CN107305636A - Target identification method, Target Identification Unit, terminal device and target identification system - Google Patents
Target identification method, Target Identification Unit, terminal device and target identification system Download PDFInfo
- Publication number
- CN107305636A CN107305636A CN201610255052.2A CN201610255052A CN107305636A CN 107305636 A CN107305636 A CN 107305636A CN 201610255052 A CN201610255052 A CN 201610255052A CN 107305636 A CN107305636 A CN 107305636A
- Authority
- CN
- China
- Prior art keywords
- grader
- target
- deep learning
- training sample
- storage
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/243—Classification techniques relating to the number of classes
Abstract
The invention provides the target identification method based on on-line automatic deep learning, Target Identification Unit, terminal device and target identification system.Wherein, this method includes：Real-time image acquisition data；Using the first grader of storage, target identification is carried out to the view data collected, to generate the view data with class label, class label includes target classification and target posterior probability；In the case where target posterior probability meets predetermined storage condition, according to target classification, training sample is stored as；And in the case where predetermined entry condition is satisfied, start online deep learning processing, to obtain the 3rd grader, and the first grader of storage is updated with the 3rd grader obtained.By using the above method of the present invention, the target identification of high-accuracy can be realized.
Description
Technical field
The present invention relates to target identification technology, more particularly to target identification method based on on-line automatic deep learning, target
Identifying device, terminal device and target identification system.
Background technology
Target identification technology all has wide practical use in fields such as video monitoring, robot, intelligent transportation.But by
Need to be related to the calculating of mass data and analysis in target identification, the interference of the environmental factor such as additional smooth visual angle, tradition is known
Other algorithm can not extract the preferred feature of image, cause discrimination limited.
A kind of method of target identification can use traditional off-line learning method.But, in traditional off-line learning method
In, it is only applicable to the specific objective in specific environment using the disaggregated model of offline classifier training.When in video image
Target sizes and when environment etc. and serious inconsistent training sample, it is impossible to carry out accurate Activity recognition, it is portable not
It is good.Although this defect can be made up to a certain extent by the method for large sample training grader.However, this kind of calculate
Method, which generally requires to set up, includes the big-sample data storehouse of different condition, different scenes, while needing to the data in database
Manual markings are carried out, so as to bring huge workload and inconvenience.
Another method of target identification is to use existing on-line study method.Compared to traditional off-line learning method,
Existing on-line study method can not only ensure the correctness of model in the way of model modification, moreover it is possible to save substantial amounts of storage
Space.On-line study method can be weakened in learning process significantly marks this cumbersome step by hand, is not usually required to open
Data are moved, or need to only start low volume data, i.e. only need one less sample set of manual mark to come for the first of grader
Begin to train.Then, the grader can constantly obtain new samples when performing classification task, so that lasting self
Training and improvement, to improve nicety of grading.Although subproblem can be solved using existing on-line study method,
It is that existing on-line study method needs to carry out the new samples of acquisition the automatic marking of classification, otherwise can not realizes identification system
The intellectuality of system, and the correctness of training sample mark determines the validity of whole training process.Thus, initially
The low grader of precision is not suitable for response extremely rapidly and required precision is reliably applied.
Notification number is in CN102915453B Chinese patent disclosed in 1 day July in 2015, it is proposed that a kind of real-time
Feed back the vehicle identification method updated, including off-line learning processing, Real time identification processing and on-line study processing.First,
1~K frame the pictures collected in real-time recognition process are entered in the offline strong classifier obtained in being handled using off-line learning
Row classification, is identified result.Then, in on-line study processing, according to obtained recognition result intercepted samples, profit
Vehicle identification is carried out with online strong classifier, target is identified.On-line study processing is constantly carried out to online strong classifier
Update.But, in the patent document, off-line learning is selected dependent on feature extraction and grader.Feature Selection needs
Make special design for different application, and the grader AdaBoost used again relies on Weak Classifier initial selected.And
And, the training sample of on-line study is derived only from 1~K frame pictures, and data type can not simulate more scenes.
The content of the invention
The present invention is directed to the shortcomings of the prior art, it is proposed that the target identification based on on-line automatic deep learning
Method, Target Identification Unit, terminal device and target identification system.Target identification method, the target identification dress of the present invention
Put, terminal device and target identification system, realize the target identification of high-accuracy.
According to an aspect of the invention, there is provided a kind of target identification method, the target identification method is based on online certainly
Dynamic deep learning, methods described includes：
Real-time image acquisition data；
Using the first grader of storage, target identification is carried out to the view data collected, class label is carried to generate
View data, the class label include target classification and target posterior probability；
In the case where the target posterior probability meets predetermined storage condition, according to the target classification, carried described
The view data of class label is stored as training sample；And
In the case where predetermined entry condition is satisfied, start online deep learning processing, the online deep learning processing
Network model is built including the first grader based on the storage, and whole training samples of storage are input to structure
Network model in carry out deep learning processing, to obtain the 3rd grader, and with the 3rd grader of acquisition come
Update the first grader of the storage.
Further, the preliminary classification device that the first grader of the storage is used be by class label by manually marking
The training sample of note carries out the offline deep learning grader that deep learning is obtained.
Further, the target identification method further comprises：
Periodically by whole training samples of the storage via network transmission to far-end server, to carry out offline deep learning
Processing；And
Received via the network from the far-end server and the offline deep learning processing is carried out by the far-end server
The second grader obtained, and when receiving second grader, with second grader received come
Update the first grader of the storage.
Further, before target identification is carried out, methods described further comprises entering the view data collected
Row image preprocessing, to improve image definition, and extracts area-of-interest figure from the view data collected
As data；And
When carrying out target identification, mesh is carried out to the area-of-interest view data in the view data collected
Mark is other.
Further, before deep learning is carried out, the online deep learning processing further comprises to the storage
Whole training samples carry out image preprocessing, to improve image definition.
Further, when the quantity for the training sample for belonging to the other storage of the target class not yet reaches predetermined quantity
When, the predetermined storage condition uses the first storage condition, and first storage condition is that the target posterior probability is more than
Equal to predetermined threshold；
It is described when belonging to the quantity of training sample of the other storage of the target class and having reached the predetermined quantity
Predetermined storage condition uses the second storage condition, and second storage condition is described in the target posterior probability is more than or equal to
Predetermined threshold, and the target posterior probability is minimum more than the training sample for belonging to the other storage of the target class
Target posterior probability；And
When the target posterior probability meets second storage condition, the target identification method further comprises, deletes
Belong to the other training sample with minimum target posterior probability of the target class except oldest stored.
Further, the predetermined entry condition is to be in idle condition using the equipment of the target identification method, and
The quantity of the training sample of the other storage of each target class reaches the predetermined quantity.
Further, the predetermined quantity is sum × 10/ mesh to be trained to of the network parameter number of the network model
Mark the sum of classification.
According to another aspect of the present invention there is provided a kind of Target Identification Unit, the Target Identification Unit is based on online
Automatic depth learns, and the Target Identification Unit includes：Real-time detection apparatus, first storage device and on-line study dress
Put, wherein
The real-time detection apparatus includes：
Image data acquiring unit, described image data acquisition unit real-time image acquisition data；
Object-recognition unit, the object-recognition unit utilizes the first grader being stored in the first storage device,
Target identification is carried out to the view data collected, to generate the view data with class label, the class label bag
Classification containing target and target posterior probability；
The first storage device includes：
First grader memory cell, the first grader memory cell stores first grader；And
First training sample memory cell, the first training sample memory cell meets predetermined in the target posterior probability
In the case of storage condition, according to the target classification, the view data with class label is stored as to train sample
This；
The on-line study device includes：
On-line study start unit, the on-line study start unit starts in the case where predetermined entry condition is satisfied
Online deep learning processing；And
Online deep learning unit, it is described when the on-line study start unit starts the online deep learning processing
Online deep learning unit builds network mould based on first grader being stored in the first grader memory cell
Type, and the whole training samples that will be stored in the first training sample memory cell are input to the network model of structure
Middle progress deep learning processing, to obtain the 3rd grader, and is stored in the 3rd grader obtained to update
First grader in the first grader memory cell.
Further, the preliminary classification that first grader being stored in the first grader memory cell is used
Device is by carrying out the offline deep learning classification that deep learning is obtained by the training sample manually marked to class label
Device.
Further, the Target Identification Unit further comprises first data transmission device, wherein,
All training that the first data transmission device periodically will be stored in the first training sample memory cell
Sample via network transmission to far-end server, to carry out offline deep learning processing；And
The first data transmission device is received from the far-end server via the network and carried out by the far-end server
The offline deep learning handles the second obtained grader, and when receiving second grader, with reception
To second grader update first grader being stored in the first grader memory cell.
Further, the real-time detection apparatus further comprises the first image pre-processing unit, and described first image is located in advance
Reason unit carries out image preprocessing to the view data collected, to improve image definition, and is collected from described
View data in extract area-of-interest view data；And
The object-recognition unit carries out target to the area-of-interest view data in the view data collected
Identification.
Further, the on-line study device further comprises the 3rd image pre-processing unit, and the 3rd image is located in advance
Unit is managed before the online deep learning unit carries out deep learning processing, is deposited to being stored in first training sample
Whole training samples in storage unit carry out image preprocessing, to improve image definition.
Further, the other training sample of the target class is belonged to when being stored in the first training sample memory cell
Quantity when not yet reaching predetermined quantity, the predetermined storage condition uses the first storage condition, first storage condition
It is more than or equal to predetermined threshold for the target posterior probability；
When the quantity for belonging to the other training sample of the target class being stored in the first training sample memory cell
When reaching the predetermined quantity, the predetermined storage condition uses the second storage condition, and second storage condition is described
Target posterior probability is more than or equal to the predetermined threshold, and the target posterior probability is trained more than being stored in described first
The minimum target posterior probability for belonging to the other training sample of the target class in sample storage unit；And
When the target posterior probability meets second storage condition, the first training sample memory cell is deleted most
Early storage into the first training sample memory cell to belong to the target class other with minimum target posterior probability
Training sample.
Further, the predetermined entry condition is that the equipment with the Target Identification Unit is in idle condition, and
The quantity for the other training sample of each target class being stored in the first training sample memory cell reaches described predetermined
Quantity.
Further, the predetermined quantity is sum × 10/ mesh to be trained to of the network parameter number of the network model
Mark the sum of classification.
According to a further aspect of the invention there is provided a kind of terminal installation, the terminal installation includes the target of the present invention
Identifying device.
According to a further aspect of the invention there is provided a kind of target identification system, the system include far-end server,
And the multiple terminal installations of the invention being connected via network with the far-end server, wherein,
Each terminal installation is using the first grader being stored in the first grader memory cell to collecting in real time
View data carries out target identification, and the view data for meeting predetermined storage condition is stored as into training sample, and with passing through
The 3rd classification that deep learning processing is obtained is carried out to the whole training samples being stored in the first training sample memory cell
Device or with the second grader received from the far-end server, to update, to be stored in the first grader storage single
First grader in member；
Each terminal installation includes first data transmission device, and the first data transmission device is periodically via the net
Whole training samples of storage are transferred to the far-end server by network；
The far-end server includes the second storage device, off-line learning device and the second data transmission device, wherein,
Second storage device includes storing the second grader memory cell of second grader, and storage classification
Second training sample of the training sample that label is received by the training sample that manually marks and from the multiple terminal installation is deposited
Storage unit,
The off-line learning device includes offline deep learning unit, and the offline deep learning unit is described based on being stored in
Second grader in second grader memory cell builds network model, will be stored in second training sample and deposits
Whole training samples in storage unit are input to progress deep learning processing in the network model of structure, to obtain new second
Grader, and be stored in the second new grader obtained to update in the second grader memory cell
Second grader, and
Second grader after renewal is transferred to each end by second data transmission device via the network
End device.
Further, first grader being stored in the first grader memory cell is with being stored in described second
Second grader in grader memory cell uses same preliminary classification device, the preliminary classification device be it is described from
Line deep learning unit by class label by the training sample that manually marks carry out that deep learning processing obtained it is offline
Deep learning grader.
Further, the off-line learning device further comprises the second image pre-processing unit, and second image is located in advance
Unit is managed before the offline deep learning unit carries out deep learning processing, is deposited to being stored in second training sample
Whole training samples in storage unit carry out image preprocessing, to improve image definition.
, can by using target identification method, Target Identification Unit, terminal device and the target identification system of the present invention
With the cosmetic variation and scene changes of identification target, the view data collected in real time is made full use of, after by target
Testing probability ensures to carry out on-line study on the premise of the reliability of training sample, so as to realize the target identification of high-accuracy.
In addition, being identified, further ensure that to target as preliminary classification device using offline deep learning grader
Higher recognition accuracy.
Brief description of the drawings
Fig. 1 shows the schematic diagram of the configuration structure of target identification system according to an embodiment of the invention；
Fig. 2 shows the flow chart of existing deep learning method；
Fig. 3 shows the schematic diagram of the configuration structure of Target Identification Unit according to an embodiment of the invention；And
Fig. 4 shows the flow chart of target identification method according to an embodiment of the invention.
Embodiment
Describe below with reference to the accompanying drawings according to various embodiments of the present invention.
Fig. 1 shows the schematic diagram of the configuration of target identification system 1 according to an embodiment of the invention.
As shown in figure 1, target identification system 1 includes far-end server 10 and multiple terminal installations 30.Each terminal installation
30 are connected via network 20 with far-end server 10, can mutually to carry out data exchange.
Each terminal installation 30 includes Target Identification Unit 300.The Target Identification Unit 300 of each terminal installation 30 is utilized
First grader of storage carries out target identification to the view data collected in real time, will meet the image of predetermined storage condition
Data storage is training sample, and carries out deep learning processing is obtained with by whole training samples to storage
Three graders or with the second grader received from far-end server 10, to update the first grader of storage.
Whole training samples of storage are periodically transferred to by the Target Identification Unit 300 of each terminal installation 30 via network 30
Far-end server 10.
The concrete configuration structure and processing procedure of Target Identification Unit 300 will be explained later below.Below
Fig. 1 and Fig. 2 will be combined the concrete configuration structure of far-end server 10 and place according to an embodiment of the invention is described in detail
Reason process.
As shown in figure 1, far-end server 10 includes the second data transmission device 110, the second storage device 120 and offline
Learning device 130.Second storage device 120 includes the second grader memory cell 121 and the second training sample memory cell
122.Off-line learning device 130 includes the second image pre-processing unit 131 and offline deep learning unit 132.
Second data transmission device 110 receives each terminal installation 30 and periodically transmits the training sample of coming via network 20, and
And by the training sample received storage into the second training sample memory cell 122.In addition, in the second storage device 120
The second grader memory cell 121 in after the second grader for storing is updated, the second data transmission device 110 is via net
The second grader after renewal is transferred to each terminal installation 30 by network 30.
Second grader memory cell 121 is used to store the second grader.Second training sample memory cell 122 be used for according to
Classification stores class label that class label receives by the training sample that manually marks and from multiple terminal installations 30 by target
The training sample of identifying device 300.
Second image pre-processing unit 131 is stored in the second training sample memory cell in offline 132 pairs of deep learning unit
Whole training samples in 122 are carried out before deep learning, to the whole being stored in the second training sample memory cell 122
Training sample carries out image preprocessing, noise reduction process or other processing is such as carried out, to improve image definition.
Offline deep learning unit 132 uses existing deep learning method, and can carry out two kinds of processing, one kind processing
It is to be handled based on zero offline deep learning, and another processing is at the offline deep learning fed back based on on-line study
Reason.
It is described below and is handled by what offline deep learning unit 132 was carried out based on zero offline deep learning.
When target identification system 1 is in original state, i.e. far-end server 10 is also not received by from multiple terminals
The training sample of device 30, and the training sample being stored in the second training sample memory cell 122 of far-end server 10
Originally it is class label by the training sample that manually marks (hereinafter, for convenience of description, by class label by artificial
The training sample of mark is referred to as " full supervised training sample ") when, offline 132 pairs of deep learning unit is stored in the second instruction
Practice whole full supervised training samples in sample storage unit 122 to carry out based on the processing of zero offline deep learning, to obtain
Preliminary classification device, the preliminary classification device of acquisition is offline deep learning grader.When offline deep learning unit 132 is by entering
When row obtains preliminary classification device based on the processing of zero offline deep learning, offline deep learning unit 132 is by initial point of acquisition
Class device is stored into the second grader memory cell 121, using the preliminary classification device as the second grader.Then, distal end takes
Second data transmission device 110 of business device 10 can will be stored in the second grader memory cell 121 via network 20
The preliminary classification device of second grader is transferred to each terminal installation 30, to cause each terminal installation 30 and far-end server
10 have same preliminary classification device.
Below in conjunction with the flow chart of existing deep learning method shown in Fig. 2, to describe offline deep learning unit
132 obtain the process of preliminary classification device by carrying out based on the processing of zero offline deep learning.
First, as shown in Fig. 2 in step s 201, offline deep learning unit 132 is by by the model of network model
The parameter value of parameter is initialized as unified value to build network model.The number of plies of the model parameter of network model including network,
Weighted value and receptance function between every layer of nodes, node and node.In existing deep learning method,
With the difference of the receptance function used, different network models can be built.Hereinafter, will be with convolutional Neural net
The explanation of correlation is carried out exemplified by network model.
Each layer of convolutional neural networks model constructed by offline deep learning unit 132 include input layer, output layer and
Multilayer hidden layer between input layer and output layer.Multilayer hidden layer includes wave filter group layer, correcting layer, local contrast normalizing
Change layer, average pond and sub-sampling layer and maximum pond and sub-sampling layer.Wave filter group layer include convolution filter,
Activation primitive and gain can be trained.Activation primitive uses non-linear transform function sigmoid.Convolution filter uses core
Function carries out convolutional filtering.Correcting layer is simply corrected to the output result of wave filter group layer, is used and is taken absolute value
Operation.Local contrast normalization layer takes average and normalized square mean, i.e. characteristics of image to normalize to upper strata output result.
Average pondization and sub-sampling layer cause the feature extracted to have robustness to miniature deformation, use all to sampling window
Value is averaged, and obtained value is transferred to next sample level.Maximum pondization and sub-sampling layer are the features pair that realization is extracted
The consistency of translation, uses and takes average maximum to sampling window all values, and obtained value is transferred to next adopt
Sample layer.
Then, in step S202, offline deep learning unit 132 is by by the pre- of the second image pre-processing unit 131
The full supervised training sample of a part in whole full supervised training samples of processing is input to the convolution built according to classification
In neural network model, deep learning is carried out by the full supervised training sample to input, the training side of deconvolution is utilized
Method, the model parameter to convolutional neural networks model is trained, to adjust the parameter value of model parameter.Training process bag
Containing preceding to training and backward training.Forward direction training process is unsupervised learning from bottom to top, i.e., since bottom, one layer
One layer of past top layer training.In forward direction training process, in the parameter by training the model parameter for learning to obtain (n-1)th layer
After value, using n-1 layers of output as the input of n-th layer, to train n-th layer, the model ginseng of each layer is thus respectively obtained
Several parameter values.Backward training process is the top-down transmission of top-down supervised learning, i.e. training error, to model
The parameter value of parameter is finely adjusted.
Then, in step S203, offline deep learning unit 132 is by by the pre- of the second image pre-processing unit 131
The full supervised training sample of another part in whole full supervised training samples of processing is inputted as test sample according to classification
In the convolutional neural networks model being trained to model parameter, obtained preferably by multiple convolution and time sample process
Feature.
Then, in step S204, offline deep learning unit 132 counts sample attribute classification according to preferred feature
Mapping table, so as to obtain the ginseng comprising sample attribute classification mapping table and the model parameter being trained to
The offline deep learning grader of numerical value, the offline deep learning grader is preliminary classification device.
The offline deep learning processing fed back based on on-line study carried out by offline deep learning unit 132 is described below.
When target identification system 1 is in non-initial state, i.e. far-end server 10 is had been received by from multiple terminals
The training sample of device 30, and the training sample being stored in the second training sample memory cell 122 of far-end server 10
This includes class label that class label receives by the training sample that manually marks and from multiple terminal installations 30 by mesh
Marking the training sample of the mark of identifying device 300 (hereinafter, for convenience of description, will both include class label by manually marking
The training sample of note, includes what the class label that is received from multiple terminal installations 30 was marked by Target Identification Unit 300 again
The sample set of training sample is referred to as " semi-supervised training sample ") when, offline deep learning unit 132 is based on being stored in
The second grader in second grader memory cell 121, to the whole being stored in the second training sample memory cell 122
Semi-supervised training sample carry out the offline deep learning processing fed back based on on-line study, to obtain new second classification
Device.When offline deep learning unit 132 by the offline deep learning processing fed back based on on-line study obtains new the
During two graders, offline deep learning unit 132 stores the second new grader of acquisition to the second grader memory cell
In 121, to update second point be stored in the second grader memory cell 121 with the second new grader of acquisition
Class device.
The offline deep learning that offline deep learning unit 132 fed back based on on-line study handles to obtain new second point
The process of class device, carries out obtaining preliminary classification device based on the processing of zero offline deep learning with offline deep learning unit 132
Process it is roughly the same, be all using existing deep learning method as shown in Figure 2.Differ only in, be based on
During the offline deep learning processing of on-line study feedback, offline deep learning unit 132 is by by the model parameter of network model
Parameter value be initialized as the ginseng of the model parameter that the second grader is included being stored in the second grader memory cell 121
Numerical value, i.e., the parameter value for the model parameter that last off-line learning is obtained, to build convolutional neural networks model, and
The training sample being input in the convolutional neural networks model of structure is semi-supervised training sample.The the second new classification obtained
Parameter value of the device comprising sample attribute classification mapping table and the model parameter being trained to.In order to simplify
It is bright, it is omitted here identical description.
The concrete configuration knot of Target Identification Unit 300 according to an embodiment of the invention is described below in conjunction with Fig. 3 and Fig. 4
Structure and processing procedure.Fig. 3 shows the signal of the configuration structure of Target Identification Unit 300 according to an embodiment of the invention
Figure.
As shown in figure 3, the Target Identification Unit 300 based on on-line automatic deep learning includes real-time detection apparatus 310, the
One storage device 320, on-line study device 330 and first data transmission device 340.
Real-time detection apparatus 310 includes image data acquiring unit 311 and object-recognition unit 313.In addition, such as Fig. 3
Shown, real-time detection apparatus 310 can further include the first image pre-processing unit 312.
First storage device 320 includes the first grader memory cell 322 and the first training sample memory cell 321.
First grader memory cell 322 is used to store the first grader, and the first grader maps comprising sample attribute classification to close
The parameter value of model parameter for being table and being trained to.It is stored in first point in the first grader memory cell 322
The preliminary classification device that class device is used can be stored in advance in the first grader memory cell 322 by class label
The offline deep learning grader that deep learning is obtained is carried out by the training sample that manually marks, or can also be by the
The preliminary classification device that one data transmission device 340 is received via network 20 from far-end server 10.
First training sample memory cell 321 is used to store the training that target posterior probability meets predetermined storage condition according to classification
Sample.
On-line study device 330 includes on-line study start unit 331 and online deep learning unit 333.In addition, such as
Shown in Fig. 3, on-line study device 330 can also include the 3rd image pre-processing unit 332.
Whole training samples warp that first data transmission device 340 periodically will be stored in the first training sample memory cell 321
Far-end server 30 is transferred to by network 20, to carry out offline deep learning processing.Moreover, first data transmission device
340 also receive from far-end server 10 via network 20 and carry out the processing of offline deep learning by far-end server 10 and obtained
The second grader, and when receiving the second grader, first is stored in update with the second grader received
The first grader in grader memory cell.
Fig. 4 shows the flow chart of the target identification method carried out according to an embodiment of the invention by Target Identification Unit 300.
As shown in figure 4, first, in step S401, the real-time image acquisition data of image data acquiring unit 311.Figure
As data acquisition unit 311 includes optical system and camera.Optical system has anamorphosis function, automatic focusing function etc..
Camera can be the video camera using colored CCD (charge coupled cell).
Then, in step S402, object-recognition unit 313 utilizes first point be stored in first storage device 322
Class device, target identification is carried out to the view data collected, to generate the view data with class label.Class label
Include target classification and target posterior probability.
In addition, before 313 pairs of view data collected of object-recognition unit carry out target identification, first can also be utilized
The view data that image pre-processing unit 312 is collected to image data acquiring unit 311 carries out the figure of noise reduction process etc.
As pretreatment, to improve image definition.
In addition, in order that object-recognition unit 313 can the more accurate view data to collecting carry out target identification, the
The view data that one image pre-processing unit 312 can also be collected to image data acquiring unit 311 carries out such as motion inspection
Survey method (such as optical flow method), background constructing method (such as gauss hybrid models foundation), target object candidate area are extracted (such as
DPM deform model of parts) etc. image preprocessing, to extract region of interest area image from the view data collected
Data.
Now, in step S402, object-recognition unit 313 utilizes first point be stored in first storage device 322
Class device, target identification is carried out to the area-of-interest view data in the view data that collects, and classification mark is carried to generate
The view data of label.Class label includes target classification and target posterior probability.
Object-recognition unit 313 calculates what is collected according to the sample attribute classification mapping table included in the first grader
Area-of-interest view data in view data or the view data collected corresponds respectively to the probability of each classification
Value, the sense that the classification with most probable value is defined as in the view data collected or the view data collected is emerging
The target classification of interesting region image data, and using the most probable value is as the view data collected or collects
The target posterior probability of area-of-interest view data in view data.
Sense in the recognition result of object-recognition unit 313, including the view data collected or the view data collected
The target classification and target posterior probability of interest region image data, can be output, for use in other application.Example
Such as, the recognition result of object-recognition unit 313 can carry out target following as the input of track algorithm.
Then, in step S403, the first training sample memory cell 321 in first storage device judges what is collected
Whether the target posterior probability of the area-of-interest view data in view data or the view data collected meets predetermined
Storage condition.When the target of the area-of-interest view data in the view data collected or the view data collected
When posterior probability meets predetermined storage condition (in step S403 be), the first training sample memory cell 321 is in step
In S404, according to target classification, the view data with class label is stored as training sample.When the image collected
The target posterior probability of area-of-interest view data in data or the view data collected is unsatisfactory for predetermined storage bar
(no in step S403), return to step S401 during part.
Predetermined storage condition can be set as needed.In the present embodiment, based on ensuring to be stored in the first training sample
The purpose of the reliability of training sample in memory cell, to set predetermined storage condition.
Belong to the view data collected or the picture number collected when being stored in the first training sample memory cell 321
When the quantity of the other training sample of target class of area-of-interest view data in not yet reaches predetermined quantity, using
One storage condition is used as predetermined storage condition.First storage condition is the view data collected or the picture number collected
The target posterior probability of area-of-interest view data in is more than or equal to predetermined threshold.Typically, predetermined threshold is set to
Higher, the reliability of storage to the training sample in the first training sample memory cell 321 is higher.Predetermined quantity can be set
Determine into what sum × 10/ of the network parameter number for the network model that will be built in online deep learning unit to be trained to
The other sum of target class.
Belong to the view data collected or the picture number collected when being stored in the first training sample memory cell 321
When the quantity of the other training sample of target class of area-of-interest view data in has reached predetermined quantity, using
Two storage conditions are used as predetermined storage condition.Second storage condition is the view data collected or the picture number collected
The target posterior probability of area-of-interest view data in is more than or equal to predetermined threshold, and the view data collected
Or the target posterior probability of the area-of-interest view data in the view data collected be more than be stored in the first training sample
Area-of-interest picture number in view data that belonging in this memory cell 321 collects or the view data collected
According to the other training sample of target class minimum target posterior probability.
When the target posteriority of the area-of-interest view data in the view data collected or the view data collected is general
When rate meets the second storage condition, the first training sample memory cell 321 further deletes oldest stored to the first training sample
Area-of-interest view data in view data that belonging in memory cell 321 collects or the view data collected
The other training sample with minimum target posterior probability of target class.
Then, in step S405, on-line study start unit 331 judges whether predetermined entry condition is satisfied.When
Line study start unit 331 is when judging that predetermined entry condition is satisfied (in step S405 be), and on-line study starts single
Member 331 starts online deep learning processing in step S406, to obtain the 3rd grader.When on-line study start unit
331 (no in step S405), return to step S401 when judging that predetermined entry condition is not satisfied.
In the present embodiment, predetermined entry condition is set with the terminal device 30 of Target Identification Unit 300 in sky
Not busy state, and the quantity for the other training sample of each target class being stored in the first training sample memory cell 321 reaches
To predetermined quantity.
Because the terminal device 30 with Target Identification Unit 300 can be applied to different field, accordingly, it is determined that having
The method of the equipment state of the terminal device 30 of Target Identification Unit 300 is also different.
For example, when the terminal device 30 with Target Identification Unit 300, which is applied to automatic Pilot/auxiliary, to be driven,
When i.e. Target Identification Unit 300 is installed in the vehicle as terminal device 30, Target Identification Unit 300 can pass through
Identification terminal equipment 30 (vehicle) is current, and whether long-time stop motion (having stopped working) determines (car of terminal device 30
) whether it is in idle condition.That is, whether scene where the identification of Target Identification Unit 300 moving target changes, and wraps
Include the background model for obtaining and being obtained relative to the background model of illumination variation robust with previous moment to image sequence background modeling
It is compared to identify whether occurrence scene change.So-called background modeling is exactly to extract each frame in movement destination image sequence
Without motion region part, to the part carry out mixed Gaussian background modeling, ask for image sequence present frame background model and
Background model difference value was obtained in the past to judge whether scene changes.If Target Identification Unit 300 recognizes present terminal
Equipment 30 (vehicle) is out of service, and background scene does not change, it is determined that terminal device 30 (vehicle) is in sky
Not busy state.
When the terminal device 30 with Target Identification Unit 300 is applied to video monitoring, Target Identification Unit 300 can
Currently determine whether terminal device 30 is in idle condition with the presence or absence of moving object by identification.If target identification is filled
When putting 300 identifications and there is currently no moving object, it is determined that terminal device 30 is in idle condition.
For some other applications, Target Identification Unit 300 can also be by recognizing whether current scene has begun to infrared benefit
Light come determine terminal device 30 whether be in idle condition.If Target Identification Unit 300 recognize currently have begun to it is infrared
Light filling, i.e., be currently in night, without progress Real time identification, it is determined that terminal device 30 is in idle condition.
In step S406, when on-line study start unit 331 starts online deep learning processing, the 3rd image is located in advance
Unit 332 is managed before online deep learning unit 333 carries out deep learning processing, to being stored in the storage of the first training sample
Whole training samples in unit 321 carry out image preprocessing, to improve image definition.Then, in step S407
In, online deep learning unit 333 builds network based on the first grader being stored in the first grader memory cell 321
Model, and the whole training samples that will be stored in the first training sample memory cell 321 are input to the network model of structure
Middle progress deep learning processing, to obtain the 3rd grader.
Online deep learning unit 333 is similar to offline deep learning unit 132, is also to use existing deep learning side
Method, to carry out the online deep learning processing fed back based on on-line study, to obtain the 3rd grader.
Online deep learning unit 333 carries out the online deep learning processing fed back based on on-line study, with offline deep learning
The offline deep learning that unit 132 fed back based on on-line study handles to obtain the process substantially phase of the second new grader
Together, all it is using existing deep learning method as shown in Figure 2.Differ only in, carrying out based on on-line study feedback
Online deep learning processing when, online deep learning unit 333 passes through the parameter value of the model parameter of network model is initial
Turn to the parameter value for the model parameter that the first grader being stored in the first grader memory cell 322 is included, i.e., upper one
The parameter value for the model parameter that secondary on-line study is obtained, to build convolutional neural networks model, and is input to structure
Training sample in convolutional neural networks model is the class label being stored in the first training sample memory cell 321 by target
The training sample (referred to as " unsupervised training sample ") that recognition unit 313 is marked.The 3rd grader obtained includes sample
The parameter value of this attribute classification mapping table and the model parameter being trained to.
Then, in step S 407, online deep learning unit 333 is updated with the 3rd grader obtained is stored in the
The first grader in one grader memory cell.
Target identification is carried out using the Target Identification Unit 200 of the present invention, solving needs to gather great amount of samples in the prior art
The problem of, save ample resources and time.The Target Identification Unit 200 of the present invention, can be in target by on-line study
In identification process, the grader for target identification is constantly updated and adjusted, the accuracy of identification is stepped up, well
The problems such as solving target appearance change, quickly move and block.Also, the deep learning network tool that the present invention is used
There are multiple hidden layers, possess the feature representation ability more excellent than shallow-layer network, algorithm tool insensitive to light, visual angle
There are preferable recognition effect and excellent classification performance.
In the target identification system 1 of the present invention, full supervision or semi-supervised off-line learning and unsupervised online are employed
The scheme being combined is practised, is adapted to various scenes, angle, illumination, the change of weather, solves the multiple dimensioned change of target
Drifted about loss problem with the target that causes is blocked, it is ensured that the adaptability and reliability of target classification.Off-line learning is using deep
This multilayer perceptron of degree study, it is not necessary to which special feature extraction algorithm, this part is automatically performed by network.It is online to learn
Habit is based entirely on non supervision model and obtains grader, and the supervised learning without being participated in by user trains class models,
So as to improve the automaticity of system.Also, on-line study training sample selection is using the high band classification mark of degree of belief
Signed-off sample sheet, advantageously ensures that the precision of on-line study.After study is finished after on-line training, it is high that selection does not abandon degree of belief
Sample, feed back to off-line training, be conducive to improve off-line learning precision.
The present invention is applied to high-precision in real-time the driving of automatic Pilot/auxiliary and high-precision obstacle recognition and video monitoring
Spend target identification.
Although by being described in conjunction with specific embodiments to the present invention, for the ordinary artisan of this area, root
With changing will be apparent according to many replacements, modification made after mentioned above.Therefore, when it is such substitute, repair
Change and change when falling within the spirit and scope of appended claims, it should be included in the present invention.
Claims (20)
1. a kind of target identification method, the target identification method is based on on-line automatic deep learning, it is characterised in that institute
The method of stating includes：
Real-time image acquisition data；
Using the first grader of storage, target identification is carried out to the view data collected, class label is carried to generate
View data, the class label include target classification and target posterior probability；
In the case where the target posterior probability meets predetermined storage condition, according to the target classification, carried described
The view data of class label is stored as training sample；And
In the case where predetermined entry condition is satisfied, start online deep learning processing, the online deep learning processing
Network model is built including the first grader based on the storage, and whole training samples of storage are input to structure
Network model in carry out deep learning processing, to obtain the 3rd grader, and with the 3rd grader of acquisition come
Update the first grader of the storage.
2. the method as described in claim 1, it is characterised in that what the first grader of the storage was used initially divides
Class device is by carrying out the offline deep learning classification that deep learning is obtained by the training sample manually marked to class label
Device.
3. method as claimed in claim 2, it is characterised in that the target identification method further comprises：
Periodically by whole training samples of the storage via network transmission to far-end server, to carry out offline deep learning
Processing；And
Received via the network from the far-end server and the offline deep learning processing is carried out by the far-end server
The second grader obtained, and when receiving second grader, with second grader received come
Update the first grader of the storage.
4. method as claimed in claim 3, it is characterised in that before target identification is carried out, methods described is further
Including carrying out image preprocessing to the view data that collects, to improve image definition, and collected from described
Area-of-interest view data is extracted in view data；And
When carrying out target identification, mesh is carried out to the area-of-interest view data in the view data collected
Mark is other.
5. method as claimed in claim 4, it is characterised in that before deep learning is carried out, the online depth
Habit processing further comprises carrying out image preprocessing to whole training samples of the storage, to improve image definition.
6. the method as any one of claim 1-5, it is characterised in that other described when belonging to the target class
When the quantity of the training sample of storage not yet reaches predetermined quantity, the predetermined storage condition uses the first storage condition, institute
State the first storage condition and be more than or equal to predetermined threshold for the target posterior probability；
It is described when belonging to the quantity of training sample of the other storage of the target class and having reached the predetermined quantity
Predetermined storage condition uses the second storage condition, and second storage condition is described in the target posterior probability is more than or equal to
Predetermined threshold, and the target posterior probability is minimum more than the training sample for belonging to the other storage of the target class
Target posterior probability；And
When the target posterior probability meets second storage condition, the target identification method further comprises, deletes
Belong to the other training sample with minimum target posterior probability of the target class except oldest stored.
7. method as claimed in claim 6, it is characterised in that the predetermined entry condition is to use the target identification
The equipment of method is in idle condition, and the training sample of each other storage of target class quantity reach it is described
Predetermined quantity.
8. method as claimed in claim 7, it is characterised in that the predetermined quantity is joined for the network of the network model
The other sum of sum × 10/ of the several numbers target class to be trained to.
9. a kind of Target Identification Unit, the Target Identification Unit is based on on-line automatic deep learning, it is characterised in that institute
Stating Target Identification Unit includes：Real-time detection apparatus, first storage device and on-line study device, wherein
The real-time detection apparatus includes：
Image data acquiring unit, described image data acquisition unit real-time image acquisition data；
Object-recognition unit, the object-recognition unit utilizes the first classification being stored in the first storage device
Device, target identification is carried out to the view data collected, to generate the view data with class label, the classification mark
Label include target classification and target posterior probability；
The first storage device includes：
First grader memory cell, the first grader memory cell stores first grader；And
First training sample memory cell, the first training sample memory cell is met in the target posterior probability
In the case of predetermined storage condition, according to the target classification, the view data with class label is stored as instruction
Practice sample；
The on-line study device includes：
On-line study start unit, the on-line study start unit in the case where predetermined entry condition is satisfied,
Start online deep learning processing；And
Online deep learning unit, when the on-line study start unit starts the online deep learning processing,
The online deep learning unit builds net based on first grader being stored in the first grader memory cell
Network model, and the whole training samples that will be stored in the first training sample memory cell are input to the network of structure
Deep learning processing is carried out in model, to obtain the 3rd grader, and is deposited with the 3rd grader obtained to update
Store up first grader in the first grader memory cell.
10. device as claimed in claim 9, it is characterised in that be stored in the institute in the first grader memory cell
It is by carrying out depth by the training sample manually marked to class label to state the preliminary classification device that the first grader used
Practise the offline deep learning grader obtained.
11. device as claimed in claim 10, it is characterised in that further comprise first data transmission device, its
In,
All training that the first data transmission device periodically will be stored in the first training sample memory cell
Sample via network transmission to far-end server, to carry out offline deep learning processing；And
The first data transmission device is received from the far-end server via the network and carried out by the far-end server
The offline deep learning handles the second obtained grader, and when receiving second grader, with reception
To second grader update first grader being stored in the first grader memory cell.
12. device as claimed in claim 11, it is characterised in that the real-time detection apparatus further comprises the first figure
As pretreatment unit, described first image pretreatment unit carries out image preprocessing to the view data collected, with
Image definition is improved, and area-of-interest view data is extracted from the view data collected；And
The object-recognition unit carries out target to the area-of-interest view data in the view data collected
Identification.
13. device as claimed in claim 12, it is characterised in that the on-line study device further comprises the 3rd figure
As pretreatment unit, the 3rd image pre-processing unit carries out deep learning in the online deep learning unit and handles it
Before, image preprocessing is carried out to the whole training samples being stored in the first training sample memory cell, to improve figure
Image sharpness.
14. the device as any one of claim 9-13, it is characterised in that train sample when being stored in described first
It is described to make a reservation for deposit when the quantity for belonging to the other training sample of the target class in this memory cell not yet reaches predetermined quantity
Storage condition uses the first storage condition, and first storage condition is that the target posterior probability is more than or equal to predetermined threshold；
When the quantity for belonging to the other training sample of the target class being stored in the first training sample memory cell
When reaching the predetermined quantity, the predetermined storage condition uses the second storage condition, and second storage condition is described
Target posterior probability is more than or equal to the predetermined threshold, and the target posterior probability is trained more than being stored in described first
The minimum target posterior probability for belonging to the other training sample of the target class in sample storage unit；And
When the target posterior probability meets second storage condition, the first training sample memory cell is deleted most
Early storage into the first training sample memory cell to belong to the target class other with minimum target posterior probability
Training sample.
15. device as claimed in claim 14, it is characterised in that the predetermined entry condition is to know with the target
The equipment of other device is in idle condition, and is stored in each target classification in the first training sample memory cell
The quantity of training sample reach the predetermined quantity.
16. device as claimed in claim 15, it is characterised in that the predetermined quantity is the network of the network model
The other sum of sum × 10/ of the number of parameters target class to be trained to.
17. a kind of terminal installation, it is characterised in that the terminal installation is included as any one of claim 9-16
Target Identification Unit.
18. a kind of target identification system, it is characterised in that the system include far-end server and via network with
Multiple terminal installations as claimed in claim 17 of the far-end server connection, wherein,
Each terminal installation is using the first grader being stored in the first grader memory cell to collecting in real time
View data carries out target identification, and the view data for meeting predetermined storage condition is stored as into training sample, and with passing through
The 3rd classification that deep learning processing is obtained is carried out to the whole training samples being stored in the first training sample memory cell
Device or with the second grader received from the far-end server, to update, to be stored in the first grader storage single
First grader in member；
Each terminal installation includes first data transmission device, and the first data transmission device is periodically via the net
Whole training samples of storage are transferred to the far-end server by network；
The far-end server includes the second storage device, off-line learning device and the second data transmission device, wherein,
Second storage device includes storing the second grader memory cell of second grader, and storage classification
Second training sample of the training sample that label is received by the training sample that manually marks and from the multiple terminal installation is deposited
Storage unit,
The off-line learning device includes offline deep learning unit, and the offline deep learning unit is described based on being stored in
Second grader in second grader memory cell builds network model, will be stored in second training sample and deposits
Whole training samples in storage unit are input to progress deep learning processing in the network model of structure, to obtain new second
Grader, and be stored in the second new grader obtained to update in the second grader memory cell
Second grader, and
Second grader after renewal is transferred to each end by second data transmission device via the network
End device.
19. system as claimed in claim 18, it is characterised in that be stored in the first grader memory cell
First grader and second grader being stored in the second grader memory cell use same initial
Grader, the preliminary classification device be the offline deep learning unit by class label by the training sample that manually marks
The offline deep learning grader that this progress deep learning processing is obtained.
20. system as claimed in claim 18, it is characterised in that the off-line learning device further comprises the second figure
As pretreatment unit, second image pre-processing unit carries out deep learning in the offline deep learning unit and handles it
Before, image preprocessing is carried out to the whole training samples being stored in the second training sample memory cell, to improve figure
Image sharpness.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610255052.2A CN107305636A (en) | 2016-04-22 | 2016-04-22 | Target identification method, Target Identification Unit, terminal device and target identification system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610255052.2A CN107305636A (en) | 2016-04-22 | 2016-04-22 | Target identification method, Target Identification Unit, terminal device and target identification system |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107305636A true CN107305636A (en) | 2017-10-31 |
Family
ID=60151742
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610255052.2A Pending CN107305636A (en) | 2016-04-22 | 2016-04-22 | Target identification method, Target Identification Unit, terminal device and target identification system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107305636A (en) |
Cited By (19)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107967460A (en) * | 2017-12-08 | 2018-04-27 | 重庆广睿达科技有限公司 | A kind of rubbish thing based on deep neural network burns recognition methods and system |
| CN107967491A (en) * | 2017-12-14 | 2018-04-27 | 北京木业邦科技有限公司 | Machine learning method, device, electronic equipment and the storage medium again of plank identification |
| CN108053031A (en) * | 2018-01-02 | 2018-05-18 | 安徽大学 | A kind of deep learning system and deep learning recognition methods |
| CN108205708A (en) * | 2018-01-02 | 2018-06-26 | 安徽大学 | A kind of novel expansible deep learning system and data identification method |
| CN108280453A (en) * | 2018-01-08 | 2018-07-13 | 西安电子科技大学 | The quick detection image goal approach of low-power consumption based on deep learning |
| CN108389220A (en) * | 2018-02-05 | 2018-08-10 | 湖南航升卫星科技有限公司 | Remote sensing video image motion target real-time intelligent cognitive method and its device |
| CN108427972A (en) * | 2018-04-24 | 2018-08-21 | 云南佳叶现代农业发展有限公司 | Tobacco classification method based on on-line study and its system |
| CN108510523A (en) * | 2018-03-16 | 2018-09-07 | 新智认知数据服务有限公司 | It is a kind of to establish the model for obtaining object feature and object searching method and device |
| CN108875959A (en) * | 2018-05-24 | 2018-11-23 | 四川斐讯信息技术有限公司 | A kind of intelligence sees the method and system of object knowledge name |
| CN109034380A (en) * | 2018-06-08 | 2018-12-18 | 四川斐讯信息技术有限公司 | A kind of distributed image identification system and its method |
| CN109754089A (en) * | 2018-12-04 | 2019-05-14 | 浙江大华技术股份有限公司 | A kind of model training systems and method |
| CN110135223A (en) * | 2018-02-08 | 2019-08-16 | 浙江宇视科技有限公司 | Method for detecting human face and device |
| WO2019179189A1 (en) * | 2018-03-23 | 2019-09-26 | 北京达佳互联信息技术有限公司 | Image classification model optimization method and device and terminal |
| CN111126606A (en) * | 2020-03-30 | 2020-05-08 | 同盾控股有限公司 | Platform-based deep learning model autonomous training method, device, equipment and medium |
| CN111487958A (en) * | 2019-01-28 | 2020-08-04 | 北京奇虎科技有限公司 | Control method and device of sweeping robot |
| US20200394289A1 (en) * | 2019-06-14 | 2020-12-17 | Microsoft Technology Licensing, Llc | Biometric verification framework that utilizes a convolutional neural network for feature matching |
| CN112823412A (en) * | 2018-10-08 | 2021-05-18 | 科磊股份有限公司 | Adaptive region of interest based on deep learning for critical dimension measurement of semiconductor substrates |
| CN112989084A (en) * | 2020-12-25 | 2021-06-18 | 深圳惟远智能技术有限公司 | Social behavior analysis method based on characteristic values |
| CN113614773A (en) * | 2019-03-08 | 2021-11-05 | Sk电信有限公司 | Image analysis apparatus and method, and method for generating image analysis model for image analysis apparatus and method |
Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101989327A (en) * | 2009-07-30 | 2011-03-23 | 富士胶片株式会社 | Image analyzing apparatus and image analyzing method |
| CN102024045A (en) * | 2010-12-14 | 2011-04-20 | 成都市华为赛门铁克科技有限公司 | Information classification processing method, device and terminal |
| CN102303605A (en) * | 2011-06-30 | 2012-01-04 | 中国汽车技术研究中心 | Multi-sensor information fusion-based collision and departure pre-warning device and method |
| CN102789593A (en) * | 2012-06-18 | 2012-11-21 | 北京大学 | Intrusion detection method based on incremental GHSOM (Growing Hierarchical Self-organizing Maps) neural network |
| CN102915453A (en) * | 2012-08-30 | 2013-02-06 | 华南理工大学 | Real-time feedback and update vehicle detection method |
| CN103236162A (en) * | 2013-04-11 | 2013-08-07 | 江苏大学 | Signalized intersection traffic capacity analyzing method based on based on video analysis |
| CN103793713A (en) * | 2012-10-31 | 2014-05-14 | 中国科学院沈阳自动化研究所 | Low-rank constraint online self-supervised learning scene classification method |
| CN105447498A (en) * | 2014-09-22 | 2016-03-30 | 三星电子株式会社 | A client device configured with a neural network, a system and a server system |
| CN105469400A (en) * | 2015-11-23 | 2016-04-06 | 广州视源电子科技股份有限公司 | Rapid identification and marking method of electronic component polarity direction and system thereof |
| CN105488490A (en) * | 2015-12-23 | 2016-04-13 | 天津天地伟业数码科技有限公司 | Judge dressing detection method based on video |
-
2016
- 2016-04-22 CN CN201610255052.2A patent/CN107305636A/en active Pending
Patent Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101989327A (en) * | 2009-07-30 | 2011-03-23 | 富士胶片株式会社 | Image analyzing apparatus and image analyzing method |
| CN102024045A (en) * | 2010-12-14 | 2011-04-20 | 成都市华为赛门铁克科技有限公司 | Information classification processing method, device and terminal |
| CN102303605A (en) * | 2011-06-30 | 2012-01-04 | 中国汽车技术研究中心 | Multi-sensor information fusion-based collision and departure pre-warning device and method |
| CN102789593A (en) * | 2012-06-18 | 2012-11-21 | 北京大学 | Intrusion detection method based on incremental GHSOM (Growing Hierarchical Self-organizing Maps) neural network |
| CN102915453A (en) * | 2012-08-30 | 2013-02-06 | 华南理工大学 | Real-time feedback and update vehicle detection method |
| CN103793713A (en) * | 2012-10-31 | 2014-05-14 | 中国科学院沈阳自动化研究所 | Low-rank constraint online self-supervised learning scene classification method |
| CN103236162A (en) * | 2013-04-11 | 2013-08-07 | 江苏大学 | Signalized intersection traffic capacity analyzing method based on based on video analysis |
| CN105447498A (en) * | 2014-09-22 | 2016-03-30 | 三星电子株式会社 | A client device configured with a neural network, a system and a server system |
| CN105469400A (en) * | 2015-11-23 | 2016-04-06 | 广州视源电子科技股份有限公司 | Rapid identification and marking method of electronic component polarity direction and system thereof |
| CN105488490A (en) * | 2015-12-23 | 2016-04-13 | 天津天地伟业数码科技有限公司 | Judge dressing detection method based on video |
Cited By (24)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107967460A (en) * | 2017-12-08 | 2018-04-27 | 重庆广睿达科技有限公司 | A kind of rubbish thing based on deep neural network burns recognition methods and system |
| CN107967460B (en) * | 2017-12-08 | 2020-05-08 | 重庆广睿达科技有限公司 | Deep neural network-based waste incineration identification method and system |
| CN107967491A (en) * | 2017-12-14 | 2018-04-27 | 北京木业邦科技有限公司 | Machine learning method, device, electronic equipment and the storage medium again of plank identification |
| CN108053031A (en) * | 2018-01-02 | 2018-05-18 | 安徽大学 | A kind of deep learning system and deep learning recognition methods |
| CN108205708A (en) * | 2018-01-02 | 2018-06-26 | 安徽大学 | A kind of novel expansible deep learning system and data identification method |
| CN108280453A (en) * | 2018-01-08 | 2018-07-13 | 西安电子科技大学 | The quick detection image goal approach of low-power consumption based on deep learning |
| CN108280453B (en) * | 2018-01-08 | 2020-06-16 | 西安电子科技大学 | Low-power-consumption rapid image target detection method based on deep learning |
| CN108389220A (en) * | 2018-02-05 | 2018-08-10 | 湖南航升卫星科技有限公司 | Remote sensing video image motion target real-time intelligent cognitive method and its device |
| CN110135223A (en) * | 2018-02-08 | 2019-08-16 | 浙江宇视科技有限公司 | Method for detecting human face and device |
| CN108510523A (en) * | 2018-03-16 | 2018-09-07 | 新智认知数据服务有限公司 | It is a kind of to establish the model for obtaining object feature and object searching method and device |
| WO2019179189A1 (en) * | 2018-03-23 | 2019-09-26 | 北京达佳互联信息技术有限公司 | Image classification model optimization method and device and terminal |
| US11544496B2 (en) | 2018-03-23 | 2023-01-03 | Beijing Dajia Internet Information Technology Co., Ltd. | Method for optimizing image classification model, and terminal and storage medium thereof |
| CN108427972A (en) * | 2018-04-24 | 2018-08-21 | 云南佳叶现代农业发展有限公司 | Tobacco classification method based on on-line study and its system |
| CN108875959A (en) * | 2018-05-24 | 2018-11-23 | 四川斐讯信息技术有限公司 | A kind of intelligence sees the method and system of object knowledge name |
| CN109034380A (en) * | 2018-06-08 | 2018-12-18 | 四川斐讯信息技术有限公司 | A kind of distributed image identification system and its method |
| CN112823412A (en) * | 2018-10-08 | 2021-05-18 | 科磊股份有限公司 | Adaptive region of interest based on deep learning for critical dimension measurement of semiconductor substrates |
| CN112823412B (en) * | 2018-10-08 | 2022-08-05 | 科磊股份有限公司 | Adaptive region of interest based on deep learning for critical dimension measurement of semiconductor substrates |
| CN109754089A (en) * | 2018-12-04 | 2019-05-14 | 浙江大华技术股份有限公司 | A kind of model training systems and method |
| CN109754089B (en) * | 2018-12-04 | 2021-07-20 | 浙江大华技术股份有限公司 | Model training system and method |
| CN111487958A (en) * | 2019-01-28 | 2020-08-04 | 北京奇虎科技有限公司 | Control method and device of sweeping robot |
| CN113614773A (en) * | 2019-03-08 | 2021-11-05 | Sk电信有限公司 | Image analysis apparatus and method, and method for generating image analysis model for image analysis apparatus and method |
| US20200394289A1 (en) * | 2019-06-14 | 2020-12-17 | Microsoft Technology Licensing, Llc | Biometric verification framework that utilizes a convolutional neural network for feature matching |
| CN111126606A (en) * | 2020-03-30 | 2020-05-08 | 同盾控股有限公司 | Platform-based deep learning model autonomous training method, device, equipment and medium |
| CN112989084A (en) * | 2020-12-25 | 2021-06-18 | 深圳惟远智能技术有限公司 | Social behavior analysis method based on characteristic values |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107305636A (en) | Target identification method, Target Identification Unit, terminal device and target identification system | |
| CN111353413B (en) | Low-missing-report-rate defect identification method for power transmission equipment | |
| CN110321923B (en) | Target detection method, system and medium for fusion of different-scale receptive field characteristic layers | |
| CN110163187B (en) | F-RCNN-based remote traffic sign detection and identification method | |
| CN106127204B (en) | A kind of multi-direction meter reading Region detection algorithms of full convolutional neural networks | |
| CN103324937B (en) | The method and apparatus of label target | |
| CN108921051B (en) | Pedestrian attribute identification network and technology based on cyclic neural network attention model | |
| CN109559302A (en) | Pipe video defect inspection method based on convolutional neural networks | |
| CN111783590A (en) | Multi-class small target detection method based on metric learning | |
| CN111444848A (en) | Specific scene model upgrading method and system based on federal learning | |
| CN109767422A (en) | Pipe detection recognition methods, storage medium and robot based on deep learning | |
| CN106408015A (en) | Road fork identification and depth estimation method based on convolutional neural network | |
| CN106022380A (en) | Individual identity identification method based on deep learning | |
| CN104992177A (en) | Internet porn image detection method based on deep convolution nerve network | |
| CN108596038A (en) | Erythrocyte Recognition method in the excrement with neural network is cut in a kind of combining form credit | |
| CN111368690A (en) | Deep learning-based video image ship detection method and system under influence of sea waves | |
| CN109544501A (en) | A kind of transmission facility defect inspection method based on unmanned plane multi-source image characteristic matching | |
| CN114677323A (en) | Semantic vision SLAM positioning method based on target detection in indoor dynamic scene | |
| CN114359695A (en) | Insulator breakage identification method based on uncertainty estimation | |
| CN109376677A (en) | A kind of video behavior detection method merged based on shape-movement double fluid information | |
| CN116310891A (en) | Cloud-edge cooperative transmission line defect intelligent detection system and method | |
| CN115311111A (en) | Classroom participation evaluation method and system | |
| CN111540203A (en) | Method for adjusting green light passing time based on fast-RCNN | |
| CN109685823A (en) | A kind of method for tracking target based on depth forest | |
| CN113673618A (en) | Tobacco insect target detection method fused with attention model |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20171031 |
|
| RJ01 | Rejection of invention patent application after publication |