Face Recognition Based on Artificial Neural Network: A review

The face recognition/detection process is considered one of the most popular applications in image processing and biometric pattern recognition systems. Although the face recognition approach improves the authentication procedure, many challenges still appear due to diversities in human facial expression, colossal image size, background complexity, variation in illumination, poses, blurry, etc. Therefore, the face detection procedure is classified as one of the most challenging tasks in computer vision. This research paper reviews the implementation of image processing based on the Artificial Neural Network approaches. ANNs represent it as a potential capability to enhance the method of extracting face patterns through an adaption of various ANN topologies. Furthermore, it means fundamental phases associated with the construction of any facial recognition system. Finally, it provides a comparison of different literature studies related to face recognition based on varying ANN approaches and critically analyzed them. was ranged from full connected network topology to partially connected network topology to distinguish among images that contain face pattern and those images with non-face patterns. The research results show that all three forms of architecture had achieved more than 96% correct classification of the face. The most efficient architecture achieved a 97.6% correct classification rate with 3.4% as the error rate. Bouzerdoum (2000) proposes an effective face detection system based on the use of Convolutional Feedforward Neural Network (CFNN) architecture composed of several


Introduction
Many organizations draw their attention to maintain high information security level and control the grant access to the system by authenticating the individual's identity to evade cybercrime issues. On the other hand, standard authentication techniques such as PIN, password, username, ID card number, etc., become inefficient techniques. With the excessive advances in software computing and image processing, many emerging methodologies appeared mainly to improve the authentication procedure (Bhattacharyya et al., 2009). The "Biometric" word is constructed through combing "Bio" and "metrics" words which refer to useful sciences that concern is analyzing and measuring biological information through adapting intelligence machine learning and various mathematical algorithms. Nowadays, Biometrics Pattern Recognitions (BPR) methodologies are widely useable systems. It was defined as an efficient information system that is mainly employed to detect, verify, and authenticating individuals' true identity based on some of its behavioral characteristics like body movement, keystroke writing style, and unique physiological factors such as fingerprint, eye retina, or eye iris, voice pattern, DNA, facial pattern and handshape (Bhattacharyya et al., 2009).
The recent development in machine learnings and technologies; makes it possible to generate an intelligence system based on statistical learning methodologies as a Facial Recognition (FR). Although, the FR system is considered one of the significant applications in image processing. However, this form of recognition may deliver great challenges in computer vision and pattern recognition due to many reasons such as diversities in facial expressions, orientation problems, illumination effects, and image size and background complexity (Khan et al., 2019).
Although many researchers place considerable efforts in this area to overcome the limitations of FR through the use of the Artificial Neural Network approach, many issues are still required to be solved. In general, the FR system analyzes the individual's facial characteristics from an image that is entered as input into the system. This image will go through a reprocess phase to extract all essential information based on using specific algorithms like the Deep Learning Algorithm (DLA) to recognize the target individual at the end (Khan et al.,2019).
This research paper will involve the following sections: Section 2 will briefly explain the structure of constructing any biometrics system. Besides, Section 3 will discuss the Artificial Neural Network concepts and their roles. This section will be concluded by explaining the work procedure for ANN. Section 4 will represent different literature studies about the face recognition system based on using various ANN architectures. It will clarify the strengths of architecture. Section 7 will critically analyze the obtained information of the literature studies and provide a brief comparison among different ANN topologies regarding specific factors such as performance rate, error rate, and several training data set. Finally, the research will end up with a conclusion.

Structure of Face Recognition System
The construction of any Biometrics Recognition system like face recognition consists of four main contemporary phases: face detection, preprocessing, feature extraction, and face recognition (Shaaban, 2021;Hassin & Abbood, 2021). It serves individuals' verification and identification purpose, as presented in Figure 1

Face Detection Phase
Detecting the target face image from a captured image or selected image from the DB is considered as the core function of this phase. Actually, the main purpose of the face detection process is to make sure and verify whether a given image has a face image region or not. When finish segmenting and detecting the target face area or region of concern, this output will be delivered into the preprocessing phase for further progressions (Zhao et al., 2003).

Preprocessing Phase
The image pre-processing steps usually forms as a combination of three important modules which are: histogram equalization, detection of edge, and matching of token that applied to enhance image quality, identify the edge point in the digital image, and finally perform removal and normalization based on specific algorithms (Al-Hatmi & Yousif, 2017; Hasoon, 2011). Through preprocessing technique, all undesirable image effects can be removed such as image noise, distortion, blur, shadow, or filters and it will make normalization for the image to generate smooth face image as an output which then will be utilized in extraction phase (Saudagare & Chaudhari, 2012) as shown in Figure 2.

Feature Extraction Phase
This stage will receive the detected face image region as an input. Through using feature extraction algorithms, all face characteristics will be extracted effectively from the face region such as the distances among eye, lip, and nose features (Saudagare & Chaudhari, 2012). The main purposes of feature extraction process are to perform specific functionalities including packing of information, cleaning of noise, and do salience extractions. After that, the obtained information is transferred into a vector for the subsequent process and use like comparison of obtained feature with stored data (Bhele & Mankar, 2012).

