Survey of Research on Face Recognition Methods Based on Depth Learning

The field of face recognition has recently become a quite popular area of research, which is of great significance to the development of technology. It introduced the definition of face recognition technology and its development process, as well as the technical advantages and application scenarios of the technology. with the development of deep learning, face recognition technology based on deep learning is gradually realized. firstly, the limitations of traditional face recognition technology are pointed out. Then, several popular face recognition methods based on depth learning are analyzed. Finally, this paper introduces the application of deep learning technology in face recognition, summarizes a new deep learning model based on big data, and summarizes and prospects the development of face recognition technology in the future.


Introduction
Face recognition technology refers to the processing and analysis of face images through computer programs, so as to recognize and verify face identity information .The development of face recognition technology can be traced back to the 1970s, when people began to explore how to use computer technology to complete image recognition.Over the past few decades, with the continuous development of computer technology and artificial intelligence, face recognition technology has also been continuously improved and refined.
The development of face recognition technology can be divided into three stages.The first stage is the traditional method based on geometric features, which is mainly based on the principle of geometric measurement and feature extraction, and realizes the recognition of face identity information through the feature calculation and comparison of face images.The second stage is the man-machine interactive recognition stage, which mainly uses geometric features to express the features of the front face image, but this stage still requires the operator's experience knowledge and can still achieve the goal of fully automatic recognition [1] .The third stage is a method based on deep learning, which uses deep neural networks for feature extraction and classification, and realizes accurate recognition of face identity information by learning more abstract and high-level feature information.
Face recognition technology has the advantages of high accuracy, fast recognition and no contact, and has been widely used in various fields.In the field of public security, face recognition technology can realize functions such as access control, video browsing security monitoring and so on [2] .In the financial field, face recognition technology can be used for identity verification, transaction confirmation and other functions.In the medical field, face recognition technology can be used for patient identity confirmation, medical record management and other functions.In addition, there are also applications of face recognition technology in smart home, education and other fields.
Deep learning is a branch of machine learning that attempts to deeply abstract data using algorithms that contain complex structures or multiple layers of processing made up of multiple non-linear transformations.Face recognition methods based on deep learning learn the ability to extract features in an end-to-end way, and use the extracted features for classification.Under the guidance of loss function, some optimization methods such as gradient descent and adaptive learning rate algorithm are used to optimize parameters in the neural network [3] , and finally realize image recognition.

Impact of Angle change
Traditional face recognition methods face many difficulties and challenges, one of which is the influence of Angle change.Because the Angle change will lead to the change of the shape and texture of the face image, so the traditional method is often difficult to accurately recognize the face.Especially in real life, people are in a variety of environments and angles, and it is difficult for traditional methods to meet the requirements of recognizing multiple angles.The traditional face recognition methods are basically based on the feature point method for recognition, so it will be affected by the change of face Angle, resulting in a decline in recognition accuracy.For example, in the case of face rotation, the position of the feature points will change, so it cannot be accurately matched, and the recognition accuracy will be seriously affected.In today's era, with the development and wide application of deep learning, face recognition research has made great breakthroughs, and its adaptability, accuracy, robustness and intelligence have been greatly improved [4] .

The influence of factors such as illumination, facial expression, and age on face recognition.
Factors such as illumination, expression and age are the main limitations of traditional face recognition methods [5] .Lighting factors can cause changes in the brightness and contrast of the face image, and even make the shape of the face change.Facial expression factors can also lead to the change of face image, which makes it difficult to achieve face feature extraction.The age factor also exists, which causes the details and features of the face image to gradually change over time, and the existing feature information will gradually be lost.
In recent years, due to the introduction of deep learning technology into the field of face recognition, these problems in traditional methods have been solved.By training a large number of face image data, the deep learning network can automatically extract face features that are unchanged under the influence factors such as illumination, expression, and age, thereby improving the accuracy of face recognition.At the same time, deep learning technology can also cope with the complexity that may exist in large-scale face recognition systems, making the recognition system more robust and reliable.

