CN205665729U - Cell convolutional neural network intelligent vision pays accelerator - Google Patents
Cell convolutional neural network intelligent vision pays accelerator Download PDFInfo
- Publication number
- CN205665729U CN205665729U CN201620177655.0U CN201620177655U CN205665729U CN 205665729 U CN205665729 U CN 205665729U CN 201620177655 U CN201620177655 U CN 201620177655U CN 205665729 U CN205665729 U CN 205665729U
- Authority
- CN
- China
- Prior art keywords
- cell
- convolutional neural
- image
- chip
- machine vision
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Expired - Fee Related
Links
Landscapes
- Image Analysis (AREA)
Abstract
The utility model provides a cell convolutional neural network intelligent vision pays accelerator belongs to the artificial intelligence technique field. The utility model discloses use cell convolutional neural network as the core, microprocessor is connected to cell convolutional neural network chip, microprocessor connects memory module, the camera, the communication interface module, a microphone, the speaker, necessary peripheral part such as touch display screen and power, constitute an intelligent vision payment accelerator that has the pattern recognition function, image or pronunciation to specific sign, recognition speed and precision have been improved, better security and convenience is provided for user and system, the market potential is gigantic.
Description
Technical field
This utility model belongs to field of artificial intelligence, specifically, relates to cell/convolutional neural networks
(Cellular/Convolutional Neural Networks, CNN) intelligent vision pays accelerator.
Background technology
This utility model is that artificial intelligence technology is applied in E-Payment field.
Up to now, human history has been subjected to three industrial revolution, respectively mechanized manufacturing industry revolution, electricity vapour industry leather
Life, computer industry revolution, corresponding machine power problem, energy source problem, information processing and the transmission problem of solving, all
The dramatic change of the productivity is brought for human society.At present, the whole world, still in the constant quest of the third time industrial revolution, is wrapped
Include the Internet, mobile Internet tide all belongs to one series derivatives phenomenon.The industrial revolution next time will be once can be complete
Substitute artificial intelligence's revolution of people role.Substantially, it is that the mankind replicate another oneself, utilizes the robot manufactured complete
Full replacement self is engaged in autonomous, intelligent behavior.The most so vigorous fourth industrial revolution, the most quietly to
We come up.
Having produced when artificial intelligence is at the beginning of computer, it is by as academic the earliest in 1956
Section sets up.From that time, scientific circles are considered as will there be big breakthrough in this regard every about 10 years.?
In the neurological research of 1940, people just have been found that the brain of the mankind is actually a neutral net, Alan
Turing thinks that it is feasible for building a word audio based on this concept.Nineteen fifty-one, the postgraduate Marvin of 24 years old
Minsky has built the first in the world neural network machine, and this machine is also referred to as SNARC, and it is also First people in history simultaneously
Work self-teaching machine.
Machine learning belongs to a branch of artificial intelligence, and neutral net is a branch of machine learning, is the most real
With, the branch of a maximally effective artificial intelligence.
As far back as 1988, doctor Yang Lin delivered following two " cell neural network " papers producing extensively impact and phase
Pass patent of invention:
Leon O, Chua;Lin Yang, " Cellular Neural Networks:Theory ", IEEE
Trans.Circuits and Systems, vol.35 (10) Oct.1988, pp.1257-1272.
Leon O, Chua;Lin Yang, " Cellular Neural Networks:Applications, " IEEE
Trans-Circuits and Systems, vol.35 (10) Oct.1988, pp.1273-1290.
Leon O, Chua;Lin Yang, " Cellular Neural Network ", United States Patent,
Patent Number:5,140,670, Date of Patent:Aug.18,1992.
In paper, doctor Yang Lin proposes the basic concept of several key: parallel processing, analog circuit, neighborhood are straight
Connect in succession, non-neighborhood indirect action, nonlinear device, multitiered network, convolution operator, parameter reconfigure, are applied at image
Reasons etc., the development for neutral net is had laid a good foundation, and concrete ins and outs describe the most again.
The concept of degree of depth study (Deep Learning) is proposed in 2006 by Hinton et al..In degree of depth learning areas
Breakthrough caused artificial intelligence's revolution.In recent years, the company such as Microsoft, face book, Google, IBM, Baidu is proposed the respective degree of depth
Learning system, " degree of depth study " technology of utilization proposes many voices and the identification of image, composition algorithm.These algorithms are a kind of
The algorithm of computer simulation human brain neural network.In simple terms, it is simply that build an artificial neural network with computer, then lead to
Cross existing mass data constantly train optimization it.
