MobiDev：利用深度学习提高人工智能人脸识别精确度（英文版）（15页）.pdf

《MobiDev：利用深度学习提高人工智能人脸识别精确度（英文版）（15页）.pdf》由会员分享，可在线阅读，更多相关《MobiDev：利用深度学习提高人工智能人脸识别精确度（英文版）（15页）.pdf（15页珍藏版）》请在三个皮匠报告文库上搜索。

1、Table of ContentsHow Deep Learning Upgrades Face Recognition SoftwareLIMITATIONS OF THE TRADITIONAL APPROACH TO FACIAL RECOGNITIONAI-ENHANCED FACE RECOGNITIONHighlights of AI face recognition system softwareHOW FACE RECOGNITION WORKSFace recognition accuracy and how to improve itConclusion1Biometric

2、 identification of a person by facial features is increasingly usedto solve business and technical issues.The development of relevant automatedsystems or the integration of such tools into advanced applications has becomemuch easier.First of all,this is caused by the significant progress in AI facer

3、ecognition.In this article,we will explain what the components are of a facerecognition software and how to overcome the limitations and challenges ofthese technologies.You will find out how AI,namely Deep Learning,can improve the accuracyand performance of face recognition software,and how,thanks t

4、o this,it ispossible to train an automated system to correctly identify even poorly lit andchanged faces.It will also become clear what techniques are used to trainmodels for face detection and recognition.Do you remember trying to unlock something or validate that its you,with thehelp of a selfie y

5、ou have taken,but lighting conditions didnt allow you to dothat?Do you wonder how to avoid the same problem when building your appwith a face recognition feature?How Deep Learning Upgrades FaceRecognition SoftwareTraditional face recognition methods come from using eigenfaces to form abasic set of i

6、mages.They also use a low-dimensional representation of imagesusing algebraic calculations.Then the creators of the algorithms moved indifferent ways.Part of them focused on the distinctive features of the faces andtheir spatial location relative to each other.Some experts have also researchedhow to

7、 break up the images to compare them with templates.As a rule,an automated face recognition algorithm tries to reproduce theway a person recognizes a face.However,human capabilities allow us to storeall the necessary visual data in the brain and use it when needed.In the case of acomputer,everything

8、 is much harder.To identify a human face,an automatedsystem must have access to a fairly comprehensive database and query it fordata to match what it sees.2The traditional approach has made it possible to develop face recognitionsoftware,which has proven itself satisfactorily in many cases.The stren

9、gths ofthe technology made it possible to accept even its lower accuracy compared toother methods of biometric identification-using the iris and fingerprints.Automated face recognition gained popularity due to the contactless andnon-invasive identification process.Confirmation of the persons identit

10、y in thisway is quick and inconspicuous,and also causes relatively fewer complaints,opposition,and conflicts.Among the strengths that should be noted are the speed of dataprocessing,compatibility,and the possibility of importing data from most videosystems.At the same time,the disadvantages and limi

11、tations of the traditionalapproach to facial recognition are also obvious.LIMITATIONS OF THE TRADITIONAL APPROACH TO FACIALRECOGNITIONFirst of all,it is necessary to note the low accuracy in conditions of fastmovement and poor lighting.Unsuccessful cases with the recognition of twins,as well as exam

12、ples which revealed certain racial biases,are perceivednegatively by users.The weak point was the preservation of data confidentiality.Sometimes the lack of guaranteed privacy and observance of civil rights evenbecame the reason for banning the use of such systems.Vulnerability topresentation attack

13、s(PA)is also a major concern.The need arose both toincrease the accuracy of biometric systems,and to add to them the function ofdetection of digital or physical PAs.However,the traditional approach to face recognition has largely exhausted itspotential.It does not allow using very large sets of face

14、 data.It also does notensure training and tuning identification systems at an acceptable speed.AI-ENHANCED FACE RECOGNITIONModern researchers are focusing on artificial intelligence(AI)to overcome theweaknesses and limitations of traditional methods of face recognition.Therefore,in this article we c

