CN108764465A - A kind of processing unit carrying out neural network computing - Google Patents
A kind of processing unit carrying out neural network computing Download PDFInfo
- Publication number
- CN108764465A CN108764465A CN201810486527.8A CN201810486527A CN108764465A CN 108764465 A CN108764465 A CN 108764465A CN 201810486527 A CN201810486527 A CN 201810486527A CN 108764465 A CN108764465 A CN 108764465A
- Authority
- CN
- China
- Prior art keywords
- neural network
- navigation
- data
- input
- equipment according
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
- G06N3/063—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using electronic means
Abstract
A kind of intelligence navigation equipment, including：Target Identification Unit for handling the image shot when navigation, and is compared with sample database, identifies the obstacle identity in image；Information processing unit for the obstacle identity of receives that treated image and identification, and passes through neural network computing, and output navigation selects data.The equipment of the disclosure can be applied to intelligent navigator, can save human cost, while reduce the risk during vehicles navigation.
Description
Technical field
This disclosure relates to technical field of information processing, and in particular to a kind of intelligence navigation equipment.
Background technology
It is intensive in navigation channel especially in navigator field at present in driving or navigational field, the case where water surface complex
Under, ship generally carries out laser-marking object come avoiding navigation risk by laser radar and conventional radar.And for small drift
Float then constantly observes tiny floating material under sail mainly by navigator, to the water surface and underwater barrier
Evaded.
Existing technology one kind is sonar undersea detection technology, can refer to sonar fish deteclor, but it only has submarine target
Effect, complex water areas are affected to the accuracy rate of detection；One kind is the radar exploration technique, such as conventional radar or laser thunder, but
It is that there are close-in target detection blind areas for radar, can not finds floating material existing for the surrounding water surface in time；Also one is visions
System Detection Techniques, mainly waterborne target feature extraction, but Conventional visual Detection Techniques, which can only differentiate island steamer etc., to be had
The significantly object of identification feature, can not play a role complicated aquatic environment；Have that one is Artificial Cognitions again, but manually distinguishes
It is too high to know cost, and to the more demanding of navigator.
Invention content
(1) technical problems to be solved
In view of this, the disclosure is designed to provide a kind of intelligent navigation equipment, to solve at least partly technical problem.
(2) technical solution
To achieve the above object, the disclosure provides a kind of intelligent navigation equipment, including：Target Identification Unit, for boat
The image shot when row is handled, and is compared with sample database, identifies the obstacle identity in image；Information processing apparatus
It sets, for the obstacle identity of receives that treated image and identification, and passes through neural network computing, output navigation selects number
According to.
In further embodiment, information processing unit includes：Neural network chip, for executing the neural network fortune
It calculates.
In further embodiment, the neural network chip includes：Storage unit is stored for storing input data
Neural network parameter and computations, the obstacle identity of the input data includes that treated image and identification；Control
Unit parses the computations and obtains multiple operational orders for extracting computations from the storage unit；Arithmetic element,
For navigation selection data to be calculated to input data execution according to multiple operational order.
In further embodiment, the storage unit includes：Input/output module, for obtaining input data, nerve
Network parameter and computations；Scalar data memory module, for storing scalar data；Storage medium, for storing data
Block, and the storage medium stores on piece.
In further embodiment, the neural network chip further includes：Direct memory access DMA, for being stored in storage
Data in unit, neural network parameter and computations, so that control unit and/or arithmetic element are called.
In further embodiment, the neural network chip further includes：Instruction buffer, for being deposited from the direct memory
DMA cache instructions are taken, are called for control circuit, described instruction caching is that on piece caches.
In further embodiment, the neural network parameter includes inputting neuron, output neuron and weights, described
Processor further includes：Neuron caching is inputted, for caching input neuron from the direct memory access DMA, for operation list
Metacall；Weights cache, and for caching weights from the direct memory access DMA, are called for arithmetic element；Output neuron is slow
It deposits, the output neuron after operation is obtained for storing from the arithmetic element, with output to direct memory access DMA；Wherein
The input neuron cache module, weights cache module and output neuron cache module are on piece caching.
In further embodiment, described information processing unit further includes：Input data encoder, for by treated
Image and the obstacle identity of identification are changed into the manageable digital information of neural network.
In further embodiment, described information processing unit further includes：Memory, for storing in current slot
Voyage conditions and navigation in different time periods select data；Arithmetic unit calculates navigation for selecting queuing data according to navigation
Risk and/or current flight loss.
In further embodiment, described information processing unit further includes：Data input/output terminal：Number is inputted for receiving
According to the signal of encoder, and artificial neural network chip is passed to as input；It is additionally operable to receive neural network chip output
The navigation of end output selects data, and is stored in memory；And for reading navigation selection queue and current time from memory
Voyage conditions in section, are input to arithmetic unit.
