《Paravision：2021邊緣AI及其對面部識別的影響研究報告（英文版）（12頁）.pdf》由會員分享，可在線閱讀，更多相關《Paravision：2021邊緣AI及其對面部識別的影響研究報告（英文版）（12頁）.pdf（12頁珍藏版）》請在三個皮匠報告上搜索。

1、Septebmer 2021UNDERSTANDINGEDGE AIand its impact on face recognitionparavision.aiTrusted Vision AIUnderstanding Edge AI and its impact on face recognition ParavisionPage 2Already redefining the way we work,connect,create and live,edge computing and artificial intelligence have rapidly become indispe

2、nsable technologies.Edge computing brings data processing to the edge of the network where data is being created,reducing reliance on cloud computing and,by extension,triggering a cascade of compelling benefits for a wide range of applications.Artificial intelligence,meanwhile,promises the ability t

3、o make faster,more accurate decisions from highly complex data sources.Now,edge computing and AI have been brought together to form“Edge AI,”harnessing their respective benefits while creating wholly new opportunities as a result of their combination.These enormous benefits are connecting with the m

4、arket,resulting in massive predicted growth.According to market research firm IDC,the worldwide edge computing market will reach$250.6 billion in 2024 with a compound annual growth rate(CAGR)of 12.5%over the 20192024 forecast period.In addition to professional and provisioned services,hardware will

5、account for$80.7 billion in revenue and software will reach$54 billion in 2024.Part of the market growth is attributable to the benefits of edge computing,such as the reduction or elimination of delays for data transport,enhanced privacy and more.The other part will be the enablement of applications

6、 such as Edge AI that either werent possible before or were only partially effective.Edge AI itself will have enormous implications for face recognition systems,as it allows for drastically increased efficiency and usability,lower operational costs,and enhanced security and privacy protections.As we

7、 dive deeper into the strengths,utilities,and importance of edge computing and Edge AI,we will come to understand how these technologies can pair up and boost the potential of applications such as face recognition.IntroductionUnderstanding Edge AI and its impact on face recognition ParavisionPage 3W

8、hat is Edge Computing?Understanding Edge AI requires starting with defining edge computing.Edge computing is easy to understand,but hard to define because the definition can vary based on the application.In the most basic terms,edge computing places high-performance compute,storage and network resou

9、rces as close as possible to end-users and devices.In practice,there are many different places on a network that can serve as an edge.A widely used model of edge computing comes from the Linux Foundations LF Edge group shows this as a continuum from cloud to edge.The LF Edge model focuses on the two

10、 main edge tiers that straddle last-mile networks that connect to devices and people:the“Service Provider Edge”and the“User Edge.”Those tiers are further broken down into subcategories.According to this model,the user edgethe main area of relevance to our discussion of Edge AIconsists of:Self-contai

11、ned edge devices,such as smartphones,wearables and automobiles;Gateway devices such as IoT aggregators,switching and routing devices;On-premises server platforms.The Internet of Things(IoT)refers to all devices connected to the internet that can communicate with each other;this is the smart device e

12、dge in LF Edge parlance.There are also devices in the“constrained device edge”which may be connected via private networks or protocols other than TCP/IP and may even have intermittent connectivity.Understanding Edge AI and its impact on face recognition ParavisionPage 4Older generations of edge devi

13、ces were usually proprietary and closed,meaning that they werent programmable by the customer.They also tended to lack the processing power to run complex applications such as image and data processing.To compensate for this lack of compute horsepower,data would be fed to a server,where it would the

14、n processed before returning to the local devicea process that takes extra time and can add cost in the form of bandwidth fees,for instance.The figure above provides an abstracted view of how edge devices deliver data to the cloud for processing.In some cases,such as video security,on-premises serve

15、rs and gateways are used to aggregate data before delivery to a cloud for actions such as face recognition.The time and cost involved in the process are not optimal for many applications where control or instant answers are needed.Performing operations in real-time such as using facial recognition f

16、or access control requires low latency,and that means moving the analytics engine as close as possible to the data.In other words,Edge AI is needed.Understanding Edge AI and its impact on face recognition ParavisionPage 5The market for edge AI chipsets is expected to grow from$7.7 billion in 2019 to

