Covered And Camouflaged Face Recognition Based On Variational Feature Deep Learning

As a non-contact biological facial feature recognition technology,face recognition is technical research with important theoretical and engineering value in the field of artificial intelligence.After nearly 20 years of research and development,this technology has achieved rich academic research results and practical application in people’s daily lives.At present,such as:attendance,access control,security and other face recognition problems in restricted scenarios,face recognition technology has achieved recognition capabilities that can replace the human eye or even surpass humans.However,for a series of complex and special practical application scenarios such as shooting angles of different equipment,shooting under different lights,face camouflaged or partial covered of faces,face recognition technology still faces no small challenge.These are the core difficult issues that this paper is trying to overcome.In view of the above challenges,this study proposes a series of innovative solutions to improve the accuracy and performance of face recognition models.The contributions of the paper can be summarized as followsFirst.Face recognition based on filtering algorithms and a new Variational Feature Learning（VFL）loss function:When faces are in different poses,different lighting conditions,or even faces appear at different angles,it is difficult for existing face recognition technologies to correctly identify the identity of a face.To solve this problem,this contribution proposes a new algorithm called Filter.In the training process,start by learning the face representation from the whole face,but in order to get rid of the interference of unnecessary features,this contribution proposes this new algorithm,which can filter out the most important features from the face representation,and after the filtering operation,the feature vector of the face is halved,and the calculation amount of the algorithm is significantly reduced.At the same time,this contribution improves the VFL loss function during training.In VFL,it uses two fully connected layers with the same inputs and outputs to predict the mean and standard deviation of a Gaussian distribution.Because the inputs and outputs of the two fully connected layers are the same,in view of this problem,this contribution uses only the first fully connected layer in the network to predict the mean and standard deviation of the Gaussian distribution.The resulting mean and standard deviation are then used to calculate the Kullback–Leibler（KL）divergence loss.When faces are presented at any angle on the image,most face detection technologies have some problems when locating faces,such as inaccurate positioning or even inability to locate.In response to these problems,this contribution uses rotation techniques in which the image rotates step by step from 0°to 360°until the face is found in the image,which is then passed to the model as a new image of the face.In the process of the experiment,this contribution used two common datasets to evaluate the model,namely LFW and YTF.On each dataset,good experimental results were obtained.On the LFW,the recognition accuracy of the model reached 99.70%;On the YTF,the recognition accuracy of the model reached 94.02%.To further validate the performance of model recognition of faces under different lighting or at different angles,this contribution created a new dataset with a total of 55 identities.In order to obtain face images at different angles and under different lights,this contribution rotates the face images of each identity by a corresponding number of degrees,and then applies the gamma correction method to change the lighting of each rotated image.The model was evaluated again on this new dataset,and good experimental results were obtained,and the recognition accuracy of the model reached 89.95%.Second.Face camouflaged/partial coverage face recognition based on the loss function of dual variational feature learning（DVFL）:Recognizing faces that are deliberately camouflaged or covered is a common challenge for face recognition.In response to this challenge,this contribution proposes a model based on the loss function of Double Variational Feature Learning（DVFL）to identify the above complex faces.The so-called DVFL loss function refers to two different KL divergence loss functions.The two fully connected layers in the network are used to predict the mean and standard deviation of the Gaussian distribution,and then the predicted mean and standard deviation are used to calculate the first KL divergence loss function₁.The second KL divergence loss function,₂ is the KL loss function built into the Keras package.During training,the two loss functions are merged,and the combined loss function is used as the loss of the entire network.When the face is camouflaged or the face is partially covered,the face has undergone a large transformation,if the entire face is used to extract the feature,then there will be many interference factors in the extracted features,which will reduce the recognition accuracy of the model.When people try to use makeup tools or glasses to camouflage their eyes,it is unreliable to identify the identity of the face by relying only on the eyes,so the area around the human eyes should also be selected.In response to the above two problems,this contribution uses a part of the face that contains the human eyes and the area around the human eyes to train the network,which carries a lot of information about identity and is not easy to hide.If there are still some major changes in the selected part of the face,in response to this problem,this contribution uses a L2 normalization layer in the network,which is used to reduce the impact of these changes on recognition.During the