In further embodiment, described information processing unit further includes：Output data converter, based on by arithmetic unit
The risk of the navigation of calculation and/or the lossy data of current flight are transferred to external equipment after recompiling.
In further embodiment, the result that arithmetic unit calculates is back to target identification and filled by the output data converter
It sets, and the update sample database.
In further embodiment, the neural network chip is additionally operable to make navigation selection in external equipment or crewman
Afterwards, input-output information group is collected, neural network parameter is adaptively updated, training generates new neural network parameter.
In further embodiment, the Target Identification Unit includes：Image processing unit, for carrying out input picture
Pretreatment, binaryzation and selection subregion extract feature；Comparing unit, the sample in feature and sample database for that will extract
Progress must be read, the preliminary obstacle target type identified in image.
(3) advantageous effect
(1) equipment of the disclosure can be applied to intelligent navigator, can save human cost, while reducing the vehicles
Risk during (such as ship) navigation；
(2) the intelligent navigation equipment of the disclosure realizes navigator using special neural network chip, can preferably fit
Different waters situations is answered, while solving delay issue (and the common hardware epineural network of the neural network of real time execution
The speed of service is slower, causes AI navigator's algorithm speed of service slower, causes ship that can not receive the feedback of operational configuration in time, no
Has real-time).
Description of the drawings
Fig. 1 is a kind of information processing unit block diagram of the embodiment of the present disclosure.
Fig. 2 is the block diagram of neural network chip in Fig. 1.
Fig. 3 is the block diagram of the intelligent navigation equipment of the embodiment of the present disclosure.
Fig. 4 is a kind of application scenarios schematic diagram of intelligent navigation system.
Fig. 5 is another application scenarios schematic diagram of intelligent navigation system.
Fig. 6 is a kind of operational flow diagram of intelligent method of navigation of the embodiment of the present disclosure.
Specific implementation mode
To make the purpose, technical scheme and advantage of the disclosure be more clearly understood, below in conjunction with specific embodiment, and reference
Attached drawing is described in further detail the disclosure.
Hereinafter, it will thus provide embodiment of the embodiment the disclosure is described in detail.The advantages of disclosure and effect
It will be more notable by disclosure disclosure of that.Illustrate that appended attached drawing simplified and used as illustrating herein.Attached drawing
Shown in component count, shape and size can modify according to actual conditions, and the configuration of component is likely more complexity.
Also otherwise practice or application can be carried out in the disclosure, and without departing from the condition of spirit and scope defined in the disclosure
Under, various change and adjustment can be carried out.
Fig. 1 is a kind of information processing unit block diagram of the embodiment of the present disclosure.According to a side of the embodiment of the present disclosure
Face provides a kind of information processing unit 110, for the image that receives that untreated or treated and/or the obstacle identity of identification,
And pass through neural network computing, output navigation selection data.Wherein, which can come from the vehicles shot when navigation
Video frame or image when (land or sea) is navigated by water captured by periphery；It can be the figure by the pretreated feature containing extraction
As data or raw video picture data.By carrying out neural network computing point to obstacle identity in the image and image
Analysis determines obstacle identity and then determines navigation selection, for example whether avoiding the barrier, avoids in which way.By this
Whether information processing unit can have barrier and barrier exact type in efficient analysis image, and provide navigation choosing
It selects, human cost can be saved, while reducing the risk during vehicles navigation.
The information processing unit 110 can be according to input data and in a period of time history image information and at this
Selected sail information is trained in the case of a little history images, passes through LSTM model initializations parameter and correlated inputs data
Operation result, obtain loss function, calculate the navigation selection of Least-cost in this condition, then externally return and intelligence is presented
Can navigate the optimal navigation selection judged.Simultaneously after crewman makes navigation selection, input-output information group is collected, it is adaptive
It updates the parameter (such as weights, biasing) in neural network (such as LSTM shot and long terms memory network) with answering, and then generates training
Generate new neural network parameter.
In some embodiments, information processing unit 110 includes：Neural network chip 111 is used to execute neural network
Operation.By using dedicated neural network chip, the speed of neural network computing is can speed up, plan is timely navigated by water to provide
Slightly.
In some embodiments, the navigation that information processing unit 110 provides selects data, including but not limited to：Whether have
Whether barrier the type of barrier, weather condition, should be hidden, direction of advance, forward speed and/or situation urgency level.