The training method of artificial neural network is by showing substantial amounts of training example to it, the most gradually to network parameter
It is adjusted, until it can feed back gratifying classification.One typical network is (the deepest by 10-30 layer
Reaching 150 layers) artificial neuron piles up framework.Illustrate, when a pictorial information is sent to a nerve
During network, result is exported to next level after receiving information and carrying out the process of low level by input layer, goes round and begins again, directly
To arriving last level, determine the analysis result of this image.
In order to obtain more preferable learning outcome, the scale of neutral net is increasing, and the number of plies gets more and more, and just becomes deep
Degree neutral net, deep neural network is an important branch of degree of depth study.The deep neural network of Google the earliest
1000 machines of Distblief, 16000 cores process, network size the chances are 1,000,000,000 neurons, then Andrew Ng
At Stanford university 16 station servers, 64 GPU altogether, and used the switch of a superperformance
InfiniBand, trainable network size has reached 11,200,000,000 neurons.Recently, the network size of Baidu's degree of depth study is
Through having reached the node of 20,000,000,000.Estimating the foreseeable future, the scale of deep neural network is up to 100,000,000,000 neurons, rule
Mould is the biggest, and parallel architecture, optimized algorithm propose unprecedented challenge, but ultra-large after may obtain more new recognizing
Know.
The new algorithm learnt based on the above-mentioned degree of depth and the partial properties of technology have exceeded human brain.Artificial intelligence is deeply
Degree learning areas algorithm aspect obtains certain while breaking through, and commercial opportunity has just been aimed at and how to have used by some companies both at home and abroad
Chip realizes the deep neural network algorithm of these parameters optimization, it is thus achieved that artificial intelligence's behavior.Such as, recent Massachusetts science and engineering
(MIT), high pass (Qualcomm), Intel (Intel), Nvidia, Movidius etc. are at research and development degree of depth study chip.
The Massachusetts Institute of Technology (MIT) is a few days ago at International Solid circuit conference (International Solid State
Circuits Conference, ISSCC) deliver a degree of depth study chip Eyeriss, it is used for realizing artificial neural network.
MIT declares that the usefulness of this chip is typically commonly use GPU 10 times, it is possible to directly performs intelligent algorithm on equipment, is not required to
Data is processed by network.MIT represents, promotes it is critical only that of Eyeriss usefulness and minimizes GPU core (Core) and storage
The frequency of swap date between device, and general GPU core is to share single memorizer, but each core of Eyeriss has oneself
Private memory.Additionally, each core directly can be linked up with neighbouring core, if need to share data, just need not be saturating
Cross Primary memory to transmit, when there is a lot of node at convolutional Neural networking when processing identical data, neighbouring internuclear can be straight
Connect communication critically important.And promote chip usefulness also have one it is crucial that across core distribution task special circuit, can be in difference
The class neural network of type reconfigures, or automatically configures data across core.The key point that these performances improve meets poplar then
The basic principle that doctor Lin proposes.
Along with the Internet and the development of radio network technique, ecommerce has obtained the growth attracted people's attention in China, and
E-Payment is the key link of ecommerce, is the basis that shaped up of ecommerce.There is no real-time E-Payment
Means match, and transaction just cannot realize.E-Payment be transaction client, including consumer, businessman and financial institution it
Between, using safe electronic means, it is existing that the currency carried out by network or fund flow, i.e. user obtain electronics by payment terminal
The payment information of gold, the credit card, debit card, fiscard etc., is sent to bank or corresponding processing mechanism by network security
Realize paying by mails.
The basic of ecommerce is constituted as it is shown in figure 1, wherein payment system is generally made up of following functions module: pay eventually
End, client modules, the paying server being connected with business application system, the payment gateway etc. that is connected with bank private network.
Client modules is in payment process, the respective mode that paying server triggers according to the different choice of user
Block.When the business datum of user's access critical, or when user submits payment information to, server end will activate the safety of user side
Proxy module, obtains the information such as the certificate of user, private key, account, order status, builds between user and paying server
Vertical safety chain, ensures the confidentiality and integrity of data point point to-point communication.
