CN205621017U - Intelligent vision chip - Google Patents
Intelligent vision chip Download PDFInfo
- Publication number
- CN205621017U CN205621017U CN201620130429.7U CN201620130429U CN205621017U CN 205621017 U CN205621017 U CN 205621017U CN 201620130429 U CN201620130429 U CN 201620130429U CN 205621017 U CN205621017 U CN 205621017U
- Authority
- CN
- China
- Prior art keywords
- image
- module
- vision
- honeycomb
- chip
- 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
Abstract
The utility model provides an intelligent vision chip system architecture belongs to the artificial intelligence technique field. The utility model provides a chip uses honeycomb neural network as the core, including camera, memory, image processing module and image feature extraction module etc. This chip has integrated level height, low power dissipation, fast, characteristics such as the configuration is nimble, and application scope is wide, has wide market prospect.
Description
Technical field
This utility model belongs to field of artificial intelligence, specifically, relates to intelligent vision chip (Smart Vision
Integrated Circuit, SVIC).
Background technology
Till the present, 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.
Breakthrough in degree of depth learning areas has caused artificial intelligence's revolution.In recent years, including Facebook, Google, University of Science and Technology
News fly, Baidu is proposed voice and the identification of image, conjunction in the interior well-known Internet enterprises of many families, utilization " degree of depth study " technology
Become algorithm.These algorithms are the algorithms of a kind of computer simulation human brain neural network.In simple terms, it is simply that build one with computer
Individual neutral net, is then constantly trained by conventional data and optimizes it.The partial function of these new algorithms and technology exceedes
Human brain.
Artificial intelligence's revolution obtains certain while breaking through in terms of degree of depth learning areas algorithm, some companies both at home and abroad
Just commercial opportunity is aimed at and how to have realized these algorithms with hardware.
In terms of intelligent vision chip design, the product architecture scheme with Movidius company of the U.S. as representative at present
For: carried out filming image by independent photographic head, by photoelectric conversion sensor, image is passed to an image procossing and figure
Numerical calculation and the system of identification as feature extraction.
The feature of Movidius architectural schemes is as follows:
1) photographic head and image procossing and characteristic extraction part, be made up of two pieces of integrated circuits respectively;
2) image procossing and characteristic extraction part, it is common that hundreds of GPU cores and about 10 CPU core are designed at a piece
In IC chip, to complete the calculating of substantial amounts of data.
The shortcoming of Movidius architectural schemes mainly has:
Owing to image procossing and characteristic extraction part need substantial amounts of calculating so that this IC chip power consumption is very big,
Generally this IC chip power consumption is more than 500 milliwatts, in miniaturized electronic devices, such as mobile phone, IPAD, takes when using in notebook
Electricity, heating is serious;
This block is responsible for the IC chip of image procossing and feature extraction at present, embeds hundreds of GPU cores and 10 left sides
Right CPU core, the integrated circuit chip area designed is relatively big, and the selection for semiconductor manufactures requires the highest, raw
Produce cost the highest;
Calculate speed the most on the low side.
To this end, the utility model proposes a kind of brand-new intelligent vision chip.
Summary of the invention
The power consumption existed for existing intelligent vision chip is high, chip area big and calculates the problems such as speed is slow, this practicality
Novel propose a kind of new intelligent chip system architecture, with vision honeycomb neutral net (Cellular Neural
Networks, CNN) it is core, the meter required by photographic head (Image Acquisition), memorizer, image procossing, image characteristics extraction etc.
Calculate and identification is integrated in one piece of integrated circuit, improve integrated level, reduce power consumption, it is possible to complete image recognition in real time, and
And flexible configuration, different image identification functions can be completed by programmable configuration.