Fig. 2 is a kind of neural network chip block diagram of embodiment in Fig. 1.As shown in Fig. 2, in some implementations
In example, neural network chip 200 includes storage unit 201, control unit 202 and arithmetic element 203, wherein storage unit 201
For storing input data (input neuron can be used as), neural network parameter and instruction；Control unit 202 is used for from described
Special instruction is read in storage unit, and is decoded into arithmetic element and instructs and be input to arithmetic element 203；Arithmetic element
203, for executing corresponding neural network computing to the data according to arithmetic element instruction, obtain output neuron.Wherein,
Storage unit 201 can also store the output neuron obtained after arithmetic element operation.Here neural network parameter includes
But it is not limited to weights, biasing and activation primitive.
In the data that the storage unit 201 is stored, wherein input data can be external incoming untreated or processing
The obstacle identity of image and identification afterwards；Neural network can be prestore or the later stage be stored in, which includes
But it is not limited to CNN (convolutional neural networks), RNN (Recognition with Recurrent Neural Network) or DNN (deep neural network), it is preferred that use
LSTM shot and long term memory network models in RNN (Recognition with Recurrent Neural Network)；Computations can also be to be prestored into storage unit 201
In or the later stage be stored in from outside, will the computations be carried out with further concrete explanation below.
In some embodiments, storage unit 201 includes scalar data storage unit, which is used for
The scalar data in input data, neural network parameter and instruction is stored, such as temperature humidity equal amount of data, scalar data are deposited
Storage module effect is to promote the speed that data are carried；Storage unit 201 can also include input/output module, defeated for obtaining
Enter data, neural network parameter and computations (computations obtain operational order after being compiled by control unit,
It controls arithmetic element and executes corresponding neural network computing).
In some embodiments, storage unit 201 can also include storage medium.The storage medium can be piece external storage
Device, certainly in practical applications, or on-chip memory, block, the data block are specifically as follows n dimensions for storing data
Be the integer more than or equal to 1 according to, n, be 1 dimension data for example, when n=1, i.e., it is vectorial, it is 2 dimension datas, i.e. matrix when such as n=2,
It is multidimensional tensor when such as n=3 or 3 or more.
In some embodiments, corresponding neural network computing is executed in arithmetic element 203 includes：Will input neuron and
Weight data is multiplied, and obtains multiplied result；Add tree operation is executed, for the multiplied result to be passed through add tree phase step by step
Add, obtain weighted sum, weighted sum biasing is set or is not processed；The weighted sum set or be not processed to biasing executes activation primitive
Operation obtains output neuron.Preferably, activation primitive can be sigmoid functions, tanh functions, ReLU functions or
Softmax functions.
In some embodiments, as shown in Fig. 2, neural network chip 200 can also include DMA204 (Direct
Memory Access, direct memory access), the input data for being stored in storage unit, neural network parameter and instruction,
So that control unit 202 and arithmetic element 203 are called；Further it is additionally operable to calculate output neuron in arithmetic element 203
Afterwards, the output neuron is written to storage unit 201.
In some embodiments, it as shown in Fig. 2, neural network chip 200 further includes instruction buffer 205, is used for from described
Direct memory access DMA cache instructions are called for control unit.The instruction buffer 205 can be that on piece caches, and pass through preparation
Technique is integrated on neural network chip, can be improved processing speed when instruction is transferred, be saved the integral operation time.
In some embodiments, neural network chip 200 further includes：Neuron caching 206 is inputted, is used for from described straight
Memory access DMA caching input neurons are connect, are called for arithmetic element 203；Weights caching 207 is used for from described directly interior
It accesses DMA204 and caches weights, called for arithmetic element 203；Output neuron caching 208 is used to store from the operation
Unit 203 obtains the output neuron after operation, with output to direct memory access DMA204.Above-mentioned input neuron caching
206, weights caching 207 and output neuron caching 208 or on piece caching, nerve is integrated in by semiconductor technology
On network chip 200, processing speed can be improved when being read and write for arithmetic element 203, save the integral operation time.
In some embodiments, above- mentioned information processing unit 110 further includes input in addition to including neural network chip 111
Data encoder 112, for treated image and the obstacle identity of identification to be changed into the manageable number of neural network
Word information.The manageable digital information can be used as input neuron in neural network to participate in neural network computing.
In some instances, above- mentioned information processing unit 110 can also include：Memory 114 and arithmetic unit 115, wherein
Memory 114 is used to store voyage conditions and navigation in different time periods selection data in current slot；Arithmetic unit 115
For selecting queuing data, the risk for calculating navigation (to be estimated by multi-aspect informations such as sea information, Weather informations according to navigation
Meter) and/or current flight loss (including fuel oil, the time it takes etc. consumed).