15、onsider certain aspects of AI face recognition.Thedevelopment of these technologies takes place through the application of3advances in such subfields of AI as computer vision,neural networks,andmachine learning(ML).A notable technological breakthrough is occurring in Deep Learning(DL).DeepLearning i

16、s part of ML and is based on the use of artificial neural networks.Themain difference between DL and other machine learning methods isrepresentation learning.Such learning does not require specialized algorithmsfor each specific task.Deep Learning owes its progress to convolutional neural networks(C

17、NN).Previously,artificial neural networks needed enormous computing resources forlearning and applying fully connected models with a large number of layers ofartificial neurons.With the appearance of CNN,this drawback was overcome.Inaddition,there are many more hidden layers of neurons in neural net

18、worksused in deep learning.Modern DL methods allow training and use of all layers.Among the ways of improving neural networks for face recognition systems,it isappropriate to mention the following:Knowledge distillation.A combination of two similar networks of differentsizes,where the larger one tra

19、ins the smaller one.As a result of training,asmaller network gives the same result as a large one,but it does it faster.Transfer learning.Focused on training the entire network or its specificlayers on a specific set of training data.This creates the possibility ofeliminating bottlenecks.For example

20、,we can improve accuracy by using aset of images of exactly the type that errors occur most often.Quantization.This approach aims to speed up processing by reducing thenumber of calculations and the amount of memory used.Approximationsof floating-point numbers by low-bit numbers help in this.Depthwi

21、se separable convolutions.From such layers,developers createCNNs that have fewer parameters and require fewer calculations butprovide good performance in image recognition,and in particular,faces.Regarding the topic we are considering,it is important to train a Deepconvolutional neural network(DCNN)

22、to extract from images of faces uniquefacial embeddings.In addition,it is crucial to provide DCNN with the ability tolevel the impact of displacements,different angles,and other distortions in theimage on the result of its recognition.Due to the data augmentation,the images4are modified in every way

23、 before training.This helps mitigate the risksassociated with different angles,distortions,etc.The more variety of imagesused during training,the better the model will generalize.Let us remember the main challenge of face recognition software development.This is the provision of fast and error-free

24、recognition by an automated system.In many cases,this requires training the system at optimal speed on very largedata sets.It is deep learning that helps to provide an appropriate answer to thischallenge.Highlights of AI face recognition systemsoftwareAs we said above,at the moment,when deciding how

25、 to build a face recognitionsystem,it is worth focusing on Convolutional Neural Networks(CNN).In thisarea,there are already well-proven approaches to creating architecture.In thiscontext,we can mention residual neural network(ResNet),which is a variant of avery deep feedforward neural network.And,fo

26、r example,such a solution as5EfficientNet is not only the architecture of a convolutional neural network butalso a scaling method.It allows uniform scaling of the depth and width of theCNN as well as the resolution of the input image used for training andevaluation.Periodically,thanks to the efforts

27、 of researchers,new architectures of neuralnetworks are created.As a general rule,newer architectures use more and morelayers of deep neural networks,which reduces the probability of errors.It is truethat models with more parameters may perform better,but slower.This shouldbe kept in mind.When consi

28、dering face recognition deep learning models,the topics of thealgorithms that are embedded in them and the data sets on which they aretrained come to the fore.In this regard,it is appropriate to recall how facerecognition works.HOW FACE RECOGNITION WORKSThe face recognition system is based on the se

29、quence of the followingprocesses:Face detection and capture,i.e.identification of objects in images or videoframes that can be classified as human faces,capturing faces in a givenformat and sending them for processing by the system.Normalization or alignment of images,i.e.processing to prepare forco

30、mparison with data stored in a database.Extraction of predefined unique facial embeddings.Comparison and matching,when the system calculates the distancebetween the same points on the images and then infers face recognition.6The creation of artificial neural networks and algorithms is aimed at learn

31、ingautomated systems,training them on data,and detecting and recognizingimages,including all of the above stages.Building AI face recognition systems is possible in two ways:1.Use of ready-made pre-trained face recognition deep learning models.Models such as DeepFace,FaceNet,and others are specially

