CN109522778A - A kind of image-recognizing method can reach image understanding level - Google Patents
A kind of image-recognizing method can reach image understanding level Download PDFInfo
- Publication number
- CN109522778A CN109522778A CN201710905332.8A CN201710905332A CN109522778A CN 109522778 A CN109522778 A CN 109522778A CN 201710905332 A CN201710905332 A CN 201710905332A CN 109522778 A CN109522778 A CN 109522778A
- Authority
- CN
- China
- Prior art keywords
- image
- probability
- characteristic point
- space
- scale
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
-
- 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/047—Probabilistic or stochastic networks
-
- 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
- G06N3/08—Learning methods
- G06N3/088—Non-supervised learning, e.g. competitive learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
- G06V20/588—Recognition of the road, e.g. of lane markings; Recognition of the vehicle driving pattern in relation to the road
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
- G06V30/10—Character recognition
- G06V30/19—Recognition using electronic means
- G06V30/192—Recognition using electronic means using simultaneous comparisons or correlations of the image signals with a plurality of references
- G06V30/194—References adjustable by an adaptive method, e.g. learning
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Probability & Statistics with Applications (AREA)
- Databases & Information Systems (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
Abstract
The present invention relates to the image-recognizing methods that one of field of information processing can reach image understanding level, the maximum a plurality of characteristic points of probability density in image are obtained by machine learning it is characterized by: referring to, maximum probability scale of each characteristic point by carrying including characteristic point, characteristic point is k-th, k-th of characteristic point to the vector value of+1 characteristic point of kth etc. structuring information, by non-structured image data carry out structuring characterization.In image recognition, the structured message of each characteristic point of sample image between the probability distribution data of the structured message of each image obtained by machine learning across at a distance from Euclidean space and probability space, and estimate by the fuzzy event probability of foundation of this distance, use this scale to be identified as according to realization according to the feature of structuring.Implementation result of the present invention is: the understanding to image can be achieved, still can correctly identify by the image of part.
Description
[technical field]
The invention belongs to the image-recognizing methods that one of artificial intelligence field can reach image understanding level.
[background technique]
Automatic Pilot is the main battle ground of artificial intelligence, but regrettably for the dedicated engineering of automatic Pilot application
The research achievement of habit is very few, so far still without causing widely to pay attention to.
Famous Toyota Company, Japan has delivered the patent (patent document 1) of " drive and be directed toward estimating device ", which mentions
Out according to automatic driving during, for burst the case where, even if in the case that driver is not reflected, passing through artificial intelligence
The machine learning algorithm of the inverse transmission network of energy, selects driving condition, automatically to avoid the generation etc. of driving accident.
On October 9th, 2016 Japan NHK by explain committee member room mountain wise man also explained " how cross over automatic Pilot wall
Build " (non-patent literature 1).On this special topic, room mountain wise man, which also proposed in automated driving system, at present also be can't resolve
Railway Project:
Man-machine judgement conflict is difficult: 2 months this year carried out the experiment of automatic Pilot in Google, when Google's vehicle is turned right, before
There is the barrier of a sand drift, Google's vehicle is anxious to be gone to right lane sudden strain of a muscle, a bus has just at this moment been carried out behind right lane, bus is driven
The person of sailing thinks that Google's vehicle can brake rapidly, and I had never expected can dodge toward right lane, therefore the major accident collided occurs.
The difficulty of the man-machine feeling fusion of spacing selection: there is 41% driver when vehicle is met in road by the investigation of certain company
It is to think more remoter better from this vehicle, but also it is believed that and then fixed range is relatively good, or thinks good close to a bit, somebody
Entertain the vehicle that curiosity will catch up with front.This is that man-machine feeling fusion, Yi Jiru how are solved the problems, such as autonomous driving vehicle
Closest to the driving problem of people, it is difficult that this becomes complicated automatic control for what selection.
Man-machine permission transfer problem: in the conversion stage of human-machine operation, it is man-machine between consciousness cannot transmit.Such as
In case of emergency, the moment of pilot steering, automatic Pilot emergency plan selected at that time need to be switched at once by automatic Pilot
It is different from pilot steering, it is be easy to cause accident, or affect opportunity adversely.
Trolley problem: encountering Emergency time, it is down to the quantity of victim at least? this is famous
Trolley problem had both been related to complicated ethics difficulty and technological difficulties.In the automatic Pilot theory of machine learning, at present
Still nobody proposed valuable solution.
By the paper (non-patent literature of " realizing train automatic Pilot with Fuzzy Predictive Control " that the great peace in the big island of Hitachi is delivered
2), which proposes the automatic Pilot that train may be implemented by the rule base of fuzzy deduction.
" using the train automated driving system of predicative fuzzy control " (the non-patent text delivered by University of tsukuba's peace letter sincere two
It offers 3), although proposing that traditional PID adjusts the traveling that can accurately control train automatic Pilot, round and smooth walking is automatic
The key of driving, and the key that the person of multiplying feels comfortable.And the key problem of automatic Pilot is one and needs to consider safety,
Traveling speed, the travel time between station, comfort consume more purpose control problems of electric power and stopping accuracy.
[patent document]
[patent document 1] (special open 2008-225923)
[non-patent literature 1]Http:// www.nhk.or.jp/kaisetsu-blog//255089.html
[non-patent literature 2]
https://www.jstage.jst.go.jp/article/ieejeiss1987/109/5/109_5_337/_
pdf
[non-patent literature 3]
http://ttt.akiba.coocan.jp/yasunobu/edu/intconthtms/text/Sic07a_
TrainATO.pdf#search=%27, which gives, sees that Off ァ ジ ィ mono- makes imperial train and transports translocation シ ス テ system %27 automatically
Above-mentioned (patent document 1), using the neural network algorithm of artificial intelligence, still, in neural network algorithm mainly
The information of objective function is carried on by " training " in the parameter of the weighted value of magnanimity, weighted value W, with threshold values T in study
In the process, to obtaining optimal solution, need that all states will be tested, combined total degree be (W ×
T)n} × P, the number of nodes for the neural network that n is one layer here, P are the number of plies of neural network, and the calculating of such high index is complicated
Degree keeps calculation amount huge, and required hardware spending is huge, probability used by the loss function of the learning effect of deep learning
Gradient descent method abbreviation SGD, obtained trained values are a local optimum solutions, therefore inevitably will appear "black box" and ask
Topic, then to have the threshold values in the model of neural network to belong to artificially defined, the mechanism with the neural network of the brain of people has no bearing on,
The principle of the stimulus signal of cranial nerve cannot at all embody in traditional neural network model, and the brains of people is according to neuron
Excitement degree difference caused by nerve signal does not carry out the mechanism of different judgements in the model of current neural network also not
Can embody etc., current neural network model can only be it is learned, represent a kind of theory of directionality, actually answer with reaching
Level disparity is very big.Nowadays in the stage for entering deep learning, hidden layer is merely added compared with traditional neural network
Quantity, this more increases the complexity calculated, and the fatal black box problem of traditional neural network cannot solve, in automatic Pilot
In there is security risk, the prospect of application is difficult to expect.