In some embodiments, information processing unit 110 further includes data input/output terminal 113, the input/output terminal 113
Signal for receiving input data encoder 112, and by its afferent nerve network chip 111 as input；It is additionally operable to receive
The navigation of 111 output end of neural network chip output selects data, and is stored in memory 114；And for being read from memory 114
Navigation selection queue and the voyage conditions in the current period are taken, arithmetic unit 115 is input to.
In some embodiments, information processing unit 110 further includes output data converter 116, output data conversion
Device 116 is outer for being transferred to after recompiling the risk of navigation and/or the lossy data of current flight that arithmetic unit 115 calculates
Portion's equipment.
In some embodiments, the result that arithmetic unit calculates is back to outside by the output data converter 116, and
Update sample database；And/or obtained result is back to outside, external sample database is updated with this.
In some embodiments, the neural network chip 111 is additionally operable to make navigation selection in external equipment or crewman
Afterwards, input-output information group is collected, neural network parameter is adaptively updated, training generates new neural network parameter.It is preferred that
, above-mentioned training process is handled in real time.
Fig. 3 is the block diagram of the intelligent navigation equipment of the embodiment of the present disclosure.According to another party of the embodiment of the present disclosure
Face, it further includes target identification in addition to the information processing unit 110 including above-described embodiment to provide a kind of intelligent navigation equipment 100
Device 120.Wherein, the Target Identification Unit 120 input (includes but not limited to camera, mobile phone, number from camera device is influenced
Camera) intake video frame or picture, as input picture, after can pass through series of preprocessing process (image defogging, ash
The processing such as degreeization, steady picture, smooth, segmentation), and (these features include that texture is special to selection subregion extraction feature after binaryzation
Sign, invariant moment features, geometric properties etc.), after extracting following feature, it is compared with navigation water surface sample database, tentatively obtains
Obstacle identity, the obstacle identity using feature and tentatively judged is as further input afferent message processing unit 110.
The intelligent navigation equipment 100 of the embodiment of the present disclosure realizes navigator using special neural network chip, can preferably adapt to
Different waters situations, while solving the delay issue of the neural network of real time execution.
Intelligent navigator is realized by 110 two devices of above-mentioned Target Identification Unit 120 and information processing unit.Target
Identification device handles image, and is compared with the image in sample database, the type of cognitive disorders object, and at information
Reason device receives the image that Target Identification Unit is passed to, and externally returns to the navigation scheme of ship under current state, to real
Now intelligence is navigated.
In the interaction of the two, the storage unit 20 of information processing unit 110 can be handed over Target Identification Unit 120 in real time
Mutually, receive the incoming feature of Target Identification Unit 120 and obstacle identity information, when device is according to input data and one section
In history image information and be trained in the selected sail information of these history images, pass through nerve net
The operation result of network (such as LSTM) model initialization parameter and correlated inputs data obtains loss function, calculates in the situation
The navigation of lower Least-cost selects, and then externally returns and the optimal navigation selection intelligently navigated and judged is presented.Simultaneously in ship
After member makes navigation selection, input-output information group is collected, adaptively updates neural network (such as LSTM shot and long terms memory net
Network) in parameter (such as weights, biasing), and then generate training and generate new neural network parameter.
It is not superfluous herein for the set-up mode of inside set-up mode reference above-described embodiment of information processing unit 110
It states.
In some embodiments, for Target Identification Unit 120, in subregion characteristic extraction procedure, textural characteristics are adopted
It is extracted with gray level co-occurrence matrixes, it includes four that textural characteristics, which describe the factor,；Invariant moment features may include 6 Hu not bending moments
Feature and 3 radiation invariant moment features (9 features altogether)；Above-mentioned geometric properties may include then area features, elongated
Spend feature, tight ness rating feature and convex closure feature.
In some embodiments, Target Identification Unit 120 includes image processing unit, for carrying out image to input picture
Processing unit, for input picture to be pre-processed, binaryzation and selection subregion extract feature；Comparing unit is used for
Sample progress in the feature and sample database of extraction must be read, the preliminary obstacle target type identified in image.
In some embodiments, during above-mentioned image preprocessing, image defogging uses the single scale based on edge detection
Retinex algorithm, and the gaussian filtering estimated brightness component based on marginal information is utilized, there is preferable defog effect.Electronics
Steady picture then uses Scale invariant features transform (SIFT) algorithm to extract characteristic point, seeks mending in conjunction with radiation pattern and Kalman filter
Parameter is repaid, every frame image is compensated using consecutive frame penalty method.
In some embodiments, the source of above-mentioned sample database is real scene shooting image data, network image data, and/or 3DMAX
The various 3D models of software development.