17、$51.9 billion by 2025,according to a forecast from market research firm Omdia.Its not just hardware thats seeing advancements:the software used to build and deploy AI models and the models themselves are being optimized for edge environments.In some cases,tools are used to compress models to run on

18、edge devices.In other cases,approaches such as TinyML are built from the start to accommodate use in edge devices.All told,advancements in Edge AI are resulting in advanced computer vision capabilities in compact,power-efficient edge devices ranging from intelligent video cameras to biometric termin

19、als for automated access control.Edge inference is the form of AI that is most commonly being deployed in the context of Edge AI.Inference has emerged as a key edge computing workload according to market research firm Omdia.Many companies have introduced chipset solutions to accelerate these Edge AI

20、 workloads.Instead of using general-purpose CPUs or expensive GPUs,chips are being designed specifically to accommodate the distinct data processing patterns of AI.Moving inference to the edgeEdge AI is a rapidly evolving subset of the overall market for artificial intelligence that provides a solut

21、ion to the shortcomings of cloud-dependent AI.Simply stated,Edge AI refers to running AI models on a device that has the appropriate sensors and processors.Network connectivity is not required for the device to process data and take action.Edge AI utilizes a new generation of specialized hardware an

22、d software that runs AI models locally instead of on remote servers,at once taking advantage of the benefits of AI and edge computing.What is Edge AI?Understanding how AI can work at the edge means understanding the two components of AI:training and inference.Model training requires an iterative pro

23、cess of analyzing a large amount of historical data,detecting patterns in that data,and generating an algorithm for that kind of pattern detection.The model is checked for accuracy,and the process is repeated.The second component of AI,inference,takes the algorithm generated by training and analyzes

24、 new data to produce insights.Training vs inferenceUnderstanding Edge AI and its impact on face recognition ParavisionPage 6Edge AI has significant implications for facial recognition specifically,but it is valuable to understand the broader benefits of Edge AI for IoT and related applications.Why E

25、dge AI MattersProcessing data at the edge reduces the need to transmit information over a network.This enables more efficient usage of bandwidth,which consequently results in:Reduced operating costs High performance results even on a constrained networkBandwidth Efficiency Fully on-device processing

26、-Users no longer wait for information to return from a remote server;instead,all processing occurs on a local device,producing faster response times.Hybrid configuration-By combining edge with nearby server-side processing(on-premises,or at the access edge,in LF Edge parlance),the time required for

27、round trip communications with centralized cloud-based services is reduced.This adds latency compared to on-device processing but offers a compromise in the form of additional processing power and the potential for using on-demand edge cloud services from cloud providers or telcos,for example.Edge A

28、I enables low latency processing of data because data is either being moved a shorter distance or not at all(when processed on device):LatencyConventional processor technologies can be costly and consume vast amounts of power,while Edge AI chips can deliver results from a dramatically economized pow

29、er envelope.Edge AI chips typically consume between 1 to 5 watts(sometimes less),whereas typical CPUs and GPUs run in the range of 50W and more.This means Edge AI chips reduce the reliance on heat sinks and fans for cooling and consume less power.The finished products overall size and weight can be

30、greatly condensed,resulting in a smaller,simplified design.Reduced Size,Weight,and PowerUnderstanding Edge AI and its impact on face recognition ParavisionPage 7AdditionallyEdge AI can drive the use of dynamic,responsive,and interactive UI features,such as LCD or LED elements with reactive colors,sh

31、apes,and patterns.Edge AI diminishes the need to rely on cloud services for processing and storage,offering the potential for a lower total cost of ownership.Data points underpinning this opportunity are beginning to emerge:A study from Deloitte notes that an Edge AI chip will cost around the same a

32、mount as a smartphones processor while offering better performance and lower power consumption than traditional processor architectures.A study from HPE suggested that the total cost of ownership(TCO)of using cloud services for data analytics workloads was 1.7 to 3.4 times higher than comparable on-

33、premises deployments,with the latter offering workload throughput improvements of 45%over the cloud architecture.Hardware CostsEdge AI provides enhanced privacy protection for personal information.Processing data locally is inherently more secure than sending data across networks.Potentially sensiti

34、ve data can be managed at the source,enabling the enforcement of policies around data storage or masking.PrivacyDue to advances in chip design,Edge AI devices can now deliver advanced functionality without relying on network connectivity.This means that in cases of power or network outages(or interm