Paying server is the payment module being associated with business application system, and it is initiated and controls a payment flow
Running.Controlling application interface by the payment incoherent with concrete business department of one group of standard, paying server realizes multiple
The support of business, major function includes: certification user, the identity of payment gateway;Offer order management services;User is provided to access
Channel interface, to realize user monitoring pay status;Process various payment message;Encryption and decryption transaction data;Management is attached thereto
The payment information of business service entity;Multiple payment gateways are prepared in management;There is provided standard interface for upper-layer service application, be used for
Associated services and payment process.
Payment gateway is the payment module connecting bank's end, is responsible for the connection connecting the internal private network of bank with paying public network,
Major function includes: provide the mutual conversion between bank and payment data form;Management certificate, private key;Certification user and industry
The identity of business service unit;Encryption and decryption payment data;Judge the integrity of the payment data that paying server submits to;User is provided
Access path interface, to realize user monitoring pay status;The configuration being connected with multiple different bank in-house networks is provided.
Payment terminal is user when consumption, is connected to the various harvesters of POS, completes the different bank credit card, borrows
The information gathering of note card, fiscard, Quick Response Code, specific identifier etc. and certain information identification, give the payment of rear end or far-end
Server is further processed.
For the POS payment flow of conventional credit card, debit card, fiscard etc., everybody has been accustomed to.Two-dimensional strip
Code/Quick Response Code (Dimensional Barcode) be according to certain rules with specific geometric figure at the figure of chequered with black and white distribution
Shape plane (two-dimensional directional) identifying recording layer symbolic information, is most commonly that QR code (Quick Response, QR) at present.QR code
Being surrounded by clear area, be divided into functional graphic and coding region form two parts, the former includes position sensing figure and separator, location
Figure and correction graph;The latter includes format information, version information and data error correction information.By image input device or photoelectricity
Scanning device automatically identifying and reading, automatically processes realizing information.Quick Response Code generates the coding that process is exactly bar code, is turned by readable information
Change code word into, add the additional information of necessity, then draw bar code.In present ecommerce, two can be generated with mobile phone
Dimension bar code or QR code, then by scanner automatically identifying and reading, complete to pay.
It is the most convenient that the E-Payment activity of bank card and QR code etc. brings to the daily financial consumption of the people, but with
Time bring inconvenience, such as bank card is too much, inconvenient to carry;Leave behind bank card or mobile phone, affect normal swipe or
Sweep QR code;Bank card is lost, and not only mending card wastes time and energy, and there is great potential safety hazard.It addition, bank card or mobile phone
Password is easily stolen or is cracked by assault, such as when user is at ATM terminal input password, others only it is noted that
Observe its action and just can guess password.
Owing to there is these problems, in order to better assure that safety and the convenience of system, sight is transferred to by people
The identification of biological characteristic or specific identifier is technical, because human body some biological characteristic or specific identifier is different and very
Difficulty is lost and imitated.Currently used biological identification technology mainly has fingerprint, iris, retina, voice, face and DNA, and
The identification technology of specific identifier.Specific identifier is that some have article or ornaments, the pattern being clearly distinguishable from other features
And biological characteristic, such as sign, tatoo, necktie, ring or necklace etc., and face, voice vocal print etc., or further,
The multiple identifier combination such as image, video, word and sound to together with, constitute a multi-modal specific identifier.
Biological with in specific identifier identification technology at these, fingerprint recognition has been obtained for actual application, particularly in intelligence
In energy mobile phone, many already provided with Fingerprint Identification Unit.Fingerprint can be regarded as a kind of specific identifier, for paying the mirror of identity
Power.Fingerprint pays and utilizes finger print information specific to everyone exactly, completes the system paid by mails.Fingerprint payment terminal can be adopted
The fingerprint of collection user, characteristic information extraction after processing, then these characteristic informations and spending amount are packaged into regulation
Protocol format is sent to paying server by WIFI or network interface card, to complete payment process.The collection of fingerprint is to hardware requirement not
Height, is easier to realize, and the fingerprint image of fast and reliable processes, recognizer also develops rapidly, it is already possible to be used for paying end
End.Fingerprint collecting equipment is placed on chain-supermarket by Paybytouch company of the U.S., and consumer can carry out identity by fingerprint
Certification and payment, reduce the waiting time of client, reduces the transaction cost of businessman, and this sample loading mode can equally be well applied to ecommerce
Pay.