" man-machine judgement conflict is difficult " that above-mentioned (non-patent literature 1) is proposed, " the man-machine feeling fusion of spacing selection
It is difficult ", " man-machine permission transfer problem " and " Trolley problem ", which should become in artificial intelligence automatic Pilot, to be had to solve
The problem of, the key problem of Ying Chengwei artificial intelligence, however do not cause but widely to pay attention at present.
Above-mentioned (non-patent literature 2) main automatic Pilot problem for being to solve train, what is proposed utilizes fuzzy deduction
Rule base realize the automatic Pilot of train, the foundation of huge knowledge base needs the study of big data, can only solve two or three
Within objective function, in automatic driving using relatively difficult.
Above-mentioned (non-patent literature 3) is although propose the fuzzy control of more purposes, also in that used fuzzy control
In the control of the more objective function of correspondence relatively reluctantly, it therefore remains in for each specific objective function
Individually control.The fusion and safety of man-machine feeling especially in automatic Pilot, energy conservation comfortably wait objective functions, more mesh
Target controls simultaneously, since different objective functions is not in the same space, can not find a common optimized control point,
Even if being mapped to the same space, the traditional method of different objective functions can not obtain a common best intersection point, be
This, it is necessary to the redundancy between more purpose Optimal Controls is found, really realizes and realizes the optimal of more purposes in optimum range
Control, therefore, it is necessary to solve the foundation of the machine learning model of more purpose controls.
First technical term involved by present patent application is defined below, it is defined below in comprising the present invention in
Hold.
Probability space (Probability Space):
" probability theory is based on measure theory " based on Soviet Union mathematician Andrey Kolmogorov is theoretical, so-called
Probability space be exactly one and total estimate for the measurable space of " 1 ".
It can produce lemma according to the definition of probability space: only existing a probability distribution, probability space in probability space
Distance is centainly across the distance between plural probability space of Euclidean space.
Probability distribution (Probability Distribution):
So-called probability distribution is for probability function, by the arrangement of the size of its possibility occurrence.
Probability scale (Probability Scale):
Any one probability distribution in probability space certainly exists a probability scale, can demarcate the journey of probability distribution
Degree.
It can produce lemma according to the definition of probability space: by probability scale, the institute of inevitable available probability space
There is probability distribution information.
Probability density (Probability Density):
Probability value of the event on given area.
Fuzzy event probability estimates (Probability Measure of Fazzy Event)
In the Euclidean space S including probability space, if p (x) meets the additive property of probability measure, μA(x)
It is membership function (Membership Function) while the additive property for also meeting fuzzy mearue, then fuzzy event probability
Set A estimate P (A) are as follows:
Formula 1
P (A)=∫SμA(x)p(x)dx
Its discrete expression formula are as follows:
Formula 2
Intelligence system (Intelligent System):
Intelligence system is the system realized according to deterministic algorithm, is to realize certain objective function according to a kind of algorithm
Processing, processing result is deterministic system.
Artificial intelligence (Artificial Intelligence):
What is artificial intelligence? briefly it is exactly the brains function of realizing people with computer, i.e., people is realized by computer
Brains thinking caused by effect, problem to be solved, processing result is often probabilistic, or perhaps in advance
It is unpredictable.
It may also be said that artificial intelligence be by effect model caused by artificial intervention, it is formulation, more specifically define,
Artificial intelligence is that the most in-depth application of probabilistic model is realized by the algorithm of machine learning.
It clusters (Clustering):
It is static treatment process according to a kind of algorithm of scale aggregation data in fixed field.
Self-organizing (Self-organization):
The acquisition objective function of self-discipline, the algorithm migrated to the direction of high probability that data can be allowed to restrain oneself, is dynamic
Treatment process.
The definition of machine learning (Machine Learning):
The model of rule is obtained in the slave data of computer self-discipline.
Probability scale self-organizing (Self-organizing based on a Probability Scale):
If probability space there is the set G of a following probability distribution, contains ζ data: formula 3 in set
gf∈ G (f=1,2 ..., ζ)
In this probability distribution g of probability spacefIn (f=1,2 ..., ζ), one is certainly existed at characteristic value A (G), by
It is measure space in probability space, therefore certainly exists a probability scale M [G, A (G)] for characteristic value A (G), meets as follows
When the condition of probability scale self-organizing, set G can be allowed on the basis of probability scale(n)It is migrated towards maximum probability direction.
Formula 4
A(n)=A (G(n))
M(n)=M [G(n), A (G(n))]
G(n)=G { [A (G(n-1)), M [G(n-1), A (G(n-1))]]
When n >=β (β is the numerical value greater than 4), A (G can be set(n)) it is maximum probability characteristic value, M [G(n), A (G(n))] it is maximum probability scale centered on maximum probability characteristic value.
High-order maximum likelihood (Higher Order Maximum Probability):
The calculated average value of beyond tradition statistics institute and variance yields, i.e., closest parent are predicted flat
Mean value and variance yields.
Probability space distance (Probabilistic Spatial Distance)
If the central point w in the probability distribution W of a probability space included in the E of Euclidean space, and in Europe
Be between central point v in the probability distribution V of another probability space included in a few Reed space E one pass through Europe it is several in
Two positions in moral space and probability space reach the distance of v from w:
Formula 5
Here, Δ(V)For by the direction of w to v, after entering the probability space of probability distribution V from Euclidean space,
Generated range error, Δ in probability space(V)In (V) expression passed through since the position of w Euclidean space enter it is general
Range error caused by the probability value for the probability distribution V that rate space is passed through, that is to say, that Δ(V)It is with w by probability distribution
The related error amount of the probability value of V, and divide with the probability passed through when passing through probability distribution W since the center w of probability distribution W
The probability value of cloth W is unrelated.
Ultra-deep study (Super Deep Learning):
By sensing layer, what nervous layer and cortex were constituted, and by between node between layers with probability scale from
The new neural network model for the distributed machine learning that the machine learning of tissue is connected.