35、ittent connectivity)devices will continue to process data.Some of the most advanced Edge AI solutions have even been tailored for battery-powered operation,sipping power for months between recharge.Offline FunctionalityHigh latency can result in sluggish,non-intuitive user experiences.The low latenc

36、y brought about by Edge AI allows for immediate feedback,generating a more interactive and compelling UI/UX.UI/UX(Integration and Responsiveness)Understanding Edge AI and its impact on face recognition ParavisionPage 8How Edge AI Benefits Face RecognitionTraditional video-based face recognition solu

37、tions rapidly process video feed to check for visible faces before delivering the results back to the client.However,continuously sending video across a network consumes a large amount of bandwidth.With Edge AI,cameras can selectively record and send footage that only includes faces.Whether or not f

38、ace recognition is used,this method reduces the amount of data that is recorded and transmitted.That translates into lower bandwidth use and lower operational costs for transmission and storage compared to conventional server-side processing.Face recognition performance is heavily correlated with th

39、e quality of submitted face images.With Edge AI,real-time image quality assessment can ensure the submission of only high-quality face image samples from the edge device.Depending on the use case,if low-quality is detected,the edge device can notify the user and provide instantaneous feedback.Image

40、quality filters can be used in tandem with detection/activation technologies to further improve bandwidth efficiency.Detection/ActivationImage QualityUnderstanding Edge AI and its impact on face recognition ParavisionPage 9The enhanced power efficiency of Edge AIand,by extension,its reduced size and

41、 weightprovides product designers with a broader range of innovative industrial design options.With an Edge AI processor,it is possible to construct form factors better suited to specific architectural limitations and environments.For example,next-generation AI-powered access control devices could i

42、nclude video cameras and face recognition capabilities,all in the size of a typical card reader or consumer video doorbell.Industrial DesignIn applications or environments where the privacy of people in the field of view needs to be preserved,Edge AI can be used to detect faces and then remove,obscu

43、re,or encrypt them at the video feeds source.In short,systems would be outfitted with advanced privacy protection and eliminate personally identifiable information at the video source.PrivacyEdge AI can be used to differentiate between human beings and non-living spoofs in real-time.This security me

44、asureknown as Presentation Attack Detection(PAD)prevents attackers from potentially fooling and bypassing a biometric system.Moving PAD to the edge enables this accurate and powerful threat-detection capability to integrate more advanced multi-sensor technologies while being more readily deployable

45、and responsive.LivenessUnderstanding Edge AI and its impact on face recognition ParavisionPage 10The Edge AI and face recognition technologies outlined above can be applied to numerous automated identification and authentication scenarios.Here,we examine their use in access control,video security,an

46、d payments.Market Implications of Edge AI-Based Face Recognition Increase speed of recognition When used in conjunction with next-gen processors,offer on-device processing for 4K and 8K resolution video streams for increased detail Apply privacy filters when appropriate Fewer technical failures due

47、to low network-connectivity Decreased chance of information theft due to reduced data transmission Liveness detection and image quality filters will reinforce threat detection measures while still offering a streamlined user experienceIn controlled environments,ensuring a users identity is critical.

48、Bolstering face recognition systems with Edge AI will dramatically improve the speed,security,and reliability of access control devices.Other benefits include:Access ControlApplying Edge AI face recognition to video security can result in a more cost-effective,accurate,and user-friendly system.More

49、specifically,it can:Video Security Guarantee a productive and satisfying user experience by enabling a quick,interactive checkout process from retail stores Rapidly identify potential fraud activity due to advanced liveness detection In a world where brick and mortar stores struggle to compete with

50、the seamless experience provided by online retailers,Edge AI-based face recognition can deliver frictionless,touchless payments.For instance,it can:PaymentsUnderstanding Edge AI and its impact on face recognition ParavisionPage 11With rapid advances in imaging sensors,embedded computing,and deep lea

51、rning technologies,Edge AI-powered face recognition can support and empower businesses in a way that is more cost-effective,accurate,and user-friendly.Edge computing and Edge AI are foundational elements in computer vision-centric IoT applications,and their use will facilitate more widespread adoption and secure,appropriate use of face recognition.In ConclusionFor more information or to schedule a demo,please contact us at:infoparavision.aiTrusted Vision AI