Although fingerprint recognition pays and applied, there is also great potential safety hazard, due to everyone fingerprint only
Without two and be difficult to change, easily stay fingerprint to be stolen in some occasion, and some places have peddled fingerprint film
, thus can imitate the fingerprint stealing other people.
In order to improve the safety of system, need to use revocable specific identifier more, such as signature, ring, stricture of vagina
Body, jewelry, pattern, the image etc. of article, and the biological characteristic such as face, vocal print, or image, video, word and sound etc.
Be grouped together, constitute multi-modal feature, more difficult by others' conjecture to rather than such as fingerprint, iris homalographic are less
, fixing biological characteristic.But use sign, tatoo, the specific identifier such as jewelry, specific pattern, face, vocal print, need in a large number
Calculate feature extraction and identification, from the point of view of this is to general computer or traditional payment terminal, be extremely difficult.If
The image of camera collection is directly transmitted to rear end and do recognition processing, and transmission line can be proposed the highest requirement, even if having employed
Data compression, with existing compression algorithm, data volume is the biggest.This is the most unfavorable to marketing.
Allow computer identification and understand image, be one of most important target of artificial intelligence.Especially Internet era,
Certain special article or pattern that people are selected by photographic head catch, if it is possible to allow mobile phone, portable equipment, payment terminal
Identify and understand image Deng well, ecommerce will be produced huge application scenarios.
But see from above, be no matter the extensive deep neural network of Google, Baidu etc., or the unit such as MIT
Degree of depth study chip, due to problems such as scale, power consumption, volume, cost or computational efficiencies, present stage there is also for payment terminal
Certain difficulty.
Summary of the invention
The power consumption existed for existing deep neural network chip is high, chip area big and it is slow etc. to calculate speed, inapplicable
In the problem of payment terminal, the utility model proposes a kind of cell/convolutional neural networks intelligent vision and pay accelerator, with honeybee
Nest/convolutional neural networks (Cellular/Convolutional Neural Networks, CNN) is core, by photographic head (figure
As obtaining), internal memory, Micro-processor MCV and communication interface modules etc. constitute the neutral net of an image recognition degree of depth study
Accelerator, improves integrated level, reduces power consumption, it is provided that image recognition rate and precision, it is possible to complete image recognition in real time,
And flexible configuration, can complete different image identification functions by programmable configuration, enhance E-Payment safety and
Convenience.
The cell that the utility model proposes/convolutional neural networks intelligent vision pays accelerator, is formed by with lower module:
Photographic head, for obtaining image or the video of specific identifier, gives machine vision the image obtained or video and knows
Other chip;
Machine Vision Recognition chip, by picture signal or the sound of mike input of preset function treatment photographic head input
Tone signal, obtains corresponding eigenvalue, gives microprocessor;
Microprocessor, arranges and detects the duty of Machine Vision Recognition chip, receives Machine Vision Recognition chip and send
The image feature data come, operation image identification and identity validation algorithm, control the display of display screen;
Memory modules, preserves initial data, results of intermediate calculations and the final characteristics of image numerical value of input;
Display screen, in order to show duty and the system configuration information of Machine Vision Recognition chip or system, Yi Jizhan
Show picture concerned or video data;
Communication interface modules, completes the inside and outside order of accelerator, data exchange so that chip is by the mode of configuration
Work, and input and output result of calculation;
Mike, the acoustical signal of pickup user, give Machine Vision Recognition chip through microprocessor;
Speaker, is used for playing voice messaging;
Power module, provides stable power supply for Machine Vision Recognition chip and system;
It is characterized in that described Machine Vision Recognition chip connects photographic head and microprocessor, microprocessor connects internal memory
Module, display screen, communication interface modules, mike, speaker and power supply, constitute payment based on intelligent vision identification and accelerate
Device.
Described cell/convolutional neural networks intelligent vision pays accelerator, it is characterised in that described machine vision is known
Other chip is cell/convolutional neural networks chip.
Described cell/convolutional neural networks intelligent vision pays accelerator, it is characterised in that described cell/convolution
Neutral net pays accelerator can be built in panel computer, notebook computer, and both two-in-one equipment.