[summary of the invention]
The first purpose of this invention is: proposing a kind of automotive environment image recognition algorithm suitable for automatic Pilot, passes through
Automotive environment image is mapped to several mapping images suitable for the automotive environment image recognition of automatic Pilot by space reflection, is led to
The identification for crossing the maximum probability region to each mapping image, the structural information of non-structured automotive environment image is taken out
Out, to reach the understanding to automotive environment image.
Second object of the present invention is: proposing a kind of more purpose Optimal Control machine learning models suitable for automatic Pilot
And system and device, it can be by safe driving, vehicle headway, accurate stopping quickly reaches, comfortable to ride, more mesh such as energy saving
Objective function carry out the control of optimal machine learning model.
Third object of the present invention is: allowing automatic Pilot that can imitate the drive manner of excellent driver, improves automatic
The level of driving.
Fourth object of the present invention is: solving the problems, such as that black box and single-chip offline mode can also be normally carried out
Automated driving system,
A kind of image-recognizing method can reach image understanding level, it is characterised in that:
(1) non-structured image data is carried out to the characterization of structuring；
(2) it is identified according to the feature of structuring.
Moreover, the characterization that non-structured image data is carried out to structuring, refers to and is obtained by machine learning
Obtain the maximum a plurality of characteristic points of probability density in image；Maximum probability ruler of each characteristic point by carrying including characteristic point
Degree, the probability distribution of maximum probability scale；Or the probability distribution that characteristic point is k-th；Or k-th of characteristic point is to+1 spy of kth
Levy the information of at least one structuring including the vector value and probability distribution of point.
Moreover, described identified according to the feature of structuring, refer in image recognition, each of sample image is special
The probability distribution data for levying the structured message and the structured message by machine learning each image obtained of point, carry out
It is surveyed including solving across Euclidean space at a distance from probability space, and by the fuzzy event probability of foundation of this distance
Degree, at least one scale carries out confrontation study as foundation between the logon data of each image, obtain one it is closest
Data, image corresponding to the data is exactly recognition result, to realize the understanding to image.
A kind of image-recognizing method can reach image understanding level proposed by the present invention, its main feature is that: it can be achieved to figure
The understanding of picture still can correctly be identified by the image of part.
Detailed description of the invention
Fig. 1 solves the problems, such as two kinds of space mapping method schematic diagrames that complexity is
Fig. 2 is suitable for the method schematic diagram that the automotive environment image radiation of automatic Pilot maps
Fig. 3 realizes the recognition methods schematic diagram of road fare by unsupervised machine learning
Fig. 4 realizes the flow chart of the identification of road fare by unsupervised machine learning
Fig. 5 carries out the schematic diagram of the mapping method of burst images from automotive environment image
Fig. 6 carries out the flow chart of the mapping method of burst images from automotive environment image
Fig. 7 is adapted for the schematic diagram of the processing method of the automotive environment image understanding of automatic Pilot
Fig. 8 is by the structural information extraction method schematic diagram of unstructured image data
Fig. 9 is the definition schematic diagram for the distance for passing through the different spaces including probability space
Figure 10 is the schematic diagram of the optimally-controlled machine learning model of more purposes
Figure 11 is the schematic diagram that excellent driver " Knowledge Acquirement " is realized by machine learning
Symbol description
201 be the original image of automotive environment image
202 be a mapping image of the original image of automotive environment image
203 be another mapping image of the original image of automotive environment image
204 be to be extracted out the characteristic value for mapping image by unsupervised learning
205 be neural network sensing layer
206 be a node of neural network sensing layer
300 be the discrete dot matrix got on the bus near line position in edge image
301 be the center line of the discrete dot matrix near fare position
302 and 303 be the straight line on the both sides of center line 301
700 be a mapping image of automotive environment image
701 be the region of a maximum probability density
702 be the region of another maximum probability density
703 be the region of another maximum probability density
704 be the region for finally obtaining a maximum probability density
801 be probability density and the maximum characteristic point of area in mapping image
802 be the radius of the maximum probability density of characteristic point 801
803 be the vector value that characteristic point 801 turns to next characteristic point 805
804 be the probability space that is passed through of vector value of the next characteristic point 805 of 801 steerings
805 be next characteristic point
806 be the radius of the maximum probability density of second characteristic point 805
807 be the vector value that characteristic point 805 will turn to next characteristic point
808 be the probability space that second characteristic point 805 must pass through the vector value for turning to next characteristic point
809 be the probability space that is passed through of vector value 810 for the final characteristic point 811 that a upper characteristic point is up to
810 be the vector value of final characteristic point 811
811 be final characteristic point
812 be the radius of the maximum probability density of final characteristic point 811
901 be an Euclidean space for covering probability space
902 be the central point w of a probability distribution of probability spacej
903 be first scale M of more probability scales of a probability distribution of probability space1
904 be second scale M of more probability scales of a probability distribution of probability space2
905 be the third scale M of more probability scales of a probability distribution of probability space3
909 be the point v of Euclidean space onej
Specific embodiment
The embodiment of the present invention is further described below in conjunction with attached drawing, but embodiment of the present invention is illustrative
, without being restrictive.
Fig. 1 is two kinds of space mapping method schematic diagrames for solving the problems, such as complexity and being.
If (a) of Fig. 1 is the schematic diagram for being mapped to n space the problem of being by a complexity, radiation mapping can be described as.
Such as the identification for image, one automotive environment image can be mapped to several mapping graphs by the model of physics
Picture, in order to extract image the deep information out.
It is problem if (b) of Fig. 1 is by the complexity of several different spaces, is mapped to the schematic diagram in the same space, it can
It is known as collecting shape mapping.Such as the Optimal Control problem for more purposes, there are multiple objective functions, respectively not in a space,
It is difficult to establish an optimal control point, by collecting shape space reflection, so that it may which the objective function of more purposes is mapped to one
A MemberShip fuzzy space, therefore the optimally-controlled method of more purposes can be sought.
Fig. 2 is adapted for the method schematic diagram of the automotive environment image radiation mapping of automatic Pilot.
The accuracy key of image recognition is exactly to pay attention to the processing of two links, one be image information extraction link, one
A is the pattern-recognition link of image.Here the image information extraction side of the main image recognition for proposing to be suitable in automatic Pilot
Method.
As shown in Figure 2: 201 be the original image of automotive environment image, and 202 be a mapping of the original image of automotive environment image
Image, 203 be another mapping image of the original image of automotive environment image, in order to take out the deep information of automotive environment image
Out, several mapping images can be launched by mapping automotive environment image.
204 be to be extracted out the characteristic value for mapping image by unsupervised learning, and 205 be neural network sensing layer, and 206 be mind
A node through network aware layer.