The utility model proposes a kind of intelligent vision chip, including Built-in Image acquisition module, D/A converter module, figure
As processing module, system clock module, memorizer, communication control module and vision honeycomb neural network module.Wherein, described
The input of system clock module is connected to external clock reference, when providing internal or external work for described intelligent vision chip
Clock;The output of described Built-in Image acquisition module is connected to described vision honeycomb neural network module;Described D/A converter module
Input be connected to external digital image source, output be connected to described image processing module;The input of described image processing module
Be also connected to external analog image source, output is connected to described vision honeycomb neural network module;Described vision honeycomb nerve net
Network module processes the value obtaining characteristics of image to its picture signal received;Described memorizer is neural with described vision honeycomb
Mixed-media network modules mixed-media connects;Described communication control module is bi-directionally connected with described vision honeycomb neural network module, exports described image
The value of feature;The communication interface of described communication control module includes serial line interface, parallel data grabbing card, Ethernet interface and wireless
Interface.Preferably, the parameter of described vision honeycomb neural network module is to arrange in advance, and can in use pass through journey
Sequence is once again set up.
Preferably, described memorizer is the memorizer of simulation, or the memorizer of numeral.
Preferably, described wave point includes WiFi and bluetooth.
Preferably, the value of described characteristics of image is exported by WiFi, parallel data grabbing card or Ethernet interface.
Preferably, described characteristics of image includes image texture, edge, convex-concave angle, border, hole, skeleton and cutting.
As can be seen from the above scheme, this utility model, around vision honeycomb neutral net CNN, by imageing sensor, is deposited
Reservoir, image procossing and image recognition etc. constitute one chip, overcome the deficiency of existing intelligent vision chip, have integrated
Degree is high, power consumption is little, calculate the features such as fast, the flexible configuration of speed.Chip is applied widely, and market potential is huge.
Accompanying drawing explanation
Fig. 1 is the intelligent vision chip block diagram that the utility model proposes.
Fig. 2 is 4 × 4 bidimensional honeycomb neutral net schematic diagrams.
Fig. 3 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 vision honeycomb neutral net (CNN), as core, by Image Acquisition, digital to analog conversion, is
Calculating and identification needed for system clock, memorizer, image procossing and image characteristics extraction are integrated in one piece of integrated circuit, such as figure
Shown in 1.
Core of the present utility model is vision honeycomb neutral net (CNN), as in figure 2 it is shown, CNN can build present normalizing
Degree of depth study (Deep Learning) system of heat.Such as neutral net, CNN is made up of a large amount of non-linear analog circuit, it is possible to
Processing the signal of input in real time, the function of these non-linear analog circuit can also use digital circuit to realize the most now.
The unit that these non-linear analog circuit are constituted is referred to as cell (Cell), reaches the cell of millions of by certain rule row
Row, the most closest cell is just connected the most mutually, exchanges information.The cell of far-end is wielded influence indirectly by coupling.
Each cell is by groups such as linear capacitance, linear resistance, nonlinear voltage-controlled current source, independent voltage source and independent current sources
Become, as shown in Figure 3, it is also possible to by the function of digital circuit with Fig. 3 equivalence.CNN make use of analog-and digital-two worlds
Advantage, its characteristic continuous time can process signal in real time, and local interconnection characteristic makes it be easy to large scale integrated circuit
Realizing, CNN is particularly well-suited to signal parallel processing.
What Fig. 2 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 cell, different CNN layers can complete different image processing functions, such as
Different cells are respectively completed image noise reduction, image texture, rim detection, the detection of convex-concave angle, Boundary Extraction, holes filling, skeleton
Extraction, cutting etc., thus obtain the various features of image in real time, it is simple to follow-up image recognition, represent and describe.Certainly,
The identification of image, represent and describe and equally realized by different configuration of CNN.
The characteristics of image signal that CNN obtains can directly export from each cell, or exports after each cell multiplexing again.
About the more detailed principle of CNN, refer to the apllied following patent of invention of doctor Yang Lin and the science opinion delivered
Literary composition.
Leon O, Chua;Lin Yang, " Cellular Neural Network ", United States Patent,
Patent Number:5,140,670, Date of Patent:Aug.18,1992.
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.
Built-in Image acquisition module, it is made up of substantial amounts of optical sensor, and each CNN cell has a single light sensing
Device is connected, and a part for each light sensors image obtains the input data of corresponding CNN cell.CNN cell can be parallel
Receiving at a high speed input signal, each CNN cell is exclusively used in the single pixel processing input picture, in order to obtain image recognition in real time
Result.By the transparent window on chip package, image to be processed or data can be projected directly on CNN core.