As can be seen from the above scheme, this utility model around cell/convolutional neural networks CNN, and photographic head,
As can be seen from the above scheme, this utility model around cell/convolutional neural networks CNN, and photographic head, internal memory,
Microprocessor and communication interface etc., constitute an intelligent vision with pattern recognition function and pay accelerator, improve image and know
Other speed and precision, overcome existing extensive deep neural network or the deficiency of degree of depth study chip, have integrated level height, merit
Consume the features such as little, calculating fast, the flexible configuration of speed.Intelligent vision pays accelerator and makes payment terminal as human visual system one
Sample, has the ability in " understanding " world so that pays without password, not to regularly replace, not remember burden;Without carrying
Bank's card or mobile phone, it is possible to pay at any time, it is to avoid the loss of bank card or mobile phone, stolen, be replicated;Can select
Different specific identifiers is as multiple-enciphered or logins multiple account;Transmission over networks is the characteristics of image of certain specific identifier
Data, have good randomness, and these are all user and system provides more preferable safety and convenience.This utility model
Applied widely, market potential is huge.
Accompanying drawing explanation
Fig. 1 is that e-commerce system constitutes block diagram.
Fig. 2 is that the CNN intelligent vision that the utility model proposes pays accelerator theory diagram.
Fig. 3 is 4 × 4 bidimensional honeycomb neutral net schematic diagrams.
Fig. 4 is the citing of individual cells equivalent circuit.
Detailed description of the invention
Below in conjunction with accompanying drawing, specific embodiment of the utility model is described in detail.
This utility model with advanced cell/convolutional neural networks (CNN) chip as core, and photographic head (image obtains
Take), microprocessor, the intelligent vision of one degree of depth study of structure such as internal memory and communication interface pay accelerator, improve and pay eventually
The recognition speed of end and precision, accelerator theory diagram is as shown in Figure 2.
Cell/convolutional neural networks (CNN) chip: for image characteristics extraction, extract image that photographic head sends here or
The image principal character paid close attention in video.The core of CNN chip is cell/convolutional neural networks (CNN), CNN network
Theory diagram is as it is shown on figure 3, CNN network can build burning the hotest degree of depth learning system.Such as the neutral net of the mankind,
CNN is made up of a large amount of non-linear analog circuit, it is possible to process the signal of input in real time, the most now these non-linear simulation electricity
The function on road can also use digital circuit to realize.The unit that these non-linear analog circuit are constituted is referred to as cell
(Cell), reaching millions of cells by certain regularly arranged, the most closest cell is just connected the most mutually, and exchange is believed
Breath.The cell of far-end is wielded influence indirectly by coupling.Each cell is by linear capacitance, linear resistance, nonlinear voltage-controlled
Current source, independent voltage source and independent current source etc. form, as shown in Figure 4, it is also possible to digital circuit and Fig. 4 etc.
The function of effect.CNN make use of the advantage of analog-and digital-two worlds, and its characteristic continuous time can process signal in real time,
Local interconnection characteristic makes it be easy to large scale integrated circuit realization, and CNN is particularly well-suited to signal parallel processing.
What Fig. 3 was given is one layer of CNN network structure of two dimension, can construct the CNN of multilamellar further, increases the deep of study
Degree, such as present degree of depth learning network framework.The parameter of CNN cell can be by arranging in advance, and used below
Journey is once again set up by program.
Different CNN cells, different CNN layers can complete different image processing functions, and such as different cells are respectively
Complete image noise reduction, image texture, rim detection, image segmentation, the detection of convex-concave angle, Boundary Extraction, holes filling, skeleton carry
Take, cutting etc., thus obtain the various features of image in real time, it is simple to follow-up micro-process realizes image recognition, expression further
And description.Certainly, image identification, represent and describe and equally realized by different configuration of CNN.
CNN described herein is mainly from extracting characteristics of image, and certain CNN can also be used for the deep of voice or sound
Degree study, extracts voice signal or the feature extraction of acoustical signal, such as vocal print feature, is subsequently used for the knowledge of voice or sound
Not, principle is similar with image.
CNN obtain characteristics of image signal can be directly fed to Micro-processor MCV, both connect interface can be serial or
Parallel data grabbing card.
About the more detailed principle of CNN, refer to the apllied patent of invention of doctor Yang Lin that above provides and deliver
Scientific paper.Based on CNN principle, complete integrated circuit one intelligent vision chip (Smart Vision Integrated
Circuit, SViC) design and flow.