In pattern-recognition, the importance of the quality of information extraction will be much higher than the sorting algorithm of machine learning, based on sky
Between mapping theory, automotive environment image is mapped out into several images by physical model, can be by the depth of automotive environment image
Layer information is directly extracted out, each characteristic information region corresponds to a machine learning, can be in corresponding image-region
" high-order maximum likelihood " value extraction, and each characteristic area correspond to a unsupervised learning.
Non-structured image is carried out information extraction by the structure of image by present patent application, can be with using unsupervised learning
The region for automatically moving to the maximum probability distribution of image, using several maximum probability distributed areas in image as image
Structure feature carry out information extraction, so that the structural information of image is obtained, as long as image a part enters the visual field of identification
In can correctly identify the image, to achieve the effect that image understanding.
In order to identify the lane line of road, need the original image of automotive environment image carrying out edge image
Mapping.It is mapped for edge image, needs to carry out the adjacent pixel of original image the calculating of derivative value, traditional method is generally adopted
With calculus of finite differences, but as the image recognition in automatic Pilot field, need to consider the image of motor vehicle environment in bad weather
Picture quality it is very poor, therefore as far as possible using high reliability, the algorithm of noise jamming can be offset.Known solving to be directed to
Other automotive environment image carries out in the mapping problems of edge image, and present patent application proposes the derivative value of multiple adjacent pixels
Average algorithm obtain the method that the image derivative value of higher precision can be obtained by multiple adjacent pixels, such as with two
Adjacent derivative value, i.e., adjacent three pixels obtain the algorithm of edge image specific formula is as follows:
[formula 1]
Here, xi-1-xi=1, xi-2-xi=2, i=1,2 ..., n, j=1,2 ..., m can thus pass through adjacent x
Three pixels in direction obtain the derivative value in the direction x of the higher image of precision.It is same:
[formula 2]
Here yi-1-yi=1, yi-2-yi=2, i=1,2 ..., n, j=1,2 ..., m.Equally pass through the adjacent direction y
Three pixels obtain the derivative value in the direction y of the higher image of precision.Similarly can also by 3 derivatives, or more adjacent x
The derivative value of the image pixel in direction or the direction y obtains the high-precision derivative value in the region.
Image data after derivation is not ideal data, wherein each derivation value is discrete, but also
There is a large amount of noise, present patent application proposition is following to utilize unsupervised machine learning, realizes that fare knows method for distinguishing.
Fig. 3 is the recognition methods schematic diagram that road fare is realized by unsupervised machine learning.
As shown in Figure 3: 300 be the discrete dot matrix got on the bus near line position in edge image, and 301 be that be located at fare position attached
The center line of close discrete dot matrix, this line can be arbitrary about in the initial state the one of discrete dot matrix center
The straight line that item is connected by two points, this straight line can obtain " high during following probability scale self-organizings
The central line of rank maximum likelihood ", 302 and 303 be the straight line on the both sides of center line 301, equally in following probability scales
During self-organizing, this two lines is exactly the distance of the scale of one " high-order maximum likelihood " at a distance from center line, that is, is existed
The scale of the maximum probability finally obtained in probability scale self-organizing process.
Fig. 4 is the flow chart that the identification of road fare is realized by unsupervised machine learning.
Such as Fig. 3, shown in Fig. 4: the processing step for carrying out fare identification using unsupervised machine learning is as follows:
S1Be initialization step, mainly provide initial range rough present in fare, for example, Fig. 3 302 and
303。
S2It is the content for being also initialization step, in S1It can in range present in the fare produced in initialization step
At will to be constituted center line from finding out two points in range in both ends, such as the 301 of Fig. 3.
S3Into the iterative step of unsupervised learning, find out dot matrix all in given range to center line distance.
S4It is to solve for the probability scale step of unsupervised machine learning, that is, solves dot matrix all in given range in
The maximum probability scale and central value of the distance of heart line.The method for solving probability scale has: normal distribution, multivariate normal point
Cloth, logarithm normal distribution, exponential distribution, t distribution, F distribution, X2Distribution, bi-distribution, negative bi-distribution, multinomial distribution, pool
Pine distribution, Erlangian distribution (Erlang Distribution), hypergeometric distribution, geometry distribution, traffic distribution, weber point
Cloth (Weibull Distribution), angular distribution, beta are distributed (Bete Distribution), and gamma is distributed (Gamma
At least one of) Distribution the scale and central value of probability distribution are solved in probability distribution.
S5It is dot matrix processing step, i.e., with S4On the basis of the probability scale of the unsupervised machine learning solved, by probability ruler
Dot matrix within degree retains, and the dot matrix other than probability scale is rejected.
S6It is to constitute new center line and range step, by S4It solves in the probability scale step of unsupervised machine learning
Obtained variance and average value as new range such as Fig. 3 302,303 and center line such as Fig. 3 301.
S7It is to differentiate whether iteration restrains step, "No" just jumps to S3Continue the iterative processing of unsupervised learning, "Yes"
Enter S8The processing step that terminates of iteration.
By such unsupervised machine learning, can almost all noise filterings, available beyond tradition
Statistics 302,303 and average value such as Fig. 3 of calculated variance yields such as Fig. 3 301, and surmount by linear
Planning the result of calculated fare accuracy.
Present patent application also proposes how the mapping method that burst images are carried out from automotive environment image.
Fig. 5 is the schematic diagram that the mapping method of burst images is carried out from automotive environment image.
As shown in Figure 5: Fig. 5 (a) is the schematic diagram of automotive environment image under normal circumstances, this is the environment of a safety
Image, the driving that automobile can trust.Fig. 5 (b) is the ambient image before automotive environment image will face a danger, but conduct
The visual field of ambient image is still the image for belonging to safe condition, and Fig. 5 (c) is that automotive environment image has abruptly entered a danger
How does is the image of state separately separated out dangerous image, is mapped to a burst images? this is automatic Pilot image recognition one
Surely the important problem to be solved.
Solve the problems, such as that this can be carried out in terms of two.First is that referring to edge image map method, using formula 1 with
The image between image by (b) image in Fig. 5 and (c) of formula 2, and (c) three adjacent width images of image, or more
Derivation is carried out between the pixel of the image of width, obtains the set of the image pixel of a variation.Then above-mentioned probability scale is referred again to
Unsupervised machine learning shown in the algorithm of self-organizing finds out the maximum probability scale that a point is constituted with a radius and is contained
The field of lid, when this field is more than that there may be regarded as burst images when dangerous scale.
Fig. 6 is the flow chart that the mapping method of burst images is carried out from automotive environment image.
As shown in Figure 6: the mapping method that burst images are carried out from automotive environment image is divided into 8 steps, concrete operations
It is as follows:
S1It is to read adjacent pictures step, in the step Rigen according to the requirement of formula 1 and the Derivative Formula of formula 2, reads
Take the picture of adjacent 3 picture of automotive environment video image or more.
S2The step of being to solve for derivative image, for the picture of adjacent 3 picture of automotive environment video image or more
Face carries out each pixel corresponded between picture, carries out the calculating differentiated as formula 1 and formula 2, to generate prominent
Send out the mapping image of image.
S3It is setting search derivative image radius and central value step, in order to judge that the image for generating certain danger is still disturbed
Motion video needs to find out the field that maximum probability scale is covered by unsupervised machine learning, when this field is more than possible
Burst images are regarded as when generating dangerous scale, initial search range are set to derivative image for this, to derivative figure
As can entirely scan for, also it is segmented into several regions and searches for respectively.
S4It is to solve for maximum probability scale step in radius, i.e., calculates each dot matrix in for the dot matrix distribution in radius
The probability distribution of the distance of heart point, finds out the scale of the probability distribution of distance.Here the scale for calculating probability distribution can pass through:
Normal distribution, multivariate normal distribution, logarithm normal distribution, exponential distribution, t distribution, F distribution, X2Distribution, bi-distribution are born
Bi-distribution, multinomial distribution, Poisson distribution, Erlangian distribution (Erlang Distribution), hypergeometric distribution, geometry
Distribution, traffic distribution, Weibull distribution (Weibull Distribution), angular distribution, beta are distributed (Bete
Distribution), maximum probability point is solved at least one of gamma distribution (Gamma Distribution) probability distribution
The scale and central value of cloth.
S5It is the screening dot matrix step on the basis of maximum probability scale, the dot matrix belonged in range retains, other than range
Dot matrix is rejected.For the dot matrix being removed in the transition process that unsupervised machine learning is carried out, the dot matrix being removed still has can
It can enter in maximum probability range scale, participate in the operation of probability scale self-organizing.
S6It is center and the radius step to form new derivative image, the dot matrix being screened out is recalculated into central value,
S4Maximum probability scale is to solve in radius as new radius.
S7It is to judge whether iteration restrains step, it can be determined that whether maximum probability scale is stable no longer to change, or
It is believed that iteration has restrained when the very little of change, "Yes" then turns the operation of next step, and "No" is then jumped toward S4It solves in radius
Maximum probability scale step continues iterative processing.
S8It is to judge whether maximum probability scale M >=δ step, δ is a specified constant here, and "Yes" then enters next
Step S9, "No" then thinks the image for emergency event do not occur, jumps toward S3Setting search derivative image radius and central value walk
Suddenly, continue the appearance of the image of monitoring emergency event.
S9It is to return to main program, or the mapping image of emergency event is shown, and the center of automatic Pilot is notified to control
The operation that unit processed takes urgent measure etc..
In order to adapt to automatic Pilot in the image recognition of atrocious weather, needing can be right in a part for only obtaining image
Image is correctly identified that, to meet this requirement, following this patent proposes to be suitable for automatic Pilot, by non-structured image
Data carry out the information extraction of structuring, to obtain the effect of image understanding, the specific method is as follows:
Fig. 7 is adapted for the schematic diagram of the processing method of the automotive environment image understanding of automatic Pilot.
As shown in Figure 7: this is one by non-structured image data, the method for carrying out the information extraction of structuring.If
L1, L2..., Li..., LnIt is unsupervised machine learning, 700 be a mapping image of automotive environment image, uses L first1?
The unsupervised machine learning that self-organizing is carried out in entire mapping 700 range of image, obtains the region 701 of a maximum probability density
This is saved as a result, reusing L2Carry out the unsupervised machine learning of self-organizing again in entirely mapping 700 range of image, again
This is equally saved as a result, same processing method, reuses L to the region of another maximum probability density 702iSelf-organizing
Unsupervised machine learning, the region 703 for another maximum probability density of getting back equally save this as a result, finally reusing Ln
The unsupervised machine learning of self-organizing, the region 704 for finally obtaining a maximum probability density equally save this result.It is identical
The region of maximum probability density above image is that the region that can stablize identification has the repeatability of maximum probability, as being known
The structural information of non-structured image data can be extracted out, the understanding of image may be implemented by the characteristic point of other image.
Here the sequence of the characteristic point of the information with structuring is that unsupervised machine learning passes through probability scale for the first time
The characteristic point of the maximum region for the maximum probability density that self-organizing obtains, and so on the characteristic point region that learns in the backward just
It is smaller.
Fig. 8 is by the structural information extraction method schematic diagram of unstructured image data.
As shown in Figure 8: 801 be probability density and the maximum characteristic point of area in mapping image, and 802 be characteristic point
The radius of 801 maximum probability density, the here radius of maximum probability density, the both unsupervised learning of probability scale self-organizing
Maximum probability scale, it is same as below not repeat.803 be the vector value that characteristic point 801 turns to next characteristic point 805, and 804 are
The probability space that 801 vector values for turning to next characteristic point 805 are passed through, the information of probability space is by repeatedly to image
Study obtain, do not repeating equally below, 806 be the radius of the maximum probability density of second characteristic point 805, and 807 are
Characteristic point 805 will turn to the vector value of next characteristic point, and 808 be that second characteristic point 805 will turn to next characteristic point
The probability space that vector value must pass through, it is upper one that the characteristic point being finally reached by several characteristic points, which is 811,809,
Characteristic point is the probability space that the vector value 810 for the final characteristic point 811 being up to is passed through, and 812 be final characteristic point
The radius of 811 maximum probability density.
Here each characteristic value carries following information:
(1) the maximum probability scale of characteristic point, and probability distribution related with maximum probability scale size.
(2) probability distribution that characteristic point is k-th, this index can be with the stable case of evaluating characteristic point.
Vector value (length) and probability distribution of (3) k-th of the characteristic point to+1 characteristic point of kth.
Vector value (direction) and probability distribution of (4) k-th of the characteristic point to+1 characteristic point of kth.
In image recognition, each characteristic point of sample image needs to find out sample by the data of above-mentioned 4 aspects
The probability distribution of each characteristic point of image and each characteristic point by machine learning each image obtained logged in
Between data across Euclidean space at a distance from probability space, and using this distance as foundation, according to above-mentioned mould
The definition that the paste probability of happening is estimated, finds out fuzzy probability measurement value, uses this scale as foundation, with other similar characteristic points
Confrontation study is carried out, to fight learning outcome, closest image is final recognition result, to realize the reason to image
Solution.
It is i.e. described to be identified according to the feature of structuring, refer in image recognition, each feature of sample image
The probability distribution data of the structured message of point and the structured message by machine learning each image obtained, are asked
Solution is estimated at a distance from probability space, and by the fuzzy event probability of foundation of this distance across Euclidean space, uses
This scale carries out confrontation study as foundation between the logon data of each image, obtains a closest data,
Image corresponding to the data is exactly recognition result.
Here, consider that a part of image enters the visual field of image recognition and can be obtained by correct recognition result, need
Each characteristic point i, i=1,2 ... k of sample image, be the same as the listed g picture number after each study for shining sum
According to each of image data the comparison combined entirely of h characteristic point, finally obtain a fuzzy event probability scale
Maximum image is as recognition result.
Fig. 9 is the definition schematic diagram for the distance for passing through the different spaces including probability space.
As shown in Figure 9: 901 be an Euclidean space for covering probability space, and 902 be one of probability space
The central point w of probability distributionj, 903 be first scale M of more probability scales of a probability distribution of probability space1, 904 are
Second scale M of more probability scales of one probability distribution of probability space2, 905 be a probability distribution of probability space
More probability scales third scale M3, 909 be the point v of Euclidean space onej, seek vjTo wjDistance.
In realistic model identification, each characteristic value will under different conditions, by plural study,
Learning outcome forms a probability distribution, and a feature vector is made of n characteristic value, so needing to calculate between vector
Distance, here, it is w that being located in set W, which has a point 902,jHaving a point 909 in ∈ W, set V is vj∈ V, in jth=1,
The scale spacing of the scale between 902 and 903 is set in 2 ..., n element of vector as D1j=M1j, belong to wjProbability distribution it is general
Rate value is P1j (wj), the scale span between 903 and 904 is D2j=M2j-M1j, belong to wjProbability distribution probability value be P2j (wj), the scale span between 904 and 905 is D3j=M3j-M2jBelong to wjProbability distribution probability value be P3j (wj), can from Fig. 9
To find out by vjTo wjIt has passed through and belong to wjProbability distribution 3 scale zones, therefore mj (wj)=3, then between 909 to 902
Probability space distance are as follows:
[formula 3]
Here it enables,
Δj (wj)For by vjTo wjDirection, after data enter probability space from Euclidean space, in probability space
The error the distance between at a distance from actual probabilities space of distance represented by Euclidean space, with adjustment Δj (wj)Method is just
Can Euclidean space distance is unified at a distance from probability space, solve machine learning data pass through it is tight in two spaces
The definition of the distance relation of lattice.
Summarize a kind of preparation method for passing through Euclidean space at a distance from probability space that Fig. 9 is proposed, feature
It is at least the presence of this probability space, at a region for passing through probability space, the section in Euclidean space
Probability metrics are related with the probability value in the region passed through.
Above-mentioned Euclidean space extends to: including Manhattan space (Manhattan Space)；Chebyshev is empty
Between (Chebyshev Space)；Minkowski sky (Minkowski Space)；Mahalanobis space (Mahalanobis
Space)；One of included angle cosine space (Cosine Space).
Further more, the distance scale of probability space be it is related with the probability distribution value when probability space passes through, passed through, be
It has to directive, is unsatisfactory for the Symmetry Condition of common distance scale.Such as it is investigating by vjTo wjApart from when,
The distance of probability space is and vjThe position at place reaches wjPosition during, the variation of the probability for the probability distribution passed through
Process is related, this probability distribution is the w in final position to be reachedjThe probability change procedure of probability distribution, and and starting point
Position vjIt is unrelated, even if vjBe also a probability distribution, also with vjProbability distribution it is unrelated.
It needs to consider in automatic Pilot " saving oil consumption ", " fast running ", " safety traffic ", " traveling interval " is " comfortable
The Optimization control problem of traveling " and several aspects of " parking position " etc., therefore the key problem of automatic Pilot is more
The optimally-controlled machine learning model of purpose should be the problem of being concerned.
Figure 10 is the schematic diagram of the optimally-controlled machine learning model of more purposes.
It is as shown in Figure 10: " saving oil consumption " in automatic Pilot, " fast running ", " safety traffic ", " traveling interval ",
The problem of " comfortable traveling " and " parking position " etc. are different spaces, solves the optimally-controlled machine learning model of more purposes
Building, need to solve two problems, first is that the objective function of more purposes of different spaces is mapped to the same space, two
That should be flexible between more purpose objective functions, the optimized intersection point of hard straight be difficult it is existing.
Present patent application solves the problems, such as that the objective function by more purposes of different spaces is mapped to the same space, proposes
By each objective function according to preferably 100 it is the worst be 0 by Membership Function carry out fixed pattern.
Such as save the Membership Function of energy consumption:
It is excessive that parameter related with oil consumption is saved has acceleration to be not easy, and the drive speed the slow more saves oil consumption, slows down and stops
Distance the longer the better, and the strength of braking and time are the smaller the better etc., the method for establishing such Membership Function
It is: sets pickup value as ac, speed sp, deceleration stop distance is ds, and strength and the time parameter of braking are that bt is then closed
In the fixed pattern for saving oil consumption are as follows:
[formula 4]
Here, w11, w12, w13And w14For the weight for adjusting each target letter interaction, ε is braking limitation parameter.
Parameter related with fast running FT is that running speed is sp, then about the fixed pattern of fast running are as follows:
[formula 5]
Here, w21For the weight for adjusting each target letter interaction.
Parameter related with safe driving has car speed sp unsuitable too fast, and fare change number lc is unsuitable in the unit time
Excessively, the number ot that overtakes other vehicles in the unit time should not be waited excessively, then about the fixed pattern of safe driving are as follows:
[formula 6]
Here, w31, w32And w33For the weight for adjusting each target letter interaction.
And the related parameter wd of vehicle headway be it is related with running distance, then about the fixed pattern of vehicle headway are as follows:
[formula 7]
Here, w41For the weight of adjustment target letter interaction.
When preferably need to set a parking coefficient δ with stopping distance TD, has with parking coefficient, it is fixed here
Adopted stopping distance is to wish to stop to the distance actually stopped, therefore stopping distance parameter is sc
[formula 8]
Here, w51For the weight of adjustment target letter interaction.
It is that ac cannot be excessive that parameter related with comfortable traveling CD, which has acceleration value, and the strength and time bt of braking are smaller more
It is good.
[formula 9]
Here, w61And w62For the weight for adjusting each target letter interaction.
The fixed pattern of Membership Function as described above is not strict mathematical derivation, but according to artificial
The formula that experience is pieced together out, such as the fixed pattern of the speed in relation to energy saving, are exactly with w12When=150,30 kilometers per hour
Speed it is relatively good, fuzzy value 100, per hour 60 kilometers when be 50.Therefore above-mentioned fixed pattern method can have various,
Either with which kind of method, or according to which it is theoretical carry out fixed pattern, by the objective function of different spaces be mapped to one it is unified
In space, the evaluation and control that carry out more purposes are the scopes of present invention content Yu technology boundary.
The machine learning that multiple driving is carried out according to outstanding driver, by " saving energy consumption ", " fast running ", " safety
Traveling ", " vehicle headway ", the numerical value that the fixed pattern of " stopping distance " and " comfortable traveling " etc. is reflected form a probability distribution
Function passes through corresponding unsupervised machine learning ML1Generate the probability distribution data of a high-order maximum likelihood, this data
It is stored in database D B1。
Autonomous driving vehicle driving has several driving conditions, such as:
State 1: automobile moves forward, and automobile is all around all without any barrier, or colleague's automobile.
State 2: automobile moves forward, and there is vehicle in left back face.
State 3: automobile moves forward, and left front has vehicle.
State 4: automobile moves forward, and there is vehicle on the left side.
State 5: automobile moves forward, and right back has vehicle.
State 6: automobile moves forward, and there is vehicle in right front.
State 7: automobile moves forward, and there is vehicle on the right side.
State 8: automobile moves forward, behind have vehicle.
State 9: automobile moves forward, and there is vehicle in front.
… …
State n: ...
The result of each state, the machine learning of corresponding more purpose controls is different, so unsupervised engineering
Habit should first determine driving condition before study, each more purpose study corresponds to this state, that is to say, that automatic Pilot
The tasks of more purpose Optimal Control machine learning be by the driving result of excellent driver by machine learning, according to each
Driving condition, the corresponding optimally-controlled fuzzy value of purpose more than one group, probability distribution value obtained.Here it is the more mesh of automatic Pilot
Optimally-controlled " Knowledge Acquirement " method.
Figure 11 is the schematic diagram that excellent driver " Knowledge Acquirement " is realized by machine learning.
It is as shown in figure 11: by machine learning obtain excellent driver driving knowledge, realize " Knowledge Acquirement " be by
According to each state of automatic Pilot, each shape that in the plural number time driving procedure of excellent driver, will be obtained by machine learning
Under state, the probability distribution of each objective function, each multiple target corresponding to from state 1 to state n is optimally-controlled respectively
Machine learning result.
Wherein (a) is plural the carried out engineering of driving for saving the excellent driver that oil consumption objective function is passed through
It is after habit as a result, this result reflect save oil consumption probability distribution.
The plural number time that wherein (b) is the excellent driver that fast running objective function is passed through drives carried out engineering
It is after habit as a result, this result reflects the probability distribution of fast running.
The plural number time that wherein (c) is the excellent driver that safety traffic objective function is passed through drives carried out engineering
It is after habit as a result, this result reflects the probability distribution of safety traffic.
The plural number time that wherein (d) is the excellent driver that vehicular gap objective function is passed through drives carried out engineering
It is after habit as a result, this result reflects the probability distribution of vehicular gap.
The plural number time that wherein (e) is the excellent driver that comfortable traveling target function is passed through drives carried out engineering
It is after habit as a result, this result reflects the probability distribution comfortably travelled.
Wherein (f) is plural the carried out machine of driving for saving the excellent driver that parking position objective function is passed through
It is after study as a result, this result reflects the probability distribution of parking position.
Probability disperse it is bigger explanation for this state this objective function control requirement it is not stringent, otherwise probability
The dispersion of distribution is smaller, illustrates that the requirement of this objective function for this state is higher.
Automatic Pilot is also more than this 9 states, and the institute that this method covers automatic Pilot is stateful, until state n, target
Also not only from (a) to (f) this 6, this method covers all objective functions of automatic Pilot to function.
During normal automatic Pilot, a plurality of target letters of purpose more than one group will necessarily be generated for a state
Numerical value, each target function value are delivered to the sensing layer (P of Figure 101) each node (N1) on, it will be corresponding to above-mentioned state
The machine learning of each objective function is as a result, each of each objective function represented by (a)-(f) in i.e. Figure 11 passes through figure
10 machine learning (ML1) after obtained probability distribution, refer again to Fig. 9 across Euclidean space at a distance from probability space
Formula, carries out the calculating of the distance at the probability distribution center for a little arriving probability space of Euclidean space, and result is sent to figure
10 nervous layer (P2) each node (N2) on.
Machine learning (the ML of Figure 102) to nervous layer (P2) on each node (N2) on carry out maximum probability scale from
The machine learning of tissue judges the practical automatic Pilot of each objective function on the basis of the probability scale of final learning outcome
Result in, which the objective function reconciled of needs with most probable value is, while also very important person be setting one is adjusted
The threshold values ζ of optimum range, judges machine learning (ML2) the distance value of learning outcome whether be above this threshold values, if
Objective function beyond threshold values is singular, and just this objective function is adjusted, if the objective function beyond threshold values is
It is a plurality of, it is just mediated for the maximum objective function of distance value, a plurality of targets beyond threshold values can also be directed to simultaneously
Function is mediated, the direct information cortex (P to be adjusted3) node (N3) on.
Cortex (P from Figure 103) node (N3) conciliation information value, be to be provided by an adjusting unit value, can
It include that linear system controls to import, Control of Nonlinear Systems, optimum control, STOCHASTIC CONTROL, self adaptive control, fuzzy control,
Qualitative contrlol, PREDICTIVE CONTROL, one of control methods such as current events Multimode Control carry out the control of braking mechanism or accelerator mechanism
System.
But the strength of braking mechanism brake to be administered, speed related with the speed of automobile, different to be adjusted
The strength of braking mechanism be different, traditional self adaptive control will repeatedly carry out braking mechanism adaptive at the beginning every time
The adjusting answered occupies some times, declines the emergency character of automobile, using brake adjusting self adaptive control parameter as
Before every secondary control braking mechanism, the phase for adaptive machine learning of braking first is called according to car speed for the data of machine learning
Answer data, so that it may quickly the numerical value for wanting brake adjustment be made to reach requirement.
Throttle Opening Control is also in this way, according to the difference of speed, and the amount for adjusting accelerator mechanism is also different, by throttle adjustment
Self adaptive control data of the parameter as machine learning, throttle adjustment is first called according to car speed before every secondary control
The corresponding data of adaptive machine learning, so that it may quickly the numerical value of throttle to be adjusted be made to reach requirement.Above-mentioned operation is all
It is " more purpose machine learning automatic adjusument " center "s are realized.
