Mohd Dilshad Ansari

Published/Copyright: April 21, 2022

Published by

Become an author with De Gruyter Brill

Submit Manuscript Author Information

From the journal Journal of Intelligent Systems Volume 31 Issue 1

Abstract

To optimize the weak robustness of traditional face recognition algorithms, the classification accuracy rate is not high, the operation speed is slower, so a face recognition algorithm based on local binary pattern (LBP) and stacked autoencoder (AE) is proposed. The advantage of LBP texture structure feature of the face image as the initial feature of sparse autoencoder (SAE) learning, use the unified mode LBP operator to extract the histogram of the blocked face image, connect to form the LBP features of the entire image. It is used as input of the stacked AE, feature extraction is done, realize the recognition and classification of face images. Experimental results show that the recognition rate of the algorithm LBP-SAE on the Yale database has achieved 99.05%, and it further shows that the algorithm has a higher recognition rate than the classic face recognition algorithm; it has strong robustness to light changes. Experimental results on the Olivetti Research Laboratory library shows that the developed method is more robust to light changes and has better recognition effects compared to traditional face recognition algorithms and standard stack AEs.

Keywords: face recognition; stacked autoencoder; LBP operator; denoising

1 Introduction

Face recognition, as a noninvasive bioinformatics identification method, is widely used in human–computer interaction, information security, and other fields. Feature learning has received wide attention from scholars as a key issue in the field of face recognition. Further, linear feature extraction algorithm mainly includes principal component analysis (PCA) as well as linear discriminant analysis (LDA). Due to the rapid development of modern technology and computer technology, all aspects of social life are transforming toward intelligence as well as automation [1]. High-speed and higher precision automatic identification technology has become the demand of the current era, and more research scholars and scientific research institutions have joined the research of automatic identification technology. Recognition technology based on biometrics, such as face recognition, fingerprint recognition, palm-print recognition, iris recognition, and voice recognition, relatively traditional keys, and keys and passwords have incomparable advantages [2]. One: biometrics uses the user’s own characteristics for identification, no need to bring any tools. Moreover, it does not require to be kept secretly, the user experience is very good; second: biometrics can prevent unauthorized persons from tampering or stealing, cannot be forged, provides good security; third: the biological characteristics of each user are unique, not easy to lose; fourth: biological characteristics are universal, and every characteristic has strong applicability [3]. Based on these advantages, the identification technology based on biometrics has been vigorously developed, and corresponding products have gradually gained popularity in people’s lives; for example, fingerprint attendance machines and face recognition machines in offices use the user’s unique fingerprints and facial features to perform identity recognition.

The focus of research on face recognition lies in the study of core algorithms; through the improvement of the algorithm, the recognition result has a practical recognition rate and recognition speed [4]. The face recognition field, there are the following algorithms: (1) expressing the Euclidean distance through the integral method, algorithm based on geometric features, for example, the Brunelli and Poggio research teams at Massachusetts Institute of Technology (MIT). (2) Using template matching algorithm for facial feature extraction such as the feature extraction of eyebrows and eye contours, that is, the Smith-Kettlewell Eye Research center of Harvard University. However, currently the main problem is that the constant changes in the external environment have a huge impact on face recognition, increasing the difficulty of identification [5]. Research shows that the deep learning structure composed of multiple nonlinear mapping layers uses artificial neural networks with multiple hidden layers to extract data features. Furthermore, compared with the shallow mapping structure, it has a better function expression ability. Hwang and Abebe [6] proposed a robust method for recognizing red from low-resolution faces, green, and the blue depth (Red Green Blue + Depth Map) camera acquires a wide range of postures on the head, illumination, images of changes in facial expressions, and occlusion in some cases. The local binary pattern (LBP) of an Red Green Blue-Depth (RGB-d) RGB-d image with an appropriate feature size of the depth image is used to extract facial features. On the basis of the error correction output encoding, they are fed into multiclass support vector machines (MSVMs), for offline training and verification, then sort online. The developed method is called the appropriate feature size of LBP-RGB-d-MSVM and depth images [6]. Hwang and Abebe proposed the LBP operator for face recognition, where LBP has both rotational and gray scale invariance and is computationally simple and has good classification performance [6]. Yan et al. proposed the classic feature face (Eigen-faces) method to first apply PCA to face recognition. However, the feature face method ignores the local features of the human face and is sensitive to light changes, so the recognition effect is not particularly ideal [7]. Zhang et al. proposed fisher-face method, which combines PCA and LDA that reduces the computational complexity and avoids the dispersion matrix singularity but may lose some important information during dimensionality reduction to affect the discriminative effect [8].