Convolutional Neural Networks (CNN)
Convolutional neural network [6] (CNN) is a very effective deep learning network model, which is widely used in face recognition technology.CNN is a deep feedforward neural network with features such as local connection and weight sharing [7] .The basic structure is input layer, conv layer, subsampling layer (pooling), full connection layer and output layer, as shown in Figure 1.The main idea of CNN is to extract image features through multiple convolution operations and pooling operations of the input image.By stacking multiple layers, the mapping relationship between the output and input of the classifier is finally obtained.For this simulation, 6127 images of 40 individuals from the MegaFace database are used, with 4742 images for training, 1185 for validation, and 200 for testing.As shown in Figure 3, the simulation achieves over 98% accuracy in face recognition, demonstrating the excellent applicability of CNNs for this task.
The MegaFace database provides a challenging benchmark for face recognition due to its large scale and diversity.The high accuracy reached by the CNN model on this dataset highlights the capabilities of deep convolutional neural networks for robust face representation and identification, even with variations in viewing angle, lighting, expression and age.
In summary, this simulation on a subset of the MegaFace database shows promising results, with the CNN model achieving over 98% accuracy.This indicates the strong potential of CNN-based approaches for face recognition applications with real-world datasets containing uncontrolled conditions.Further research can build on these results to handle even larger datasets and more difficult scenarios.

Deep Belief Network (DBN)
Deep Belief Network (DBN) [8] is an unsupervised deep generation model for feature extraction.It consists of multiple layers, each of which is a random variable of a binary variable.DBN has certain advantages in feature extraction.It can be trained by using backpropagation algorithm between upper and lower layers, so as to learn feature representations at different levels.By using DBN, extracted features can have a higher level of abstraction, which is especially important for complex tasks such as face recognition.
Compared with traditional neural networks, an important advantage of deep learning is that it largely solves the problem of training speed and accuracy of low-level neural networks.Before the global learning of the multi-layer model, as shown in Figure 3, DBN will decompose the neural network into a cascade of multiple constrained Boltzmann machines (RBMS), and then train it layer by layer [9] .In order to accurately describe the feature structure, DBN can learn abstract features of each layer from the bottom up.Different from the algebraic feature method, the feature extraction does not need manual selection, but is completely completed by automatic learning.One defect of DBN is that it directly uses the pixels of face image as the input for learning, often ignoring the local features of portrait.Under the influence of factors such as gesture, light and noise, the output feature expression may be unfavorable to the result [10] .In order to solve this problem, literature [11] extracts Gabor features as the input of DBN to carry out face recognition, and the recognition rate is as high as 92.7%.ZHAO [12] also proposed a face recognition method based on the combination of Gabor wavelet and DBN, which can effectively extract abstract features of portraits and effectively reduce the influence of posture and light on the recognition rate, thus achieving accurate recognition of portraits.

Local binary Mode (LBP) combined with deep learning
The combination of local LBP and deep learning is a new face recognition technology.LBP is an algorithm discovered by Ojala et al. and can describe local texture [13] , but its performance is poor in terms of Angle transformation and illumination transformation.Deep learning is a technology that has attracted much attention from researchers in recent years.It can quickly and effectively extract features from face images and has good generalization ability.Combining LBP features with deep learning can improve the accuracy and robustness of face recognition.
In 2018, Li [14] extracted the texture features of face images through LBP in order to overcome the poor feature representation ability of traditional face recognition algorithms and the problems such as sensitivity to light changes and noise interference, and then used the obtained texture features as the input of the convolution al network to extract the features obtained after processing each pooled layer from the convolutional network.The extracted features are cascaded and fused in the full connection layer to obtain the final classification features, and finally the Softmax classifier is used for classification and recognition.In the experiment, the face database is rotated at different angles to expand the database and verify the robustness of the algorithm.Experiments are carried out in ORL, YALE and AR databases respectively.Finally, the correct recognition rate reaches 98.6%, 95.6% and 98.9% respectively, which is higher than the classical recognition algorithm and the robustness is better than the comparison algorithm.
In 2020, Man Zhongang et al [15] .proposed a face recognition algorithm called BPBN that divided face images into blocks, locally applied the LBP operator, and combined it with a deep belief network.Firstly, the face images were divided into blocks, and the LBP was extracted from each block to generate a histogram.The LBP histograms were then combined into a new feature vector according to a certain order.Secondly, the obtained LBP features were used as input for the DBN, and greedy layer-wise training was performed.The trained deep belief network was then optimized using backpropagation (BP) algorithm.Finally, the trained deep belief network was used to recognize faces.Experiments on the ORL face database showed an identification rate of 96.0%.This method was compared with the traditional PCA algorithm integrated with SVM, and the identification rate showed a significant improvement.
In conclusion, on the basis of LBP features, using deep learning models such as convolutional neural network and deep belief network for feature fusion and classification can effectively improve the feature expression and classification accuracy of face images, and greatly overcome the shortcomings of traditional face recognition methods.