Above-mentioned every adjusting is primary, also to return to machine learning (ML1) and (ML2) learning outcome it is whether effective? so repeatedly
Study is adjusted repeatedly, it is likely that machine learning several times and adjusting is needed to can be only achieved the driving effect of excellent driver.
" more purpose machine learning automatic adjusument " center "s simultaneously can also be for fare identification as a result, and GPS
The represented traffic route of figure, controls wheels.
Claims (3)
1. a kind of image-recognizing method that can reach image understanding level, it is characterised in that:
(1) non-structured image data is carried out to the characterization of structuring；
(2) it is identified according to the feature of structuring.
2. a kind of image-recognizing method that can reach image understanding level according to claim 1, it is characterised in that: described
The characterization that non-structured image data is carried out to structuring refers to through probability density in machine learning acquisition image most
Big a plurality of characteristic points；Each characteristic point will carrying include characteristic point maximum probability scale, maximum probability scale it is general
Rate distribution；Or the probability distribution that characteristic point is k-th；Or k-th of characteristic point be to the vector value of+1 characteristic point of kth, and general
Rate is distributed in the information of interior at least one structuring.
3. a kind of image-recognizing method that can reach image understanding level according to claim 1, it is characterised in that: described
It is identified, is referred in image recognition according to the feature of structuring, the structured message of each characteristic point of sample image
With the probability distribution data of the structured message by machine learning each image obtained, carry out including that solution is several across Europe
Reed space is at a distance from probability space；And estimate by the fuzzy event probability of foundation of this distance；At least one scale
As foundation, confrontation study is carried out between the logon data of each image, obtains a closest data, the data institute
Corresponding image is exactly recognition result, to realize the understanding to image.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710905332.8A CN109522778A (en) | 2017-09-20 | 2017-09-20 | A kind of image-recognizing method can reach image understanding level |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710905332.8A CN109522778A (en) | 2017-09-20 | 2017-09-20 | A kind of image-recognizing method can reach image understanding level |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109522778A true CN109522778A (en) | 2019-03-26 |
Family
ID=65767860
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710905332.8A Pending CN109522778A (en) | 2017-09-20 | 2017-09-20 | A kind of image-recognizing method can reach image understanding level |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109522778A (en) |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2004126774A (en) * | 2002-09-30 | 2004-04-22 | Toshiba Corp | Pattern recognition device |
| JP2005208850A (en) * | 2004-01-21 | 2005-08-04 | Nippon Hoso Kyokai <Nhk> | Face image recognition device and face image recognition program |
| US20150065803A1 (en) * | 2013-09-05 | 2015-03-05 | Erik Scott DOUGLAS | Apparatuses and methods for mobile imaging and analysis |
| US20160171328A1 (en) * | 2014-12-11 | 2016-06-16 | Xerox Corporation | Methods and systems for vehicle tag number recognition |
| WO2017032243A1 (en) * | 2015-08-26 | 2017-03-02 | 阿里巴巴集团控股有限公司 | Image feature extraction method, apparatus, terminal device, and system |
-
2017
- 2017-09-20 CN CN201710905332.8A patent/CN109522778A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2004126774A (en) * | 2002-09-30 | 2004-04-22 | Toshiba Corp | Pattern recognition device |
| JP2005208850A (en) * | 2004-01-21 | 2005-08-04 | Nippon Hoso Kyokai <Nhk> | Face image recognition device and face image recognition program |
| US20150065803A1 (en) * | 2013-09-05 | 2015-03-05 | Erik Scott DOUGLAS | Apparatuses and methods for mobile imaging and analysis |
| US20160171328A1 (en) * | 2014-12-11 | 2016-06-16 | Xerox Corporation | Methods and systems for vehicle tag number recognition |
| WO2017032243A1 (en) * | 2015-08-26 | 2017-03-02 | 阿里巴巴集团控股有限公司 | Image feature extraction method, apparatus, terminal device, and system |
Non-Patent Citations (2)
| Title |
|---|
| 杨世强等: "Log Gaussian Cox场手部指节的图像偏移特征学习与识别", 《农业机械学报》 * |
| 顾泽苍: "人工智能技术深度剖析", 《机器人技术与应用》 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US11720099B2 (en) | Control method of automatic driving imported “smart gains” model, device and program | |
| CN107229973B (en) | Method and device for generating strategy network model for automatic vehicle driving | |
| CN110796856B (en) | Vehicle lane change intention prediction method and training method of lane change intention prediction network | |
| Díaz-Álvarez et al. | Modelling the human lane-change execution behaviour through multilayer perceptrons and convolutional neural networks | |
| US11354915B2 (en) | Image extraction method importing into SDL model | |
| CN109726804B (en) | Intelligent vehicle driving behavior personification decision-making method based on driving prediction field and BP neural network | |
| CN110103956A (en) | Automatic overtaking track planning method for unmanned vehicle | |
| CN105489069B (en) | A kind of low altitude airspace navigation aircraft collision detection method based on SVM | |
| CN108944946A (en) | A kind of bend risk based on truck traffic estimates calculation method | |
| CN114170789B (en) | Intelligent network link lane change decision modeling method based on space-time diagram neural network | |
| CN110764507A (en) | Artificial intelligence automatic driving system for reinforcement learning and information fusion | |
| US11550327B2 (en) | Composition method of automatic driving machine consciousness model | |
| CN110182217A (en) | A kind of traveling task complexity quantitative estimation method towards complicated scene of overtaking other vehicles | |
| CN114312830A (en) | Intelligent vehicle coupling decision model and method considering dangerous driving conditions | |
| Yu et al. | Autonomous overtaking decision making of driverless bus based on deep Q-learning method | |
| CN109455178A (en) | A kind of road vehicles traveling active control system and method based on binocular vision | |
| CN115257819A (en) | Decision-making method for safe driving of large-scale commercial vehicle in urban low-speed environment | |
| CN113901718A (en) | Deep reinforcement learning-based driving collision avoidance optimization method in following state | |
| CN109543497A (en) | A kind of construction method of more purposes control machine learning model suitable for automatic Pilot | |
| Li | A hierarchical autonomous driving framework combining reinforcement learning and imitation learning | |
| CN111046897A (en) | Method for defining fuzzy event probability measure spanning different spaces | |
| Cao et al. | Fuzzy social force model for pedestrian evacuation under view-limited condition | |
| CN109522778A (en) | A kind of image-recognizing method can reach image understanding level | |
| CN109670597A (en) | A kind of more purpose control methods of the machine learning of automatic Pilot | |
| CN111126612A (en) | Automatic machine learning composition 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 |