Recent studies indicate that the application of the fuzzy set theory introduced by Zadeh [21] on the extraction of texture spectrum features as well as their efficient successors, the LBP features, can possibly improve their robustness to noise. Moreover, this is also applied in various application areas by numerous researchers. Furthermore, Ansari et al. used intuitionistic fuzzy sets-local binary pattern for the extraction of features from digital image [11,13,15,18]. Additional, Talab et al. applied LBP for face recognition [17]. Similarly, other authors have used LBP as well other variants of LBP along with fuzzy set theory and their extensions such as Fuzzy LBP and Intuitionistic Fuzzy Set LBP [11–20]. Fuzzy LBP for texture representations is more robust to the existence of noise. Moreover, it is also applied in various application areas through fuzzification of a variety of BP approaches, including the LBP, LBP/C, local edge patterns, and median binary pattern.

In summary, a face recognition algorithm based on LBP and stacked AE is developed. The LBP operator is an effective texture descriptor that can characterize the local micropatterns and their distribution of face images, such as their highlights, dark points, and edges. Moreover, it has simple computation and fast operation speed. LBP operators have been widely used in dynamic texture recognition, expression analysis, facial recognition, and so aspects. Considering the local texture features of the face image extracted by LBP as the input of the first layer of autoencoder (AE), sparse autoencoder (SAE) automatically learns features layer by layer. Finally, the classification and recognition are automatically performed by Soft-max. Experimental results were displayed and compared with the classic algorithm; the extracted features are more discriminative, strong robustness to light changes; it can recognize and classify face images more accurately.

The article is structured as follows: existing method and proposed model are given in Section 2, and Section 3 displays the analysis of results. Section 4 discusses the conclusion and future work.

2 Research methods

2.1 Local binary mode and stacked AE

2.1.1 Local binary mode

The basic LBP operator works on a 3 × 3 window, let the center pixel of the window be P c , the other eight pixels are P ₁, P ₂, …, P ₈, compare P ₁, P ₂, …, P ₈ with the center pixel point P c of the window, if P 1 > P c , then mark the pixel position of P ₁ as 1, otherwise, it is 0. Additionally, connect 8 binary numbers in clockwise order, and calculate the corresponding decimal value. To adapt to the texture characteristics of different scales, Ojala et al. [22] developed an extended LBP operator, replaced square neighborhoods with circular neighborhoods, and picked P equidistant points in the circular neighborhood, the radius of the circular neighborhood is R, which is represented by the scale (P, R). For any LBP operator, its encoding is shown in formula (1):

(1) LBP P , R = ∑ i = 0 P − 1 s ( g i − g c ) 2 i , s ( x ) = 1 , x ≥ 0 , 0 , x < 0 ,

where g c represents the gray value of the central pixel, and g _i represents the gray value of the i-th point in the neighborhood. This operator will get 2 P different binary patterns. To solve the problem of too many patterns, a unified model LBP is proposed. Sum the absolute value of the difference between the basic LBP binary pattern and the binary pattern shifted by one bit, and the definition is shown in formula (2):

(2) U ( G P ) = ∣ S ( g P − 1 − g c ) − S ( g 0 − g c ) ∣ + ∑ P = 1 P − 1 ∣ S ( g P − g c ) − S ( g P − 1 − g c ) ∣ .