experiment,the model was evaluated using four public datasets:LFW,FEI,MFR2,and PKU-Masked-Face.On each dataset,good experimental results were obtained.On the LFW,the recognition accuracy of the model is 95.62%;On FEI,the recognition accuracy of the model is98.3%,which is considered the most advanced result;On the MFR2,the recognition accuracy of the model is 93.31%;On PKU-Masked-Face,the recognition accuracy of the model is 87.96%.To further validate the performance of the model,this contribution creates a new dataset that is used to evaluate the model again.The dataset contains normal face images,partially covered face images,and faces camouflaged,in which 5 camouflage effects are used to achieve facial camouflage,namely crayon painted faces,painted faces,vampire tooth faces,wildcat faces,and cartoon faces.During the experiment,normal face images are used for training,and faces that are camouflaged and partially covered are used for testing,where each special type of face image is tested separately as a group,and then the corresponding experimental results are obtained separately.Experimental results show that when recognizing partially covered faces,the recognition accuracy of the model is 95.47%;When recognizing cartoon faces,the recognition accuracy of the model is 86.80%;When recognizing crayon painted faces,the recognition accuracy of the model is 90.91%;When identifying vampire tooth faces,the recognition accuracy of the model is 87.04%;When recognizing a feral cat face,the recognition accuracy of the model is 75.51%;When recognizing a painted face,the recognition accuracy of the model is 98.15%Third.Gender recognition of normal/facial camouflage/partially covered faces based on the loss function of dual variational feature learning（DVFL）:Gender carries a plentiful of information related to male and female social activities,such as in some places that restrict gender,such as baths,temples,gender is very critical,so gender recognition has always received widespread attention from scholars.Most of the previous gender recognition research was conducted on the whole face,but when encountering some complex situations,such as:the face is covered by a mask,scarf,etc.,or people deliberately camouflage their faces through some makeup techniques,then gender recognition for the whole face will become unreliable.In response to this problem,this contribution proposes gender recognition based on the loss function of dual variational feature learning（DVFL）,which aims to study the gender recognition of faces that are camouflaged or partially covered on the basis of recognizing the gender of normal faces.As with Contribution 2,instead of using the entire face training network,this contribution trains the network using a part of the face that contains the human eyes and the area around the human eyes,which carries a lot of information about identity and is not easy to hide.Also like Contribution 2,the DVFL loss function proposed in this contribution refers to 2 different KL divergence loss functions.The two fully connected layers in the network are used to predict the mean and standard deviation of the Gaussian distribution,and then the predicted mean and standard deviation are used to calculate the first KL divergence loss function₁.The second KL divergence loss function,₂ is the KL loss function that is built into the Keras package.The two loss functions are merged,and the combined loss function is the loss of the entire network.During the experiment,this contribution used five public datasets to evaluate the model,namely FEI,SCIEN,AR FACES,LFW,ADIENCE.On each dataset,good experimental results were obtained.On the FEI dataset,the recognition accuracy of the model is 99.51%;On the SCIEN dataset,the recognition accuracy of the model is 96.22%;On the AR FACES dataset,the recognition accuracy of the model is 99.39%;On the LFW dataset,the recognition accuracy of the model is 97.90%;On the ADIENCE dataset,the recognition accuracy of the model is 98.23%.To further validate the model’s ability to recognize the gender of a face camouflaged,this contribution created a new dataset that was used to evaluate the model again.First,some images were randomly selected from the AR FACES dataset,including a certain number of male images and a certain number of female images.Then,in order to achieve the purpose of facial camouflage,five camouflage effects are applied to each extracted image,namely crayon painted face,painted face,vampire tooth face,wild cat face and cartoon face.Finally,the camouflaged face images needed for the experiment were obtained.During the test,face images of each camouflage effect were tested as a group individually,and then the corresponding experimental results were obtained separately.Experimental results show that when recognizing cartoon faces,the recognition accuracy of the model is 80.18%;When recognizing crayon painted faces,the recognition accuracy of the model is 89.50%;When identifying vampire tooth faces,the recognition accuracy of the model is 76.60%;When recognizing the face of a wildcat,the recognition accuracy of the model is 95.13%;When recognizing painted faces,the recognition accuracy of the model is 94.96%,which is considered to be the most advanced results.Forth.ethnics recognition based on variational feature learning（VFL）loss function:At present,in face recognition,there is a lack of research on classifying people’s ethnics by face.At the same time,in order to verify that the new VFL loss function proposed in contribution 1 can indeed improve the recognition accuracy of the model.In response to the above two questions,this contribution tests the new VFL loss function in the field of ethnics identification.During training,the model is trained with a new VFL loss function,only the first fully connected layer in the network is used to