In some embodiments, the obstacle in above-mentioned sample database includes following at least one：Ship, large ocean biology,
Buoy, reef, glacier, island, farm and fishing net etc..
In some embodiments, the navigation selection cost of above-mentioned loss function can be defined by following several points, and ship can
(quantitative measurement is come with the price of maintenance of ships) is damaged caused by energy, the influence caused by environmental ecology, hours underway are prolonged
Accidentally, the oil mass of ship consumption, more than quantitative measurement after several standards, is weighted the above standard, obtains navigating by water selection
Loss function.
The embodiment of the present disclosure also provides a kind of intelligent method of navigation, including：By Target Identification Unit to being shot when navigation
Image handled, and be compared with sample database, identify the obstacle identity in image；And pass through information processing apparatus
The obstacle identity for receiving treated image and identification is set, and passes through neural network computing, output navigation selection data.
Fig. 6 is a kind of operational flow diagram of intelligent method of navigation of the embodiment of the present disclosure.The artificial finger of the embodiment of the present disclosure
Enabling the overall process that navigation equipment works can be：
Step S1, Target Identification Unit 120 handle image, and are compared with the image in sample database, know
The type of other barrier forms input data；
Input data is stored in storage unit by step S2 after the coding of input data encoder 112 through input/output terminal 113
201 or direct incoming storage unit 201；
The finger that step S3, DMA204 (Direct Memory Access, direct memory access) store storage unit 201
It enables, inputs neuron (comprising the above-mentioned data for meeting neural network input format) and weights caching, is passed to finger in batches respectively
Caching is enabled, is inputted in neuron caching and weights caching；
Step S4, control unit read instruction from instruction buffer, and operation is passed to after being decoded as arithmetic element instruction
Unit；
Step S5, instructs according to arithmetic element, and arithmetic element executes corresponding operation：In each layer of neural network,
Operation can be divided into three steps：Corresponding input neuron is multiplied by step S5.1 with weights；Step S5.2 executes add tree fortune
It calculates, i.e., the result of step S5.1 is added step by step by add tree, obtains weighted sum, weighted sum biasing is set as needed or not
It processes；Step S5.3 executes activation primitive operation to the result that step S5.2 is obtained, obtains output neuron, and passed
Enter in output neuron caching.
Obtained result is back to Target Identification Unit 120 by step S6, output signal converter 116, updates boat with this
Row water surface sample database, while obtained digital signal being recompiled, by showing that equipment is presented to crewman.
Specific functional modules/device/unit involved in step etc. can refer to corresponding function element in above-described embodiment
It is configured.
Fig. 4 is a kind of application scenarios schematic diagram of intelligent navigation system.As shown in figure 4, the embodiment of the present disclosure provides one kind
Intelligent navigation system can be applied to marine intelligence and navigate, and a steamer can be equipped with a kind of intelligent navigation system comprising
Image camera device 410 (the image camera device can be camera or video camera), in front of shooting is navigated by water in navigation
Image；And Target Identification Unit, for handling the image shot when navigation, and it is compared with sample database,
Identify the obstacle identity in image；Information processing unit, for the obstacle identity of receive that treated image and identification, and
Pass through neural network computing, output navigation selection data；And display device, the boat for exporting described information processing unit
Row selection data show user.Setting for above-mentioned apparatus, it is right in intelligent navigation equipment to be referred in above-described embodiment
Device is answered, it will not be described here.It should be noted that for target device device, information processing unit and display device, three
Main body is can be different, can also be become one, such as computer 420 shown in Fig. 4.When image camera device 410
When taking barrier 430, it can be calculated by computer 420 shown in Fig. 4 and show user (such as crewman), to assist carrying out
Flight operations.
Similar, Fig. 5 is another application scenarios schematic diagram of intelligent navigation system.Slightly different, the intelligence in Fig. 5
Energy navigation system is applied in the navigation on land, and the image camera device 510 in Fig. 5 can be installed on shield glass
(such as installing with camera in windshield), for information processing unit, display device and target device device, three
Main body is can be different, can also be become one, such as (it can be integrated to the behaviour of automobile to host 520 shown in Fig. 4
Make in system), when image camera device 510 takes barrier 530, it can calculate and carry by host 520 shown in fig. 5
Show driver, to assist driving.
In the embodiment that the disclosure is provided, it should be noted that, disclosed relevant apparatus and method can pass through others
Mode is realized.For example, the apparatus embodiments described above are merely exemplary, for example, the part or module division,
Only a kind of division of logic function, formula that in actual implementation, there may be another division manner, such as multiple portions or module can be with
In conjunction with being either desirably integrated into a system or some features can be ignored or does not execute.