Photographic head: paid close attention to or the image of specific identifier to be processed or video for shooting us, the image of acquisition
Or video gives CNN chip.On market, photographic head is the most, typically can meet the requirement of the present embodiment, in the present embodiment
Have employed Sony IMX135, use back-illuminated type imaging sensor, resolution is 4224 × 3176.
Micro-processor MCV: micro-process is connected to CNN chip, and CNN chip carries out communication, arranges and checks CNN chip
Duty.Arranging the initial operation mode of CNN chip, at system run duration, it can read the work shape of CNN chip
State, and reset the configuration parameter of each cell in CNN, it is thus achieved that different characteristics of image.According to selected microprocessor
Ability, allows the characteristics of image that microprocessor obtains according to CNN, participation parts of images identification work, and microprocessor and CNN chip are joined
Close and accelerated different image identification functions, improve recognition speed and recognition performance, play the work that image recognition is accelerated together
With.Rely on the picture material identified, user authentication and discriminating can be completed;Or microprocessor is correlated characteristic and preliminary knowledge
Other result uploads to high in the clouds by wireless network, high in the clouds return result after processing, then pay.
In the present embodiment, the processor of ARM kernel has been selected in micro-process, have low in energy consumption, the speed of service is fast, kind is many
Etc. characteristic, current occuping market leading position, high pass, Samsung, Lian Fake, Huawei, Quan Zhi, auspicious core micro-, brilliant morning etc. numerous companies
ARM microprocessor can be provided.The S5PV210 having selected Samsung in the present embodiment adopts as microprocessor, S5PV210
With ARM CortexTM-A8 kernel, ARM V7 instruction set, dominant frequency up to 1GHZ, 64/32 inside bus structure, 32/32KB
Data/commands level cache, the L2 cache of 512KB, it is possible to achieve 2000DMIPS's (200,000,000 instruction set of computing per second)
High performance computation ability.S5PV210 has a powerful hardware compression function, built-in MFC (Multi Format Codec) and
PowerVR SGX5403D graphics engine and 2D graphics engine, it would be preferable to support the PC ranks such as DX9, SM3.0, OpenGL2.0 show
Technology.Possess IVA3 hardware accelerator, built-in HDMIv1.3, HD video can be exported on external display.
The storage control of S5PV210 supports that the RAM, Flash of LPDDR1, LPDDR2 and DDR2 type supports Nandflash,
Norflash, OneNand etc..
Internal memory: for preserve the initial data of input that CNN chip and microprocessor etc. relate to, results of intermediate calculations and
Final characteristics of image numerical value etc..CNN chip self also has memorizer, it is also possible to part relates to during preserving characteristics of image
Data.In the present embodiment, internal memory has selected Micron Technology MT47H64M16HR, is DDR2 SDRAM, capacity 1Gbit.Microprocessor
Flash extension have employed Samsung K9F1G08U0E, belongs to NAND Flash, and capacity is 1Gb.
Communication interface modules: complete accelerator and outside order, data exchange, support serial data interface, including
USB, I2C etc., parallel data grabbing card, such as Samsung S5PV210 etc., general arm processor inherently has USB, I2C, parallel
The interface such as data-interface, OTG.Also support the Ethernet of RJ45 interface, and wireless network WiFi interface.Ether in the present embodiment
Net controller has selected the W5500 of WIZnet company, supports the communication by wired network Yu paying server.WiFi module is selected
Peace letter can ESP8266.
Touching liquid-crystal display screen LCD: microprocessor is also associated with touching liquid-crystal display screen, in order to show CNN chip or to be
The information such as the duty of system and system configuration, and the input of user is accepted by touch screen.The touch LCD of the present embodiment
Display screen have employed naughty crystalline substance and speeds TJC4024T032_011R screen, and touch manner is resistance-type.
Mike: for the acoustical signal of user is converted to the signal of telecommunication, gives CNN chip by Micro-processor MCV, enter
Row sound characteristic extracts, and then completes voice recognition, it is achieved user authentication and payment based on sound, it is also possible to and image, regard
Feel etc. is grouped together and constitutes multi-modal payment.The present embodiment does not has particular/special requirement, the mike on open market to mike
All can meet requirement.