predict the mean and standard deviation of the Gaussian distribution.The resulting mean and standard deviation are then used to calculate the Kullback–Leibler（KL）divergence loss.Different from Contribution 2 and Contribution 3,this contribution inputs the entire face into the model to obtain the features of the face,and then uses the features of the whole face to train the model and perform the final racia classification.Although there are many large databases of facial images on the Internet,none of these databases can meet the research purposes of this contribution,so this contribution uses a private data set from the published paper to evaluate the model.The dataset contains three ethnic groups of faces,namely China,Pakistan,and Russia.On this dataset,the classification accuracy of the model reached 96.85%,which is considered to be high classification accuracy compared with previous related studies.From the above four contributions,it can be summarized that the VFL loss function and DVFL loss function proposed in this paper can improve the accuracy and performance of model recognition of faces in some complex scenarios,especially the recognition of those faces that are camouflaged or partially covered,which has certain application value in many fields.First.Security field:primary and secondary schools are densely populated,students have weak ability to prevent injuries,once there are criminals provoking trouble,it is inevitable that they will cause certain harm to students,especially those who deliberately cover or camouflage their faces in order to hide their identities,and the harm caused by them will be greater.Therefore,for the face recognition system deployed in the crowded area of the campus,it is necessary to improve the ability of the system to recognize the face that is camouflaged or partially covered,so that it can quickly and accurately identify the above special types of faces.This is exactly the focus of this paper,so the technique presented in this paper can be applied here.Finally,the results of the system identification are compared with the data of the Public Security System,so as to find the criminals,avoid their harm to the campus,and achieve early detection and early prevention.Second.Life field:Due to the epidemic,for the safety of themselves and others,everyone must wear a mask to go out,and it is not easy to take off the mask,so in the face payment,access control and other occasions that require face recognition,the corresponding equipment can correctly and quickly identify the face covered by the mask is very important.The VFL loss function and DVFL loss function proposed in this paper can improve the accuracy of the model to recognize faces that are partially covered,so the technology in this paper can be applied here,which shows that this new technology has great application value in the field of life.Third.Medical field:This paper also attempts to train an image dataset related to the field of medical image classification using NN4 with the VFL loss function and the DVFL loss function.The model is evaluated on the test image,and the experimental results show that the model has high recognition and classification performance.This shows that the technology proposed in this paper can not only be applied to the field of face recognition,but also has certain application value in the field of medicine.Face recognition in complex scenes is a challenging subject,and different equipment shooting angles,shooting under different lights,face camouflaged,and local covered will bring certain difficulties to face recognition.Based on these difficulties,this paper has carried out a series of studies,but the techniques proposed in this paper and other existing technologies cannot completely solve above difficult problems.Therefore,there is still a lot of work to be further done in the future.First.Improve face detection algorithms:There are many face detection technologies,but there are still some shortcomings,such as when the face appears on the image at different angles or when there is a face-like interference on the image,most face detection models cannot accurately detect the face.Correctly detecting a face on the image is one of the prerequisites for the face to be correctly recognized,so the above deficiencies will affect the accuracy of the final face recognition.Therefore,it is necessary to design and implement a reliable face detection algorithm,which is the next work to be done in this paper.Second.Improve the generalization of special face recognition:In this paper,for the face whose face is partially covered,the default is that people’s eyes are not covered,but once the human eyes are covered,neither the technology proposed in this paper nor the previous related research can accurately identify the identity of the corresponding face according to the characteristics of other parts of the face.Therefore,identifying the identity of a face based on any part of the face,rather than just relying on the human eyes,is a task that needs to be challenged in the future.Third.Try to use face image completion technology:This paper mainly relies on the characteristics of the eyes when identifying faces that are camouflaged or partially covered,but in general,the identity of the face will be more accurately recognized according to the characteristics of the entire face,so this paper wants to try to apply the face image completion technology to the new technology proposed in this paper in the future.When identifying a face that is camouflaged or partially covered,the features are analyzed according to the non-missing part of the face and the missing part of the face is predicted,so as to achieve face completion,so as to obtain a complete face,and finally the corresponding identity is more accurately identified according to the complete face.