Exterior view image source sum modular transformation module, intelligent vision chip SVIC is except above-mentioned built-in image collection module
Outward, the input of exterior view image source is also supported.Exterior view image source is probably video camera, mobile phone, computer or video player etc., thing
Reason interface support BNC, separate video signal interface YUV/RGB, S-Video terminal, composite video signal CVBS, RCA, USB,
HDMI etc..That the picture signal that exterior view image source produces is probably simulation or numeral, if simulation, then it is directly fed to CNN
Unit, if but numeral, then give digital to analog conversion module, complete the data image signal conversion to analogue signal, obtain phase
The analogue signal answered, gives CNN unit after image processing module again.
Image processing module, carries out some technical finesses to the picture signal of outside input, such as carries out input signal
Amplitude amplitude limiting processing, is normalized between [-1,1], in order to meet the CNN requirement to input range;Gamma correction, to obtain more
Good DYNAMIC DISTRIBUTION, etc..
System clock module, intelligent vision chip SVIC supports onboard clock and external clock.
Memorizer, for preserving the initial data of input, results of intermediate calculations and final characteristics of image numerical value etc., storage
Device is built in inside SVIC, and this is different from existing intelligent vision chip.Memorizer uses digital memory or simulation to deposit
Reservoir, uses analog memory more can reduce power consumption.
Communication/control module, completes SVIC and outside order, data exchange, supports serial data interface, including USB,
I2C etc.;Parallel data grabbing card, the gigabit Ethernet of RJ45 interface, and radio network interface, including WiFi and bluetooth.CNN obtains
The image feature value obtained can be transmitted by these communication interfaces, such as by parallel data grabbing card, Ethernet or WiFi etc..
The intelligent vision chip solution that the utility model proposes and existing Movidius architectural schemes be relatively shown in Table 1,
Therefrom it can be seen that the architectural schemes that the utility model proposes has the advantage that includes that electronic device integration degree is high, chip merit
Consuming little, IC chip power consumption is less than 50 milliwatts, has effect of significant low-power consumption, high performance-price ratio.The chip scope of application
Extensively, also being adapted in small intelligent equipment using, market potential is huge;Use vision honeycomb nerual network technique, in real time from institute
Extracting image graphics characteristic in the visual pattern obtained, extracting operation time is the computing extraction time that the current world announces
About 1/10th;The design of this chip is internal millions of optic nerve cell factory, and employing can arrange (preset, network
Amendment) the nerve cell network number of plies, and the functional parameter of every confluent monolayer cells, the range of application of this chip can be changed flexibly.
Table 1 products scheme compares
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 (6)
1. intelligent vision chip, it is characterised in that include Built-in Image acquisition module, D/A converter module, image processing module,
System clock module, memorizer, communication control module and vision honeycomb neural network module, wherein:
The input of described system clock module is connected to external clock reference, provides internal or external for described intelligent vision chip
Work clock;
The output of described Built-in Image acquisition module is connected to described vision honeycomb neural network module;
The input of described D/A converter module is connected to external digital image source, output is connected to described image processing module;
The input of described image processing module is also connected to external analog image source, output is connected to described vision honeycomb nerve net
Network module;
Described vision honeycomb neural network module processes the value obtaining characteristics of image to its picture signal received;
Described memorizer is connected with described vision honeycomb neural network module;
Described communication control module is bi-directionally connected with described vision honeycomb neural network module, exports the value of described characteristics of image;
The communication interface of described communication control module includes serial line interface, parallel data grabbing card, Ethernet interface and wave point.
Intelligent vision chip the most according to claim 1, it is characterised in that described vision honeycomb neural network module
Parameter is to arrange in advance, and can be in use once again set up by program.
Intelligent vision chip the most according to claim 1, it is characterised in that described memorizer is the memorizer of simulation, or
The memorizer of person's numeral.
Intelligent vision chip the most according to claim 1, it is characterised in that described wave point includes WiFi and bluetooth.