If U ( G P ) is less than or equal to 2, then the mode is a unified mode. Using the unified mode, the LBP operator can reduce 2^P binary modes to P(P − 1) + 2, reduce the feature dimension, and improve the efficiency of the algorithm. Other unifications that are not unified are called mixed modes. The image encoded by the LBP operator, its overall texture and contour characteristics are described, but its highlights, the distribution of micro-patterns such as dark spots has not been fully characterized. Additionally, for this problem, proposed a block mode LBP calculation. Moreover, implementation process is as follows: first divide the original face image f(x, y) into m × n sub-images, extract the LBP of each sub-image separately, then statistically generate the LBP histogram of each sub-image, finally, connect these histograms in order, the local texture features that make up the entire face image. The LBP features of the block face image extracted by the unified mode, strong robustness to light changes, and the calculation is simple and easy to implement.

2.2 Stacked AE

AE is composed of input layer; the hidden layer and the output layer constitute a network. The automatic encoder mainly realizes the identity process of h θ ( x ( i ) ) ≈ x ( i ) , making the output y as close to the input x as possible. Assuming the sample set { ( x ( 1 ) , y ( 1 ) ) , ( x ( 2 ) , y ( 2 ) ) , ⋯ , ( x ( m ) , y ( m ) ) } , contains m samples. For a single example (x, y), the cost function is shown in formula (3):

(3) J ( W , b ; x , y ) = 1 2 ∥ h w , b ( x ) − y ∥ 2 .

This is a variance cost function. For a data set containing m samples, the overall cost function is defined as shown in formula (4):

(4) J ( W , b ) = 1 m ∑ i = 1 m J ( W , b ; x ( i ) , y ( i ) ) + λ 2 ∑ l = 1 n l − 1 ∑ i = 1 s l ∑ j = 1 s l + 1 ( W j i l ) 2 = 1 m ∑ i = 1 m 1 2 ∥ h w , b ( x ( i ) ) − y ( i ) ∥ 2 + λ 2 ∑ l = 1 n l − 1 ∑ i = 1 s l ∑ j = 1 s l + 1 ( W j i l ) 2 ,

where W j i l represents the connection weight between the j-th unit of the l-th layer and the i-th unit of the l + 1th layer, λ is the weight attenuation parameter, n _l represents the number of layers of the AE, and s _l represents the number of units in the first layer. If the AE has a three-layer structure, then h w , b ( x ) = f ( z ( 3 ) ) = f ( W ( 2 ) a ( 2 ) + b ( 2 ) ) , where a ⁽²⁾ denotes the output value of the second layer, b ⁽²⁾ represents the bias term of the second layer, and z ⁽³⁾ represents the input weighted sum of the third layer (including the bias unit). The sigmoid function is usually selected as the activation function, then f ( z ) = 1 1 + e − z .

Gradient descent method can be used to calculate the minimum value of the overall cost function J(W, b) for the parameters W and b, in each iteration of the gradient descent method, the parameters W and b are updated as shown in formulas (5) and (6):

(5) W j i ( l ) = W j i ( l ) − α ∂ ∂ W j i ( l ) J ( W , b ) ,

(6) b i ( l ) = b i ( l ) − α ∂ ∂ b i ( l ) J ( W , b ) ,

where α is the learning rate. The reverse conduction algorithm is used to calculate the partial derivative. The residual for the n _l th layer (output layer) is shown in formula (7):

(7) δ i ( n l ) = ∂ ∂ z i ( n l ) J ( W , b ; x , y ) = ∂ ∂ z i ( n l ) ∥ y − h W , b ( x ) ∥ 2 = − ( y i − a i ( n l ) ) f ′ ( z i ( n l ) ) .

The residuals for the second to nl − 1 layers (hidden layers) are shown in formula (8):

(8) δ i ( l ) = ∑ j = 1 s l + 1 W j i ( l ) δ j ( l + 1 ) f ′ ( z i ( n l ) ) .

Finally, the partial derivative is calculated as shown in formulas (9) and (10):

(9) ∂ ∂ W j i ( l ) J ( W , b ; x , y ) = a j ( l ) δ i ( l + 1 ) ,

(10) ∂ ∂ b i ( l ) J ( W , b ; x , y ) = δ i ( l + 1 ) .