Speaker: be used for playing voice messaging, such as prompt tone.The present embodiment does not has particular/special requirement to speaker, typically
Speaker on market all can meet requirement.
Power supply: the stable power-supplying needed for providing to CNN chip and system.
Above-mentioned memory chip, communication interface chip or module, mike, speaker and touching liquid-crystal display screen all pass through
Suitably pin is connected on micro-chip processor.
This embodiment is characterized in that described cell/convolutional neural networks (CNN) chip and photographic head, internal memory, Wei Chu
Feature extraction and identification accelerator in the pie graph pictures such as reason device, communication interface modules and power supply or video payment.
The present embodiment have employed Android operation system, involved software, just repeats no more at this.
This utility model is that core constitutes intelligent vision payment terminal with cell/convolutional neural networks (CNN) chip, can
For market, supermarket etc., and Online Store, there is prominent advantage: without using password, password is also without the most more
Change, be not afraid of and have forgotten, do not remember burden;Without carrying bank card or mobile phone, it is possible to pay at any time, bring greatly
Convenient, it is to avoid the unsafe act such as bank card loss, stolen, duplication;Multiple specific identifier can be utilized to constitute multiple-enciphered, carry
High security of system, or login multiple account;The length of characteristics of image is that ordinary password is incomparable, and Cipher Strength can be non-
Chang Gao;Transmission over networks is the image feature data of certain specific identifier, has good randomness, even if having obtained these
Data, decode difficulty the biggest.These are all user and system provides more preferable safety and convenience.
Being in above-described embodiment CNN built-in chip type in payment terminal, certainly, described CNN accelerator can also be interior
It is placed in the smart mobile phone in addition to payment terminal and other portable equipment, such as, described CNN accelerator is built in flat board electricity
Brain, as IPAD, and notebook computer, and both two-in-one notebook computers etc..
Similarly, CNN chip can make the computer card with respective standard interface, such as the most extensive degree of depth
GPU board commonly used in neutral net, inserts personal desktop computer (PC), server, work station and big-and-middle-sized
The inside such as computer, builds extensive deep neural network system based on CNN chip, and the degree of depth study completing big data pays
System.
Cell/convolutional neural networks (the Cellular/Convolutional Neural related in above-described embodiment
Networks, CNN) chip, in actual implementation is executed, it is also possible to replace with other Machine Vision Recognition chip, such as based on
The Machine Vision Recognition chip of either shallow study (Shallow Learning), based on autocoder (AutoEncoder), sparse
Coding (Sparse Coding), restriction Boltzmann machine (Restricted Boltzmann Machine, RBM), deep Belief Network
The Machine Vision Recognition chip of degree of depth study (Deep Learning) of network (Deep Belief Networks) scheduling algorithm.
This utility model is illustrated by above-mentioned detailed description of the invention with preferred embodiment, but this is only to facilitate manage
The example of one visualization of Xie Erju, is not considered as the restriction to this utility model scope.Equally, new according to this practicality
The technical scheme of type and the description of preferred embodiment thereof, can make various possible equivalent and change or replace, and all these
Change or replace and all should belong to this utility model scope of the claims.
Claims (2)
1. cell/convolutional neural networks intelligent vision pays accelerator, is formed by with lower module:
Photographic head, for obtaining image or the video of specific identifier, gives Machine Vision Recognition core the image obtained or video
Sheet;
Machine Vision Recognition chip, by picture signal or the sound of mike input of preset function treatment photographic head input
Signal, obtains corresponding eigenvalue, gives microprocessor;
Microprocessor, arranges and detects the duty of Machine Vision Recognition chip, receives what Machine Vision Recognition chip was sent here
Image feature data, operation image identification and identity validation algorithm, control the display of display screen;
Memory modules, preserves initial data, results of intermediate calculations and the final characteristics of image numerical value of input;
Display screen, in order to show duty and the system configuration information of Machine Vision Recognition chip or system, and shows phase
Close picture or video data;
Communication interface modules, completes the exchange of order, data;
Mike, the acoustical signal of pickup user, give Machine Vision Recognition chip through microprocessor;
Speaker, is used for playing voice messaging;
Power module, provides stable power supply for Machine Vision Recognition chip and system;
It is characterized in that described Machine Vision Recognition chip connects photographic head and microprocessor, microprocessor connects internal memory mould
Block, display screen, communication interface, mike, speaker and power supply, constitute payment accelerator based on intelligent vision identification.