In the disclosure, term "and/or" may have been used.As used herein, term "and/or" means one
Or other or both (for example, A and/or B mean A or B or both A and B).
In the above description, for purpose of explanation, elaborate numerous details in order to provide each reality to the disclosure
Apply the comprehensive understanding of example.However, the skilled person will be apparent that, without certain in these details
It can implement one or more other embodiments.Described specific embodiment be not limited to the disclosure but in order to illustrate.
The scope of the present disclosure is not determined by specific example provide above, is only determined by following claim.At other
In the case of, in form of a block diagram, rather than it is illustrated in detail in known circuit, structure, equipment, and operation is so as not to as making to retouching
The understanding stated thickens.Thinking suitable for place, the ending of reference numeral or reference numeral is weighed in all attached drawings
It is multiple to indicate optionally correspondence or similar element with similar characteristics or same characteristic features, unless otherwise specifying or
Obviously.
Various operations and methods have been described.Certain methods are carried out in a manner of comparative basis in way of flowchart
Description, but these operations are optionally added to these methods and/or are removed from these methods.In addition, although flow
The particular order of the operation according to each example embodiment is illustrated, it is to be understood that, which is exemplary.It replaces real
These operations can optionally be executed, combine certain operations, staggeredly certain operations etc. in different ways by applying example.Equipment is herein
Described component, feature and specific optional details can also may be optionally applied to method described herein, in each reality
It applies in example, these methods can be executed by such equipment and/or be executed in such equipment.
Each functional unit/subelement/module/submodule can be hardware in the disclosure, for example the hardware can be electricity
Road, including digital circuit, analog circuit etc..The physics realization of hardware configuration includes but is not limited to physical device, physics device
Part includes but is not limited to transistor, memristor etc..Computing module in the computing device can be any appropriate hard
Part neural network chip, such as CPU, GPU, FPGA, DSP and ASIC etc..The storage unit can be any magnetic appropriate
Storage medium or magnetic-optical storage medium, such as RRAM, DRAM, SRAM, EDRAM, HBM, HMC etc..
It is apparent to those skilled in the art that for convenience and simplicity of description, only with above-mentioned each function
The division progress of module, can be as needed and by above-mentioned function distribution by different function moulds for example, in practical application
Block is completed, i.e., the internal structure of device is divided into different function modules, to complete all or part of work(described above
Energy.
Particular embodiments described above has carried out further in detail the purpose, technical solution and advantageous effect of the disclosure
Describe in detail bright, it should be understood that the foregoing is merely the specific embodiment of the disclosure, be not limited to the disclosure, it is all
Within the spirit and principle of the disclosure, any modification, equivalent substitution, improvement and etc. done should be included in the protection of the disclosure
Within the scope of.
Claims (14)
1. a kind of intelligence navigation equipment, it is characterised in that including：
Target Identification Unit for handling the image shot when navigation, and is compared with sample database, identifies image
In obstacle identity；
Information processing unit, it is defeated for the obstacle identity of receive that treated image and identification, and by neural network computing
Go out navigation selection data.
2. intelligence navigation equipment according to claim 1, which is characterized in that described information processing unit includes：Nerve net
Network chip, for executing the neural network computing.
3. intelligence navigation equipment according to claim 2, which is characterized in that the neural network chip includes：
Storage unit stores neural network parameter and computations, the input data includes place for storing input data
The obstacle identity of image and identification after reason；
Control unit parses the computations and obtains multiple operational orders for extracting computations from the storage unit；
Arithmetic element, for navigation selection data to be calculated to input data execution according to multiple operational order.
4. intelligence navigation equipment according to claim 3, which is characterized in that the storage unit includes：
Input/output module, for obtaining input data, neural network parameter and computations；
Scalar data memory module, for storing scalar data；
Storage medium, for storing data block, and the storage medium are on piece storage.
5. intelligence navigation equipment according to claim 3, which is characterized in that the neural network chip further includes：
Direct memory access DMA, the data for being stored in storage unit, neural network parameter and computations, for control
Unit and/or arithmetic element are called.
6. intelligence navigation equipment according to claim 5, which is characterized in that the neural network chip further includes：
Instruction buffer, for from the direct memory access DMA cache instructions, being called for control circuit, described instruction caching is
On piece caches.
7. intelligence navigation equipment according to claim 5, which is characterized in that the neural network parameter includes input nerve
Member, output neuron and weights, the processor further include：
Neuron caching is inputted, for caching input neuron from the direct memory access DMA, is called for arithmetic element；
Weights cache, and for caching weights from the direct memory access DMA, are called for arithmetic element；
Output neuron caches, and the output neuron after operation is obtained for storing from the arithmetic element, to export to direct
Memory access DMA；
The wherein described input neuron cache module, weights cache module and output neuron cache module are on piece caching.