After calculating the partial derivative, the gradient descent technique can be used to update the weights and get the optimal parameters of the network. Multiple AEs are connected together, and the features learned by the AE of the previous layer are used as the input of the latter layer; in this way, a stack type AE is formed. Moreover, with layer-by-layer unsupervised pretraining, the network has a better initial value. On the basis of this initial value, the gradient descent method is used for supervised fine-tuning; the network parameters can be converted to a better local extreme point. This is because the unsupervised training process provides a priori information about the patterns contained in a large amount of input data. In the supervised fine-tuning stage, the training goal is to minimize the prediction error; therefore, a logistic regression layer can be added to the top layer of the network to form a classification. A logical classifier is added at the top level, the obtained high-level features is taken as the input of the classifier, forming supervised learning. The parameters of all layers of the network can be fine-tuned using the gradient descent method to achieve the best recognition effect.

2.3 Face recognition algorithm based on LBP and stacked autoencoder

Combine LBP algorithm with SAE method for face recognition. Unlike traditional AEs that directly use images as input data, the features obtained by the LBP algorithm are used as the input of SAE. The local binary mode is a texture feature extraction algorithm; the extracted features have gray-level translation invariance, rotation invariance, and other advantages. When the face image is affected by uniform illumination changes and tiny rotations, the local texture structure characteristics will not change much. Therefore, the LBP algorithm can better characterize the facial features. However, the dimension of the facial features extracted by the LBP algorithm is too high, difficult to calculate, and the SAE algorithm can not only learn features automatically but can effectively reduce dimensionality [9]. Further, combining these two algorithms can get more structured and hierarchical abstract features. At the same time, the LBP texture feature is used as the input of the SAE network; it can strengthen the learning ability of the deep network. Algorithm steps are as follows: (1) read in the face picture, and divide the face image into blocks, use 4 × 4 blocks. (2) Perform a unified mode LBP operation on each subimage, get the histogram of each subimage, and connect the histograms of all subimages, get the texture characteristics of the entire image. (3) Use the texture feature obtained in step (2) as the input of the SAE network to train the network, and this process is unsupervised learning. (4) Use BP neural network and gradient descent algorithm to fine-tune the network because this process takes labeled data as the expected output, and so the process is supervised learning. (5) Use the LBP texture feature of the test sample as the input of the SAE network, the abstract features of the face are automatically learned and extracted by the network. (6) Perform Soft-Max regression classification at the end, obtain the standard value of the test sample, and calculate the recognition error rate.

3 Result analysis

3.1 LBP operator selection experiment

To better extract the LBP texture structure features of the face image, first, the appropriate LBP operator is selected. This experiment selects the unified mode LBP operator of the block to be carried out on the Olivetti Research Laboratory (ORL) database. The original image is blocked, the LBP features of each subimage are extracted separately, and then a histogram of the entire image is formed. Figure 1 shows that, for the same face image, the feature dimension of the face image extracted by the block LBP operator is much larger than the face image feature extracted by the standard LBP operator.

Figure 1

Histogram of the original image.

The image feature dimension represents the amount of image information extracted; as the key information of the extracted face image increases, the ability to express the details of facial images is enhanced. Face can be highlighted, detailed features such as dark spots and edges are more fully characterized, the recognition rate of face images will increase, but the amount of system calculations has also increased. Additionally, excessive calculations during training will cause the problem of “overtraining,” and this will also affect the recognition rate of the image; therefore, choosing a suitable LBP operator is the key to this experiment. The different mode of LBP Operator combined with SAE to conduct experimental analysis on the ORL face database, the result obtained is shown in Figure 2.

Figure 2

The recognition rate of the unified mode LBP operator of different blocks.

As can be seen from Figure 2, when using 4 × 4 block unified mode LBP operator combined with SAE for experiment, we can get a higher recognition rate, and as the number of iterations increases, the recognition rate is significantly improved. When using 5 × 5 blocks and 6 × 6 blocks for experiments, the recognition rate has decreased. Therefore, the 4 × 4 block unified LBP operator combined with SAE is used for the face recognition. The LBP operator has a good recognition rate, significantly higher than other methods, and with the increasing number of sample categories selected during the identification process, the recognition rate of the LBP algorithm and the weighted LBP algorithm is relatively much lower for other algorithms. The LBP weighted based on prior knowledge achieves better recognition rates. Through the comparison of this experiment, the recognition rate of the weighted LBP is better than the unweighted LBP, and it also shows that the contribution degree of the different subwindows of different parts in the recognition process of the face images is different. We conclude that the eyes, mouth, and nose contribute greatly, and the contribution of other regions is not so obvious. First, the weight improvement of prior knowledge can improve the recognition rate of the LBP algorithm.