Intelligent vision chip the most according to claim 4, it is characterised in that the value of described characteristics of image by WiFi and
Row data-interface or Ethernet interface output.
Intelligent vision chip the most according to claim 1, it is characterised in that described characteristics of image includes image texture, limit
Edge, convex-concave angle, border, hole, skeleton and cutting.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201620130429.7U CN205621017U (en) | 2016-02-22 | 2016-02-22 | Intelligent vision chip |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201620130429.7U CN205621017U (en) | 2016-02-22 | 2016-02-22 | Intelligent vision chip |
Publications (1)
Publication Number | Publication Date |
---|---|
CN205621017U true CN205621017U (en) | 2016-10-05 |
Family
ID=57033515
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201620130429.7U Expired - Fee Related CN205621017U (en) | 2016-02-22 | 2016-02-22 | Intelligent vision chip |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN205621017U (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107623827A (en) * | 2017-08-15 | 2018-01-23 | 上海集成电路研发中心有限公司 | A kind of intelligent CMOS image sensor chip and its manufacture method |
CN108712630A (en) * | 2018-04-19 | 2018-10-26 | 安凯(广州)微电子技术有限公司 | A kind of internet camera system and its implementation based on deep learning |
-
2016
- 2016-02-22 CN CN201620130429.7U patent/CN205621017U/en not_active Expired - Fee Related
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107623827A (en) * | 2017-08-15 | 2018-01-23 | 上海集成电路研发中心有限公司 | A kind of intelligent CMOS image sensor chip and its manufacture method |
CN107623827B (en) * | 2017-08-15 | 2020-06-09 | 上海集成电路研发中心有限公司 | Intelligent CMOS image sensor chip and manufacturing method thereof |
CN108712630A (en) * | 2018-04-19 | 2018-10-26 | 安凯(广州)微电子技术有限公司 | A kind of internet camera system and its implementation based on deep learning |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106203621B (en) | The processor calculated for convolutional neural networks | |
CN205621018U (en) | Cell -phone cell convolutional neural network accelerator | |
Wang et al. | Improved multi-order distributed HOSVD with its incremental computing for smart city services | |
CN102981477B (en) | Based on the refrigerator inside inspection method of multi-cam handoff technique | |
CN109740739A (en) | Neural computing device, neural computing method and Related product | |
CN104899921A (en) | Single-view video human body posture recovery method based on multi-mode self-coding model | |
CN110163355A (en) | A kind of computing device and method | |
CN103106672B (en) | A kind of interesting image regions detection method based on color characteristic | |
CN205621017U (en) | Intelligent vision chip | |
CN105094720A (en) | Cloud platform based 3D printing manufacturing system and implementing method thereof | |
CN101916429A (en) | Geometric correction and disparity extraction device of binocular camera | |
CN104657434B (en) | A kind of social network structure construction method | |
CN104657832A (en) | Whole-process monitoring and traceability tracking method for quality safety of agricultural products | |
CN108536844A (en) | A kind of network representation learning method of Text enhancement | |
CN110515587A (en) | Multiplier, data processing method, chip and electronic equipment | |
CN104319897A (en) | Intelligent micro-grid control system for achieving high-speed communication based on FPGA | |
Jabbar et al. | Ambient light energy harvesting and numerical modeling of non-linear phenomena | |
CN206556690U (en) | Temperature Humidity Sensor | |
CN117034189A (en) | Multi-source internet of things data fusion method | |
CN102916691A (en) | BCD (binary-coded decimal) decimal counter based on reversible logic | |
CN113034298A (en) | Important user identification method based on node attraction in mobile social network | |
CN205302607U (en) | Wireless sensor node data collection system under water | |
CN105426574A (en) | Method for fast switching from process and performance requirements to design parameter of injection molding equipment | |
CN103065176A (en) | Method for generating image by one-dimensional bar code and system | |
CN204480307U (en) | A kind of agricultural product quality and safety complete monitoring and tracing device |
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 |
Granted publication date: 20161005 Termination date: 20180222 |
|
CF01 | Termination of patent right due to non-payment of annual fee |