Face Recognition Phase
This is the last phase and it is utilized to achieve automatic authentication and identification of the individuals.
To achieve this goal, each face recognition system should maintain a face DB that stores information about all extracted faces features in which for each individual several images should be taken and then extracted features stored in this dedicated DB (Bhele & Mankar, 2012). Consequently, Figure 3 shows the extracted features information that is received from the previous phase will be compared to each face class that stored in DB to perform authentication and recognize the person and the algorithm return the identity (Raheja & Kumar U, 2010).

Artificial Neural Network Background
In the last decade, various models and architectures of Artificial Neural Network (ANN) have been developed and widely utilized for face detection and recognition based on the neural network (Kasar et al., 2016). An Artificial Neural Network (ANN) is an information processing paradigm that behaves like biological nervous systems. ANN is a powerful mathematical tool that processes input information to efficiently simulate or predicate the desired output The effectiveness of NN and its increased use could be due to its ability to work in a non-linear network (Boughrara et al., 2012). Therefore, the feature extraction phase of face characteristics through using NN is more effective and efficient than using the linear karhunnen-loeve technique (Lawrence et al., 1997). The first ANN technique utilized for face recognition is a single-layer adaptive network known as "WISARD" (Stonham et al., 1986). The WISARD comprises a distinct network for each stored person. Constructing the NN structure is critical for making a successful face recognition system, and the model that should be applied depends on the intended application objectives (Stonham et al., 1986). Commonly, Convolutional Neural Network (CNN), as well as Multi-layers Perception (MLP) structure, have been employed for the aim of face detection (Sung & Poggio,1995). On the other hand, the Multi-Resolution Pyramid structure has been applied efficiently for face verification purposes.
In normal, ANN is formed as a base of Deep Learning and subset of Machine Learning where the algorithms are inspired by the structure of the human brain and it is made up of many layers of neurons known as "artificial nodes" . These neurons are core passing units of the network. ANN is composed of three fundamental layers. The first layer is the input layer which receives the pixels as input, in between exist the hidden layers which perform most of the computation required by the network, and the final layer is the output layer that anticipates the ultimate output (Lawrence et al., 1997). In generic representation, ANN takes a set of data, train themself to recognize the parent that available in this data, then predict the output or new set of similar data (Tolba et al., 2006). The digital image is composed of many pixels as shown in Figure 4, each pixel expresses a numerical value that is fit as input to each neuron in the first layer. Neurons of one layer are connected to the subsequent layer through "channels" that sign specific numerical value known as "weight" (Agarwal et al., 2010). In case of the wrong predication, the information will be passed back to neurons via the "backpropagation" process with adjustment of inputs and weights in which these passes continue until the network predicate and recognize the face correctly (Agarwal et al., 2010).

Related Work
In the present study we examined patterns of responses for online and paper-based SET scores at a midsized, regional, comprehensive university in the United States.

Text body
The concept of face recognition isn't the newest subject in computer vision and for many past decades researcher's grape their attention in this area due to its practical importance and great abilities in enhancing the way of verification and authenticating an individual's identity. Despite the fact that state other identification methods like fingerprint, iris scanning, etc., are classified as more accurate identification tools, "face recognition has always remained as a major focus of most researchers because it is people's primary technique for identifying individual " (Kasar et al., 2016).
There are many of literature studies that focus on studying the modern ANN architectures and models which could be applied for constructing a successful face recognition system as shown in Table 1.

Results & Discussion
In Table 1 is composed of information related to face recognition/detection system based on different ANN topologies, these data have been collected from many previous related works that were conducted among the year 1997 to 2010. Each of these research studies describes specific ANN architecture that was applied to develop a face   Nevertheless, the detection rate is seemed to be the most used factor for assessing system efficiency compared with other measurement standards as shown in Figure 6. According to the diagram as illustrated in Figure 7 ). Despite, the graph shows that the highest detection level can be gained through the adaption of the CNN approach (Matsugu et al., 2003), and BPNN that was utilized in (Bojkovic & Samcovic, 2006) also acquire a good detection rate level. Furthermore, it was observed that each of these research papers have specific limitations as well as specific strengths over other topologies (Kaushal & Raina, 2010;Bouzerdoum, 2000). Table 3 presented the detection rate of all reviewed methods. Therefore, we can't specify exactly what is the most efficient architectures for building up a powerful face recognition/detection system with the highest accuracy and performance level.

Conclusion
Recently, different emerging techniques have been generated to improve identification and authenticating procedure for individuals and overcome weaknesses of traditional identification techniques. This paper aims to review and critically analyze the current methods to improve detection and validate the images. Also, the results show that the Biometric pattern recognitions are one of these modern techniques that have been applied widely to improve and enhance security levels and access control.