32、 designed forface recognition tasks.2.Custom model development.When starting the development of a new model,it is necessary to define severalmore parameters.First of all,this concerns the inference time for which theoptimal range is set.You will also have to deal with the loss function.With itshelp,

33、you can,by calculating the difference between predicted and actual data,evaluate how successfully the algorithm models the data set.Triplet loss andAM-Softmax are most often used for this purpose.The triplet loss functionrequires two images anchor and positive of one person,and one more image negati

34、ve of another person.The parameters of the network are studied inorder to approximate the same faces in the functionality space,and conversely,to separate the faces of different people.The standard softmax function uses7particular regularization based on an additive margin.AM-Softmax is one of thead

35、vanced modifications of this function and allows you to increase the level ofaccuracy of the face recognition system thanks to better class separation.For most projects,the use of pre-trained models is fully justified withoutrequiring a large budget and duration.Provided you have a project team ofde

36、velopers with the necessary level of technical expertise,you can create yourown face recognition deep learning model.This approach will provide thedesired parameters and functionality of the system,based on which it will bepossible to create a whole line of face recognition-driven software products.

37、Atthe same time,the significant cost and duration of such a project should betaken into account.In addition,it should be remembered how facial recognitionAI is trained and that the formation of a training data set is often a stumblingblock.Next,we will touch on one of the main potentials that rely o

38、n face recognitionmachine learning.We will consider how accurate facial recognition is and how toimprove it.Face recognition accuracy and how toimprove itWhat factors affect the accuracy of facial recognition?These factors are,first ofall,poor lighting,fast and sharp movements,poses and angles,and f

39、acialexpressions,including those that reflect a persons emotional state.It is quite easy to accurately recognize a frontal image that is evenly lit and alsotaken on a neutral background.But not everything is so simple in real-lifesituations.The success of recognition can be complicated by any change

40、s inappearance,for example,hairstyle and hair color,the use of cosmetics andmakeup,and the consequences of plastic surgery.The presence in images ofsuch items as hats,headbands,etc.,also plays a role.The key to correct recognition is an AI face recognition model that has anefficient architecture and

41、 must be trained on as large a dataset as possible.Thisallows you to level the influence of extraneous factors on the results of imageanalysis.Advanced automated systems can already correctly assess the8appearance regardless of,for instance,the mood of the recognized person,closed eyes,hair color ch

42、ange,etc.Face recognition accuracy can be considered in two planes.First of all,we aretalking about the embeddings matching level set for specific software,which issufficient for a conclusion about identification.Secondly,an indicator of theaccuracy of AI face recognition systems is the probability

43、of their obtaining acorrect result.Lets consider both aspects in turn.We noted above that the comparison ofimages is based on checking the coincidence of facial embeddings.A completematch is possible only when comparing exactly the same images.In all othercases,the calculation of the distance betwee

44、n the same points of the imagesallows for obtaining a similarity score.The fact is that most automated facerecognition systems are probabilistic and make predictions.The essence ofthese predictions is to determine the level of probability that the two comparedimages belong to the same person.Lets as

45、sume that based on the results of the image analysis,the systemcalculated the similarity score at the level of,for example,0.99(99%).This meansthat there are grounds for non-absolute,but 99 percent confidence that theperson is correctly recognized.But most scenarios of automatic face recognitionrequ

46、ire the formation of an unambiguous conclusion about identification aswell.For this,before the development,a limit is set,which will be consideredminimally sufficient for a positive conclusion.It is often called the similaritythreshold.The choice of the threshold is usually left to the software deve

47、lopmentcustomer.A high threshold may be accompanied by certain inconveniences forusers.Lowering the similarity threshold will reduce the number ofmisunderstandings and delays,but will increase the likelihood of a falseconclusion.The customer chooses according to priorities,specifics of theindustry,a

48、nd scenarios of using the automated system.Commonly,it is recommended to set the similarity threshold at the level of 0.99(99%)in those areas where errors are unacceptable.It is about security andprotection of public order,etc.For less responsible and risky situations,it is9possible to set a lower t