8. intelligence navigation equipment according to claim 2, which is characterized in that described information processing unit includes：
Input data encoder can be handled for the obstacle identity of treated image and identification to be changed into neural network
Digital information.
9. intelligence navigation equipment according to claim 8, which is characterized in that described information processing unit further includes：
Memory, for storing voyage conditions and navigation in different time periods selection data in current slot；
Arithmetic unit calculates the risk of navigation and/or the loss of current flight for selecting queuing data according to navigation.
10. intelligence navigation equipment according to claim 9, which is characterized in that described information processing unit further includes：
Data input/output terminal：Signal for receiving input data encoder, and it is passed to artificial neural network chip work
For input；It is additionally operable to receive the navigation selection data of neural network chip output end output, and is stored in memory；And for from
Memory reads navigation selection queue and the voyage conditions in the current period, is input to arithmetic unit.
11. intelligence navigation equipment according to claim 10, which is characterized in that described information processing unit further includes：
Output data converter, the risk of the navigation for calculating arithmetic unit and/or the lossy data of current flight are compiled again
It is transferred to external equipment after code.
12. intelligence navigation equipment according to claim 11, which is characterized in that the output data converter is by arithmetic unit
The result of calculating is back to Target Identification Unit, and the update sample database.
13. intelligence navigation equipment according to claim 12, which is characterized in that the neural network chip is additionally operable to outside
After portion's equipment or crewman make navigation selection, input-output information group is collected, neural network parameter is adaptively updated, training
Generate new neural network parameter.
14. intelligence navigation equipment according to claim 1, which is characterized in that the Target Identification Unit includes：
Image processing unit, for input picture to be pre-processed, binaryzation and selection subregion extract feature；
Comparing unit must be read, the preliminary obstacle mesh identified in image for the sample progress in the feature and sample database by extraction
Mark type.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810486527.8A CN108764465B (en) | 2018-05-18 | 2018-05-18 | Processing device for neural network operation |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810486527.8A CN108764465B (en) | 2018-05-18 | 2018-05-18 | Processing device for neural network operation |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108764465A true CN108764465A (en) | 2018-11-06 |
| CN108764465B CN108764465B (en) | 2021-09-24 |
Family
ID=64007269
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810486527.8A Active CN108764465B (en) | 2018-05-18 | 2018-05-18 | Processing device for neural network operation |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108764465B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110597756A (en) * | 2019-08-26 | 2019-12-20 | 光子算数(北京)科技有限责任公司 | Calculation circuit and data operation method |
| WO2023279701A1 (en) * | 2021-07-08 | 2023-01-12 | 嘉楠明芯(北京)科技有限公司 | Hardware acceleration apparatus and acceleration method for neural network computing |
| CN116395105A (en) * | 2023-06-06 | 2023-07-07 | 海云联科技(苏州)有限公司 | Automatic lifting compensation method and system for unmanned ship |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20140371912A1 (en) * | 2013-06-14 | 2014-12-18 | Brain Corporation | Hierarchical robotic controller apparatus and methods |
| CN105512723A (en) * | 2016-01-20 | 2016-04-20 | 南京艾溪信息科技有限公司 | Artificial neural network calculating device and method for sparse connection |
| US20170061625A1 (en) * | 2015-08-26 | 2017-03-02 | Digitalglobe, Inc. | Synthesizing training data for broad area geospatial object detection |
| CN106919915A (en) * | 2017-02-22 | 2017-07-04 | 武汉极目智能技术有限公司 | Map road mark and road quality harvester and method based on ADAS systems |
| CN107092252A (en) * | 2017-04-11 | 2017-08-25 | 杭州光珀智能科技有限公司 | A kind of robot automatic obstacle avoidance method and its device based on machine vision |
| US9760806B1 (en) * | 2016-05-11 | 2017-09-12 | TCL Research America Inc. | Method and system for vision-centric deep-learning-based road situation analysis |
| CN107563332A (en) * | 2017-09-05 | 2018-01-09 | 百度在线网络技术（北京）有限公司 | For the method and apparatus for the driving behavior for determining unmanned vehicle |
| CN107886099A (en) * | 2017-11-09 | 2018-04-06 | 电子科技大学 | Synergetic neural network and its construction method and aircraft automatic obstacle avoiding method |