Feature extraction based on color and texture
Feature extraction method based on color and texture [16] .For face images, there are often different color and texture areas.Therefore, we use the feature extraction method of color and texture to obtain more accurate face recognition results.Specifically, we first quantize the color of the image, converting the color value of each pixel point into a discrete color to reduce the amount of computation.Then, we extract the texture features of the face image, including texture direction, gradient and histogram.Through the extraction of these features, the different regions of the face image can be effectively distinguished, so as to improve the accuracy of face recognition.
Feature extraction based on color and texture is a relatively simple but effective face recognition technology.It can help us extract more representative features from images and improve the accuracy of face recognition.

Multi-scale feature extraction
In face recognition technology, feature extraction is one of the most important steps.Nowadays, feature extraction methods based on deep learning have been widely adopted.Among them, multi-scale feature extraction method can effectively improve the accuracy of face recognition.
The multi-scale feature extraction method can extract multi-level image features by convolution and pooling of input images in different proportions.These features reflect the information of different detail levels, which can effectively improve the scale change problem in face images.
A common multi-scale feature extraction method is feature pyramid structure [17] .The structure consists of a number of images of different proportions, each image feature extraction, and then the features of different scales are fused together.This method can enrich the diversity of feature representation and thus improve the accuracy.
In addition, deep neural networks can also realize multi-scale feature extraction.For example, in a convolutional neural network, multiple convolution checks of different sizes can be used to carry out convolution operations on input images to obtain convolutional feature graphs of different scales [18] .Then, these feature maps are processed by aggregation and activation functions to obtain higher-level features.This method can extract the details of the image more precisely and make the features more distinguishable.

Summary and prospect
According to the above review, there are some defects in traditional face recognition methods, and the introduction of deep learning technology into face recognition is mainly discussed in the current two most popular deep models DBN, CNN and LBP combined with deep learning.At the same time, some techniques of feature extraction are discussed.From the research status at home and abroad, the application of face recognition technology based on DBN and CNN has become mature and achieved good results.However, these two models also found the same problem, it is difficult to completely solve the problem that the recognition rate is generally low in the case of relatively small data, but the third model also obtained better results in the case of relatively small data sets.
Face recognition technology based on deep learning has become one of the most popular research directions in the field of computer vision, and has been widely used.As the demand for face recognition technology continues to increase, related research is also continuing to advance.The multi-angle face recognition method based on deep learning described in this paper overcomes the problem that traditional face recognition methods are easily affected by Angle changes.At the same time, the introduction of deep learning technology has also improved the accuracy of face recognition.In the future, multi-modal fusion is expected to become the development direction of face recognition technology.Multimodal fusion refers to the fusion of information from different sensors or different modes to improve the accuracy and robustness of recognition.For example, multiple information such as face images, face textures, voices, and gestures can be fused together for a more comprehensive and accurate recognition of faces.On this basis, we can develop more intelligent and detailed face recognition applications to provide more reliable identity recognition solutions for various fields.

Figure 1 .
Figure 1.basic structure of CNN The application of CNN in face recognition technology mainly includes two aspects: feature extraction and face classification.In terms of feature extraction, CNN converts face images into feature maps, and performs convolution and pooling operations on the feature maps to obtain highly characterized image representations.

Figure 2 and
Figure2and Figure3shows the simulation results of a CNN-based face recognition system, including the training loss curve.It can be seen that the model achieves good performance.The dataset is based on a subset of the MegaFace face database, which has a large quantity of images and random variations.

Figure 2 .
Figure 2. Training model loss plot based on CNN.

Figure 3 .
Figure 3. Simulation results of face recognition data based on CNN

Figure 4 .
Figure 4. Structure diagram of DBN model