3.2 Extended yale face database B face library experiment

Proposed algorithm is compared with the classic algorithm; the experimental results of each algorithm are the good results in experiments with multiple sets of parameters. Among them, the LBP algorithm uses a 4 × 4 block unified mode LBP operator, compare the size of the current pixel and surrounding pixels, set the one greater than the current pixel to 1, set to 0 if it is less than. Set the number of subblock divisions in the WLD algorithm M = 10, N = 10, differential excitation quantization level L1 = 4, the gradient direction quantization level L2 = 8. Sparse representation classification algorithm (SRC) sets the upper limit of the number of iterations t _max = 20, tolerable error ε = 0.001. The number of iterations in the stacked AE is 1,000, learning rate α = 1, set the SAE network to a three-tier structure of 944-500-100-38. Table 1 shows the correct recognition rate of different algorithms on the ExtendedYaleB library.

Table 1

Correct recognition rate of different algorithms on ExtendedYaleB library

Algorithm	Correct recognition rate (%)
PCA	59.00
SVM	72.67
LBP	63.00
WLD	67.01
SRC	90.13
SAE	83.52
LBP-SAE	95.15

It can be observed from the results in Table 1 that for facial image recognition with obvious changes in illumination, the recognition effect of SAE and LBP-SAE based on deep neural network framework is better than that of shallow structure PCA and support vector machine (SVM). The accuracy of the Proposed model LBP-SAE on the ExtendedYaleB library reached 95.15%, than PCA, SVMs were 36.15% higher, 22.48%. Compared with algorithms, such as PCA and SVM, the features extracted by the algorithm are more discriminative [8]. The recognition effect of WLD algorithm is better than that of LBP algorithm; however, the computational complexity of the WLD algorithm is relatively large, the time required to process a photo is 0.0027 s, the time required for the LBP algorithm is 0.0015 s, and it can be seen that the LBP algorithm is more efficient. The method based on sparse representation is essentially still based on the gray features of the sample, and sparse linear expression is used to realize the separation of change components within the class, but for practical problems, intraclass changes are relatively complicated, and it is difficult to separate it simply by linear expression [10]. The recognition rate of SAE and LBP-SAE based on deep neural network framework is higher than that of sparse representation algorithm. The LBP algorithm has strong robustness to illumination and posture changes and enhances the detailed features of face images, and it can more effectively characterize the feature information of the face image under the condition of light changes and make more accurate predictions. The LBP texture structure feature is used as the input of the stacked AE, and it can reduce SAE’s associative memory of redundant information, so as to realize the accurate recognition of the face image under the condition of changing illumination.

3.3 ORL face database experiment

There are 40 people in the ORL face database, 10 images per person, a total of 400 face images; the image resolution is 112 × 92. Randomly five out of ten images of each person were selected as training samples, and the other five were used as test samples. The SAE network is set to a three-tier structure of 944-500-100-40. The last layer is connected to the Soft-max classifier for recognition and classification to get the recognition result. The algorithm is compared with the classic algorithm; the experimental results of each algorithm are the good results in experiments with multiple sets of parameters. Table 2 shows the correct recognition rate of different algorithms on the ORL library.

Table 2

Correct recognition rate of different algorithms on ORL library

Algorithm	Correct recognition rate (%)
PCA	87.89
SVM	92.13
LBP	91.01
WLD	91.12
SRC	96.03
SAE	94.10
LBP-SAE	99.20

As can be seen from Table 2, the recognition rate of the algorithm LBP-SAE in this study with the introduction of the LBP operator on the ORL library reached 99.20%; compared with the traditional SAE algorithm, it is improved by 5.1%, and it is because the LBP algorithm has strong robustness to illumination and posture changes and enhances the detailed features of face images. The experiment shows that the algorithm has a better recognition effect on face images.