Cell the most according to claim 1/convolutional neural networks intelligent vision pays accelerator, it is characterised in that described
Machine Vision Recognition chip be cell/convolutional neural networks chip.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201620177655.0U CN205665729U (en) | 2016-03-09 | 2016-03-09 | Cell convolutional neural network intelligent vision pays accelerator |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201620177655.0U CN205665729U (en) | 2016-03-09 | 2016-03-09 | Cell convolutional neural network intelligent vision pays accelerator |
Publications (1)
Publication Number | Publication Date |
---|---|
CN205665729U true CN205665729U (en) | 2016-10-26 |
Family
ID=57162455
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201620177655.0U Expired - Fee Related CN205665729U (en) | 2016-03-09 | 2016-03-09 | Cell convolutional neural network intelligent vision pays accelerator |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN205665729U (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108960017A (en) * | 2017-05-26 | 2018-12-07 | 广州智慧城市发展研究院 | A kind of RF type fingerprint recognition sensing chip framework |
CN109922361A (en) * | 2019-03-29 | 2019-06-21 | 深圳海青联合技术有限公司 | A kind of artificial intelligence box |
CN110045960A (en) * | 2018-01-16 | 2019-07-23 | 腾讯科技(深圳)有限公司 | Instruction set processing method, device and storage medium based on chip |
-
2016
- 2016-03-09 CN CN201620177655.0U patent/CN205665729U/en not_active Expired - Fee Related
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108960017A (en) * | 2017-05-26 | 2018-12-07 | 广州智慧城市发展研究院 | A kind of RF type fingerprint recognition sensing chip framework |
CN110045960A (en) * | 2018-01-16 | 2019-07-23 | 腾讯科技(深圳)有限公司 | Instruction set processing method, device and storage medium based on chip |
CN109922361A (en) * | 2019-03-29 | 2019-06-21 | 深圳海青联合技术有限公司 | A kind of artificial intelligence box |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Tong et al. | An efficient deep model for day-ahead electricity load forecasting with stacked denoising auto-encoders | |
JP7322044B2 (en) | Highly Efficient Convolutional Networks for Recommender Systems | |
CN205680247U (en) | Cell/convolutional neural networks intelligent vision driving fatigue monitoring accelerator | |
CN109522945B (en) | Group emotion recognition method and device, intelligent device and storage medium | |
CN109934249A (en) | Data processing method, device, medium and calculating equipment | |
Blanco‐Gonzalo et al. | Performance evaluation of handwritten signature recognition in mobile environments | |
CN107995428A (en) | Image processing method, device and storage medium and mobile terminal | |
CN109034206A (en) | Image classification recognition methods, device, electronic equipment and computer-readable medium | |
CN109754068A (en) | Transfer learning method and terminal device based on deep learning pre-training model | |
CN105740808B (en) | Face identification method and device | |
CN205665729U (en) | Cell convolutional neural network intelligent vision pays accelerator | |
CN113327191B (en) | Face image synthesis method and device | |
CN110070484A (en) | Image procossing, image beautification method, device and storage medium | |
CN109670927A (en) | The method of adjustment and its device of credit line, equipment, storage medium | |
CN110414550A (en) | Training method, device, system and the computer-readable medium of human face recognition model | |
CN113254491A (en) | Information recommendation method and device, computer equipment and storage medium | |
CN115512005A (en) | Data processing method and device | |
CN112418059A (en) | Emotion recognition method and device, computer equipment and storage medium | |
Praseetha et al. | Secure fingerprint authentication using deep learning and minutiae verification | |
CN116524125A (en) | Meta universe aggregation method and platform | |
Capece et al. | Implementation of a coin recognition system for mobile devices with deep learning | |
US20220254188A1 (en) | Methods for Creating Personalized Items Using Images Associated with a Subject and Related Systems and Computers | |
CN108737491A (en) | Information-pushing method and device and storage medium, electronic device | |
CN110097246A (en) | Intelligent reviewing method and system, computer installation, computer readable storage medium | |
CN108875496A (en) | The generation of pedestrian's portrait and the pedestrian based on portrait identify |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C14 | Grant of patent or utility model | ||
GR01 | Patent grant | ||
CF01 | Termination of patent right due to non-payment of annual fee | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20161026 Termination date: 20170309 |