49、hreshold.As a rule,this parameter ranges from 0.80(80%)to 0.99(99 percent).The approach to the similarity threshold influences the choice of technology anddevelopment parameters.In other words,automated facial recognition systemsfor access to a vault with the states gold reserve and admission to a c

50、orporateparty are developed differently.Lets move on to the accuracy of AI face recognition in terms of the proportion ofcorrect and incorrect identifications.First of all,we should note that the resultsof many studies show that AI facial recognition technology copes with its tasks atleast no worse,

51、and often better than a human does.As for the level ofrecognition accuracy,the National Institute of Standards and Technologyprovides convincing up-to-date data in the Face Recognition Vendor Test(FRVT).According to reports from this source,face recognition accuracy can be over99%,thus significantly

52、 exceeding the capabilities of an average person.By the way,current FRVT results also contain data to answer common questionsabout which algorithms are used and which algorithm is best for facerecognition.When familiarizing with examples of practical use of the technologies,the clientaudience is oft

53、en curious about whether face recognition can be fooled orhacked.Of course,every information system can have vulnerabilities that haveto be eliminated.At the moment,in the areas of security and law enforcement,where the life andhealth of people may depend on the accuracy of the conclusion about thei

54、dentification of a person,automated systems do not yet work completelyautonomously,without the participation of people.The results of the automatedimage search and matching are used for the final analysis by specialists.For example,the International Criminal Police Organization(INTERPOL)uses theIFRS

55、 face recognition system.Thanks to this software,almost 1,500 criminalsand missing persons have already been identified.At the same time,INTERPOLnotes that its officers always carry out a manual check of the conclusions ofcomputer systems.10Either way,the AI face recognition software helps a lot by

56、quickly samplingimages that potentially match what is being tested.This facilitates the task ofpeople who will assess the degree of identity of faces.To minimize possibleerrors,multifactor identification of persons is used in many fields,where otherparameters are evaluated in addition to the face.In

57、 general,in the world of technology,there is always a kind of race betweenthose who seek to exploit technological innovations illegally and those whooppose them by protecting peoples data and assets.For example,the surge ofspoofing attacks leads to the improvement of anti-spoofing techniques andtool

58、s,the development of which has already become a separate specialization.Read also:Anti-spoofing techniques for liveness detection in face recognitionVarious tricks and devices have been invented recently for computer visiondazzle.Sometimes such masking is done to protect privacy and ensure thepsycho

59、logical comfort of people,and sometimes with malicious purposes.However,automated biometric identification through the face can undoubtedlyovercome such obstacles.The developers include in the algorithms methods ofneutralization of common techniques of combating face recognition.In this context,it i

60、s useful to recall the relatively high accuracy of neuralnetworks facial recognition for people wearing medical masks,demonstratedduring the recent COVID-19 pandemic.Such examples instill confidence in thereality of achieving high face recognition accuracy even under unfavorablecircumstances.The way

61、s to increase the accuracy of facial recognition technology are throughthe enhancement of neural network architectures,and the improvement ofdeep learning models due to their continuous training on new datasets,whichare often larger and of higher quality.Significant challenges in the development of

62、automated systems are also theneed to reduce the recognition time and the number of system resources,without losing accuracy.At the moment,the technical level of advanced applications already allows toanalyze the image and compare it with millions of records within a few seconds.11An important role

63、is played by the use of improved graphical interfaces.Performing face recognition directly on peripheral devices is also promisingbecause it allows you to do without servers and maintain user data security bynot sending it over the Internet.ConclusionSo,we considered how facial recognition uses AI a

64、nd,in particular,machinelearning.We have listed the main areas of development of these technologies.Touching on the technical aspects of creating automated systems for neuralnetworks facial recognition,we identified common problems that arise in thisprocess and promising ways to solve them.From this

65、 article,you learned how AI face recognition works and whatcomponents it consists of.Also,we did not overlook the topic of the accuracy ofthis process.In particular,we revealed how to improve face recognitionaccuracy.You will be able to use the knowledge obtained from this article toimplement your ideas in the research field.Get in touch with our AI team and use their expertise to prepare your facerecognition deep learning projects.12