3.4 Yale face database experiment

Yale face library contains 165 pictures of 15 volunteers, contains lighting, and changes in expression and posture. The image resolution is 100 × 100, and everyone has six different expressions and three different lighting. Randomly five out of ten images of each person were selected as training samples, the other six were used as test samples. The SAE network is set to a three-tier structure of 944-500-100-15. The last layer is connected to the Soft-max classifier for recognition and classification to get the recognition result. Table 3 shows the correct recognition rate of different algorithms on the Yale face database.

Table 3

Correct recognition rate of different algorithms on Yale library

Algorithm	Correct recognition rate (%)
PCA	80.89
SVM	92.14
LBP	90.13
WLD	91.33
SRC	94.76
SAE	93.21
LBP-SAE	99.05

As can be seen from Table 3, the recognition rate of the algorithm LBP-SAE on the Yale database reached 99.05%; it further shows that the algorithm has a better recognition ability in face recognition.

The LBP method uses only the size relationship of the neighborhood and the central pixels, does not exploit the size relationship between the upper neighborhoods, and has equal weights for the neighbors. Therefore, how to use the size relationship between the neighbors for encoding is a key technical problem to solve. Considering that the local ordinal pattern algorithm can order the neighbors according to certain relationships, this algorithm can improve the coding of the mode.

4 Conclusion

Stacked AE has a better function expression ability than shallow mapping structure. The multi-layer structure can be useful to learn the characteristics of the images, has more powerful image feature extraction and classification and recognition functions. Experiments on the ExtendedYaleB face database show that when compared with the classic face recognition algorithm, the proposed method has a higher recognition rate. It has strong robustness to illumination changes and can enhance the detailed features of the face sample image, and it can more effectively characterize the feature information of the face sample image under nonideal conditions, and make more accurate Peng’s prediction. Combining the adaptive LBP algorithm with the stacked AE can enhance the image details that are ignored by the network structure, making the algorithm more robust to changes in lighting, and improving the recognition rate of the algorithm. In this article, the study of the stack AE is relatively preliminary as the selection of the number of network layers, the choice of the number of hidden layer nodes, and the choice of the top-level classifier are all based on the prior experience, and the subsequent studies can do further research and analysis in these aspects.

Acknowledgments

Zhengzhou excellent teaching team. Zhengzhou Education Bureau (zjmd [2019] No. 598).

Conflict of interest: Authors state no conflict of interest.

References

[1] Kas M, El-Merabet Y, Ruichek Y, Messoussi R. A comprehensive comparative study of handcrafted methods for face recognition lbp-like and non lbp operators. Multimed Tools Appl. 2020;79(1/2):375–413.10.1007/s11042-019-08049-3Search in Google Scholar

[2] Li T, Wang L, Chen Y, Ren Y, Xia J. A face recognition algorithm based on lbp-ehmm. J Artif Intell. 2019;1(2):61–8.10.32604/jai.2019.06346Search in Google Scholar

[3] Benitez-Garcia G, Nakano-Miyatake M, Olivares-Mercado J, Perez-Meana H, Sanchez-Perez G, Toscano-Medina K. A low complexity face recognition scheme based on down sampled local binary patterns. Int Arab J Inf Technol. 2019;16(3):338–47.Search in Google Scholar

[4] Reddy N, Rao M, Satyanarayana C. A novel face recognition system by the combination of multiple feature descriptors. Int Arab J Inf Technol. 2019;16(4):669–76.Search in Google Scholar

[5] Zhou L, Wang H, Lin S, Hao S, Lu ZM. Face recognition based on local binary pattern and improved pairwise-constrained multiple metric learning. Multimed Tools Appl. 2020;79(1):675–91.10.1007/s11042-019-08157-0Search in Google Scholar

[6] Hwang CL, Abebe HB. Rgb-d face recognition using lbp with suitable feature dimension of depth image. IET Cyber-Phys Syst: Theory Appl. 2019;4(3):189–97.10.1049/iet-cps.2018.5045Search in Google Scholar

[7] Yang W, Zhang X, Li J. A local multiple patterns feature descriptor for face recognition. Neurocomputing. 2020;373(Jan.15):109–22.10.1016/j.neucom.2019.09.102Search in Google Scholar

[8] Zhang Q, Li H, Li M, Ding L. Feature extraction of face image based on lbp and 2-d gabor wavelet transform. Math Biosci Eng. 2020;17(2):1578–92.10.3934/mbe.2020082Search in Google Scholar PubMed

[9] Dinesh K, Rao PV. Implementing and analysing far and frr for face and voice recognition (multimodal) using knn classifier. Int J Intell Unmanned Syst. 2020;8(1):55–67.10.1108/IJIUS-02-2019-0015Search in Google Scholar

[10] Chakraborty S, Singh SK, Chakraborty P. R-theta local neighborhood pattern for unconstrained facial image recognition and retrieval. Multimed Tools Appl. 2019;78(11):14799–822.10.1007/s11042-018-6846-zSearch in Google Scholar

[11] Ansari MD, Koppula VK, Ghrera SP. Fuzzy and Entropy based approach for Feature Extraction from Digital Image. Pertanika J Sci Technol. 2019;27(2):829–46.Search in Google Scholar

[12] Srivastava R, Tomar R, Sharma A, Dhiman G, Chilamkurti N, Kim B-G. Real-time multimodal biometric authentication of human using face feature analysis. Comput, Mater Continua. 2021;49(1):1–19.10.32604/cmc.2021.015466Search in Google Scholar

[13] Ansari MD, Ghrera SP, Mishra AR. Texture feature extraction using intuitionistic fuzzy local binary pattern. J Intell Syst. 2020;29(1):19–34.10.1515/jisys-2016-0155Search in Google Scholar

[14] Meethongjan K, Dzulkifli M, Rehman A, Altameem A, Saba T. An intelligent fused approach for face recognition. J Intell Syst. 2013;22(2):197–212.10.1515/jisys-2013-0010Search in Google Scholar

[15] Ansari MD, Ghrera SP. Feature extraction method for digital images based on intuitionistic fuzzy local binary pattern. 2016 International Conference System Modeling & Advancement in Research Trends (SMART); IEEE. 2016, pp. 345–349.10.1109/SYSMART.2016.7894547Search in Google Scholar

[16] Ghorbel A, Aydi W, Tajouri I, Masmoudi N. Hybrid approach for face recognition from a single sample per person by combining VLC and GOM. J Intell Syst. 2020;29(1):1523–34.10.1515/jisys-2018-0380Search in Google Scholar

[17] Talab MA, Awang S, Ansari MD. A novel statistical feature analysis-based global and local method for face recognition. Int J Opt. 2020;2020:17.10.1155/2020/4967034Search in Google Scholar

[18] Ansari MD, Ghrera SP. Intuitionistic fuzzy local binary pattern for features extraction. Int J Inf Commun Technol. 2018;13(1):83–98.10.1504/IJICT.2018.090435Search in Google Scholar

[19] Dhiman G, Oliva D, Kaur A, Singh KK, Vimal S, Sharma A, et al. BEPO: a novel binary emperor penguin optimizer for automatic feature selection. Knowl Syst. 2021;211:106560.10.1016/j.knosys.2020.106560Search in Google Scholar

[20] Malhotra P, Kumar D. An optimized face recognition system using cuckoo search. J Intell Syst. 2019;28(2):321–32.10.1515/jisys-2017-0127Search in Google Scholar

[21] Zadeh LA. Fuzzy Sets. J Inf Control. 1965;8(3):338–53.10.21236/AD0608981Search in Google Scholar

[22] Ojala T, Pietikäinen M, Harwood D. A comparative study of texture measures with classification based on featured distributions. Pattern Recognit. 1996;29(1):51–9.10.1016/0031-3203(95)00067-4Search in Google Scholar

Received: 2021-10-09

Revised: 2021-11-30

Accepted: 2021-12-07

Published Online: 2022-04-21

This work is licensed under the Creative Commons Attribution 4.0 International License.

Articles in the same Issue

https://doi.org/10.1515/jisys-2022-0011

Keywords for this article

face recognition; stacked autoencoder; LBP operator; denoising

Creative Commons

BY 4.0