-
2018
- 2018-05-18 CN CN201810486527.8A patent/CN108764465B/en active Active
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20140371912A1 (en) * | 2013-06-14 | 2014-12-18 | Brain Corporation | Hierarchical robotic controller apparatus and methods |
| US20170061625A1 (en) * | 2015-08-26 | 2017-03-02 | Digitalglobe, Inc. | Synthesizing training data for broad area geospatial object detection |
| CN105512723A (en) * | 2016-01-20 | 2016-04-20 | 南京艾溪信息科技有限公司 | Artificial neural network calculating device and method for sparse connection |
| US9760806B1 (en) * | 2016-05-11 | 2017-09-12 | TCL Research America Inc. | Method and system for vision-centric deep-learning-based road situation analysis |
| CN106919915A (en) * | 2017-02-22 | 2017-07-04 | 武汉极目智能技术有限公司 | Map road mark and road quality harvester and method based on ADAS systems |
| CN107092252A (en) * | 2017-04-11 | 2017-08-25 | 杭州光珀智能科技有限公司 | A kind of robot automatic obstacle avoidance method and its device based on machine vision |
| CN107563332A (en) * | 2017-09-05 | 2018-01-09 | 百度在线网络技术（北京）有限公司 | For the method and apparatus for the driving behavior for determining unmanned vehicle |
| CN107886099A (en) * | 2017-11-09 | 2018-04-06 | 电子科技大学 | Synergetic neural network and its construction method and aircraft automatic obstacle avoiding method |
Non-Patent Citations (2)
| Title |
|---|
| ZHANG S 等: "Cambricon-X: An accelerator for sparse neural networks", 《2016 49TH ANNUAL IEEE/ACM INTERNATIONAL SYMPOSIUM ON MICROARCHITECTURE》 * |
| 陈云霁: "从人工智能到神经网络处理器", 《领导科学论坛》 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110597756A (en) * | 2019-08-26 | 2019-12-20 | 光子算数(北京)科技有限责任公司 | Calculation circuit and data operation method |
| WO2023279701A1 (en) * | 2021-07-08 | 2023-01-12 | 嘉楠明芯(北京)科技有限公司 | Hardware acceleration apparatus and acceleration method for neural network computing |
| CN116395105A (en) * | 2023-06-06 | 2023-07-07 | 海云联科技(苏州)有限公司 | Automatic lifting compensation method and system for unmanned ship |
| CN116395105B (en) * | 2023-06-06 | 2023-10-20 | 海云联科技(苏州)有限公司 | Automatic lifting compensation method and system for unmanned ship |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108764465B (en) | 2021-09-24 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108764470A (en) | A kind of processing method of artificial neural network operation | |
| CN109635744B (en) | Lane line detection method based on deep segmentation network | |
| CN108805064A (en) | A kind of fish detection and localization and recognition methods and system based on deep learning | |
| CN109508634B (en) | Ship type identification method and system based on transfer learning | |
| CN111507993A (en) | Image segmentation method and device based on generation countermeasure network and storage medium | |
| CN108764465A (en) | A kind of processing unit carrying out neural network computing | |
| CN113850848B (en) | Marine multi-target long-term detection and tracking method based on cooperation of unmanned ship carrying navigation radar and visual image | |
| CN109131348A (en) | A kind of intelligent vehicle Driving Decision-making method based on production confrontation network | |
| CN107731011B (en) | Port berthing monitoring method and system and electronic equipment | |
| Song et al. | Semi-supervised dim and small infrared ship detection network based on haar wavelet | |
| CN109466725B (en) | Intelligent water surface floater fishing system based on neural network and image recognition | |
| CN110163207A (en) | One kind is based on Mask-RCNN ship target localization method and storage equipment | |
| CN109583343A (en) | A kind of fish image processing system and method | |
| CN108647781A (en) | A kind of artificial intelligence chip processing device | |
| CN110991257A (en) | Polarization SAR oil spill detection method based on feature fusion and SVM | |
| Zhao et al. | CRAS-YOLO: A novel multi-category vessel detection and classification model based on YOLOv5s algorithm | |
| CN115527103A (en) | Unmanned ship perception experiment platform system | |
| Sun et al. | IRDCLNet: Instance segmentation of ship images based on interference reduction and dynamic contour learning in foggy scenes | |
| Yi et al. | Research on Underwater small target Detection Algorithm based on improved YOLOv7 | |
| CN113837924A (en) | Water bank line detection method based on unmanned ship sensing system | |
| CN113280820A (en) | Orchard visual navigation path extraction method and system based on neural network | |
| Qiu et al. | Underwater sea cucumbers detection based on pruned SSD | |
| CN116863293A (en) | Marine target detection method under visible light based on improved YOLOv7 algorithm | |
| CN116246139A (en) | Target identification method based on multi-sensor fusion for unmanned ship navigation environment | |
| CN115661657A (en) | Lightweight unmanned ship target detection method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |