Face Recognition System: Literature Review And Techniques

Applications of Face Recognition System

The face recognition system is the technology that has the capability of verifying or identifying any individual from his digital image or video frame from the video sources. Various significant methodologies are present for face recognition system and they work by proper comparison of the selected facial characteristics from the provided image with faces in the database [4]. This facial recognition is even descried as the application of biometric artificial intelligence, which comprises of the core capability of identifying an individual uniquely after proper analysis of the patterns on the basis of the facial textures or the shapes of that specific individual. It is considered as the basic form of computer application initially and hence is being utilized widely in every other form of technology like robotics [11].

The face recognition system is utilized as the access control within the security systems and could be compared to any other type of biometric system like fingerprint recognition system and iris recognition system [8]. Since, the total accuracy of this particular system as the biometric technology is lower than the iris recognition system or fingerprint recognition system; this is being often widely adopted for the non-invasive as well as contactless procedure. In recent studies, it has been observed that the facial recognition system has become the most popular as well as commercial verification or marketing tool.

According to Han et al. [20], some of the most significant applications of this face recognition system are video surveillance, advanced interaction between human and computer, video databases, and automated indexing of the images and several others. All of these above mentioned applications are extremely important and noteworthy for the users of face recognition system. For the proper verification and identification of payment, the facial biometrics could be integrated with all the physical devices or objects [10]. With the subsequent utilization of passcodes, the mobile phones and other customer electronics could be easily accessed with the facial features of the owners for making payments. Moreover, criminal identification is also easier with this particular system. Other applications of face recognition system mainly include advertising, health care and many others.

                                   

                                              Figure 1: Block Diagram of Face Recognition System

Techniques for Face Recognition System

                                                                       (Source: Haig [5])

Parkhi, Andrea and Andrew [2], state that the entire procedure of the face recognition method could be performed in two distinct steps. The first and the foremost step of this system includes the feature extraction as well as feature selection. The second step in this system includes the proper classification of the objects. The recent developments have been introduced by varying the technologies to this process of face recognition. Some of the major and the most notable techniques of the face recognition system are as follows:

1. i) Traditional Technique: Few of the algorithms of this face recognition system help in the identification of the facial features of the individual by simply extracting the features or landmarks from the image of that specific individual’s face [9]. As an example, the algorithm could analyse the relative positions, shapes and sizes of the nose, eyes, jaws and cheek bones. Each and every above mentioned feature is then utilized for the purpose of searching the other images with the help of the matching features. The other important algorithms eventually normalize the entire gallery of the face images as well as compressing the facial data and hence saving that data within the image, which could be utilized for the purpose of face recognition [13]. A specific probe image is next subsequently compared with the collected facial data. The earliest and the most successful system is thus based on the several techniques of template matching and is being applied to the collection of the important facial features. A sort of the respective compacted face representation is provided in this manner.

As per Cui et al. [12], the recognition algorithms hence could be sub divided into two major approaches, known as the geometric or checking the distinguishing characteristics and the photometric; a specified statistical approach, which is responsible for distilling any image into reasonable values and then comparing these values with the various templates for the core purpose of eliminating the variances. Most of these algorithms could be properly classified into two broader categories like holistic model and feature based model [15]. The holistic model attempts in recognizing the face in the entirely and the feature based model can sub divide these elements like as per the features and analysis of each of the feature in comparison to any other feature. The most popular algorithms of the face recognition system are linear discriminant analyses, hidden Markov model, principal component analyses with eigen faces, dynamic link matching, multi linear subspace learning with tensor representation, Fisher face algorithm and various others [19].

                               

                                             Figure 2: Process of Video Based Face Recognition System

                                                                 (Source: Chan et al. [18])

1. ii) 3 Dimensional Recognition Technique: Drira et al. [9] state that this particular technique of face recognition is majorly responsible for utilizing the 3D sensors to eventually capture the information regarding the distinct shape of the face of an individual. This specific information is afterwards utilized for the purpose of utilizing to identify and verify the distinct characteristics on the face surface like contouring of eye sockets, chin and nose. The major benefit of the 3D face recognition system would be that this is not at all affected by the various changes or alterations within lighting similar to other techniques. This technique could even identify any face from the range of overviewing angles involving the profile view. The respective three dimensional data points from the face substantially improvises the overall precision of the face recognition. As per a recent research of Amos, Bartosz and Mahadev [14], the three dimensional system is being enhanced by properly developing the sophisticated sensors and then capturing the three dimensional image of the face. These typical sensors could work by simply projecting the structured light on to the face of the individual. More than a dozen of the image sensors could be eventually placed on one CMOS chip and each sensor then captures the different part of the spectrum [10].

A perfect three dimension matching technique can also be sensitive for expressions. For this purpose, the respective tools could be applied from the metric geometry for treating each and every expression as isometric. The new methodology for the three dimensional face recognition system is introducing the method of capturing the 3D image by utilizing the three cameras for tracking and these cameras would be pointing at separate angles [16]. These cameras can work altogether for tracking the face of the person and detect him or her properly.

iii) Analysis of Skin Textures: The next significant trend that utilizes the several visual details of skin as being captured within the standard scanned or digitalized images. This particular technique, known as the analysis of skin texture turns the respective unique patterns, spots and lines within the skin of the individual to a mathematical space. According to Lu, Yap-Peng and Gang [6], the other form of this face recognition system technique is the analysis of surface texture. The analysis of surface texture works in the similar manner of face recognition method. An image of the skin patches is being undertaken and these are termed as skin prints. This skin patch is then sub divided into small specific blocks. With the help of face recognition algorithms for turning the patch to the measurable or mathematical space, this particular system eventually distinguishes the lines, skin texture and pores [2]. It could identify the major differences between the identical twins that is not yet possible with the face recognition method.

                                               

                        Figure 3: Surface Texture Analysis Model of Face Recognition Method

                                                             (Source: Sun et al. [3])

1. iv) FaceNet Technique: According to Schroff, Dmitry and James [1], a recent advancement has come into account for the field of face recognition. This particular advancement efficiently resolves the issues of current approaches. This new technique of FaceNet technique directly learns the mapping from the facial images to the compact Euclidean space, in which the distances are directly corresponded to the specific measure of similarity of the face [1]. When this space is being produced, the various tasks like the face recognition, clustering or verification could be easily deployed with the standardized techniques with the embedding of FaceNet as the characteristic vectors.

This specific methodology of FaceNet eventually utilizes the deeper convolutional network training to the direct optimization of the embedding itself and not on the layer of intermediate bottleneck in the previously existing approaches of deep learning [1]. For this purpose of training, triplets of the roughly aligned matching as well as non matching face patches are being generated with the help of a novel online triplet mining methodology. FaceNet technique is extremely effective and efficient and the performance of state of the art face recognition is being achieved by utilizing 128 bytes of one face. The accuracy of this technique is around 99.63%. The triplet loss is calculated by the following process.

The embedding is being represented by f(x) ∈ R^d. FaceNet embeds the image of x to the d dimensional Euclidean space. Moreover, the embedding is constrained for living on the respective d dimensional hyper sphere, that is, ||f(x)||₂ = 1. The loss is then motivated within the context of the nearest neighbour classifications [1]. Hence, the image of x^a is ensured for the specified individual, which is much closer to the other images of positive (x^p) and negative (xⁿ) of that person. This helps in determining the triplet loss and thus FaceNet technique is utilized easily, without much issues. Due to the integration of face clustering technique, the false acceptance rate is much lower in this technique and thus it is extremely popular. The future work for this technique would check the error cases by reducing the size of the model and requirements of CPU [1].

1. v) Fast l₁Minimization Algorithm: Yang et al. [17] in their paper has stated that this algorithm helps in finding the minimum l₁norm solutions for the underdetermined linear system b = Ax. The numerical deployment of the sparsity based classified framework within the robust face recognition method is checked here. This sparse representation could be sought for recovering the human identity from the higher dimensional face images, which might be corrupted by facial disguises, pose variations and illuminations [17]. This fast l₁ minimization algorithm uses the framework of classical convex optimization, called the ALM or Augmented Lagrangian Method. The face alignment is easily checked with this technique. The primal augmented algorithm is considered as one of the most efficient and the fastest methodology for solving the issues of face alignment.
2. vi) Heterogeneous Face Recognition Technique: As per Klare and Anil [7], the heterogeneous face recognition technique or HFR eventually involves the subsequent matching of two images of face from the alternative imaging modalities like the infrared images of the photographs or even the distinct sketch of that photograph of an individual. The most accurate system of the heterogeneous face recognition technique comprises of the greater value within the several applications like surveillance and forensics. The databases are updated specifically; however the respective probe images are also restricted to the some of the alternative modalities [7]. The generic framework of heterogeneous face recognition technique is being proposed in this paper, where the probe images and database images are also represented in respect to the non linear kernel similarities with the specific collection of the prototype face images. This prototype then subjects the images within the modalities and measures the similarities of images within the prototype images. The significant accuracy of this specific technique of HFR is improvised with the help of linear transformations from the representation of probe prototype by following with the framework of random subspace [7].

With the help of geometric normalization, image filtering and localized description representations, the respective image processing is easily completed and the framework of heterogeneous prototypes is considered. For the experiments Klare and Anil [7] have taken five data sets within the database. The first data set is near infrared to visible, where the spectrum of 780 to 1100 nm is checked. The next data set is thermal to visible, where the sensitivity is being checked within the range of 3 to 5 nm an 8 to 12 nm. The third data set is viewed sketch to visible, where a viewed sketch is checked properly. The fourth data set is forensic sketch to visible. This is the most important data set of this particular application of heterogeneous face recognition technique [7]. The final and the fifth data set is the standardized face recognition.

                           

                                  Figure 5: Viewed Sketch and Forensic Sketch Experiment Results

                                                           (Source: Klare and Anil [7])

Hence, the method of P-RS or prototype random subspaces is being found for the heterogeneous face recognition system.

As per Galbally, Sébastien and Julian [10], the face recognition system has a major contribution of the resilient model of face recognition on the basis of the behavioural features with the characteristics of physiological biometric. These physiological features of the face of the individuals with proper relevance to the several expressions like sadness, fear, happiness, ager, disgust and surprise are solely linked with the respective geometrical structures that are restored as the basic matching template for this recognition system [5]. The respective behavioural aspect of the system is related to the attitude behind the various expressions as the property base. These property bases are then aligned and exposed as the hidden categories in the genetic algorithms. The specific set of gene training helps in the evaluation of the expressional uniqueness of several faces of individuals for providing the robust expressional model of recognition for the biometric security [9]. The basic experimental analysis and the procedure of hierarchical security structure is quite efficient within the identification of geometric shapes for the physiological traits.

The future scope of this face recognition method is extremely high and better in comparison to other biometric systems. The utilization of the spherical canonical images would be substantially enabling the users in performing the matching within the domain of spherical harmonic transform [1]. This type of spherical canonical image does not need any preliminary image alignment. The various errors that could be introduced by the embedding of expressional spaces with the previously defined geometry can be eradicated without any issue. Moreover, the measurement of the facial surface cropping of the larger distance positions within each and every point could be evaluated within the help of a parallelized parametric version and this is considered as another important and significant future scope of the methodology of face recognition [5]. Further studies or researches could also be laid down in the proper direction of the allele of the gene matching with the geometric factors of these facial expressions. Better tools could also be utilized in the future of face recognition since, the integration of this technique is done easily and promptly. The next version of this technique that could be effective and efficient for the future generation is the full body recognition technique [2]. All of these above mentioned works could bring major scope for the future technological world.

References:

[1] Schroff, Florian, Dmitry Kalenichenko, and James Philbin. “Facenet: A unified embedding for face recognition and clustering.” In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 815-823. 2015.

[2] Parkhi, Omkar M., Andrea Vedaldi, and Andrew Zisserman. “Deep face recognition.” In BMVC, vol. 1, no. 3, p. 6. 2015.

[3] Sun, Yi, Ding Liang, Xiaogang Wang, and Xiaoou Tang. “Deepid3: Face recognition with very deep neural networks.” arXiv preprint arXiv: 1502.00873 (2015).

[4] Wen, Yandong, Kaipeng Zhang, Zhifeng Li, and Yu Qiao. “A discriminative feature learning approach for deep face recognition.” In European Conference on Computer Vision, pp. 499-515. Springer, Cham, 2016.

[5] Haig, Nigel D. “The effect of feature displacement on face recognition.” Perception 42, no. 11 (2013): 1158-1165.

[6] Lu, Jiwen, Yap-Peng Tan, and Gang Wang. “Discriminative multimanifold analysis for face recognition from a single training sample per person.” IEEE transactions on pattern analysis and machine intelligence 35, no. 1 (2013): 39-51.

[7] Klare, Brendan F., and Anil K. Jain. “Heterogeneous face recognition using kernel prototype similarities.” IEEE transactions on pattern analysis and machine intelligence 35, no. 6 (2013): 1410-1422.

[8] Liao, Shengcai, Anil K. Jain, and Stan Z. Li. “Partial face recognition: Alignment-free approach.” IEEE Transactions on pattern analysis and machine intelligence 35, no. 5 (2013): 1193-1205.

[9] Drira, Hassen, Boulbaba Ben Amor, Anuj Srivastava, Mohamed Daoudi, and Rim Slama. “3D face recognition under expressions, occlusions, and pose variations.” IEEE Transactions on Pattern Analysis and Machine Intelligence 35, no. 9 (2013): 2270-2283.

[10] Galbally, Javier, Sébastien Marcel, and Julian Fierrez. “Image quality assessment for fake biometric detection: Application to iris, fingerprint, and face recognition.” IEEE transactions on image processing 23, no. 2 (2014): 710-724.

[11] Li, Billy YL, Ajmal S. Mian, Wanquan Liu, and Aneesh Krishna. “Using kinect for face recognition under varying poses, expressions, illumination and disguise.” In Applications of Computer Vision (WACV), 2013 IEEE Workshop on, pp. 186-192. IEEE, 2013.

[12] Cui, Zhen, Wen Li, Dong Xu, Shiguang Shan, and Xilin Chen. “Fusing robust face region descriptors via multiple metric learning for face recognition in the wild.” In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 3554-3561. 2013.

[13] Taigman, Yaniv, Ming Yang, Marc’Aurelio Ranzato, and Lior Wolf. “Deepface: Closing the gap to human-level performance in face verification.” In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 1701-1708. 2014.

[14] Amos, Brandon, Bartosz Ludwiczuk, and Mahadev Satyanarayanan. “Openface: A general-purpose face recognition library with mobile applications.” CMU School of Computer Science (2016).

[15] Yi, Dong, Zhen Lei, and Stan Z. Li. “Towards pose robust face recognition.” In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 3539-3545. 2013.

[16] Lu, Can-Yi, Hai Min, Jie Gui, Lin Zhu, and Ying-Ke Lei. “Face recognition via weighted sparse representation.” Journal of Visual Communication and Image Representation 24, no. 2 (2013): 111-116.

[17] Yang, Allen Y., Zihan Zhou, Arvind Ganesh Balasubramanian, S. Shankar Sastry, and Yi Ma. “Fast $ell_ {1} $-Minimization Algorithms for Robust Face Recognition.” IEEE Transactions on Image Processing 22, no. 8 (2013): 3234-3246.

[18] Chan, Chi Ho, Muhammad Atif Tahir, Josef Kittler, and Matti Pietikäinen. “Multiscale local phase quantization for robust component-based face recognition using kernel fusion of multiple descriptors.” IEEE Transactions on Pattern Analysis and Machine Intelligence 35, no. 5 (2013): 1164-1177.

[19] Xu, Yong, Xingjie Zhu, Zhengming Li, Guanghai Liu, Yuwu Lu, and Hong Liu. “Using the original and ‘symmetrical face’ training samples to perform representation based two-step face recognition.” Pattern Recognition 46, no. 4 (2013): 1151-1158.

[20] Han, Hu, Shiguang Shan, Xilin Chen, and Wen Gao. “A comparative study on illumination preprocessing in face recognition.” Pattern Recognition 46, no. 6 (2013): 1691-1699.