CN105224947B - classifier training method and system - Google Patents
classifier training method and system Download PDFInfo
- Publication number
- CN105224947B CN105224947B CN201410250540.5A CN201410250540A CN105224947B CN 105224947 B CN105224947 B CN 105224947B CN 201410250540 A CN201410250540 A CN 201410250540A CN 105224947 B CN105224947 B CN 105224947B
- Authority
- CN
- China
- Prior art keywords
- sample
- negative
- positive
- difficult
- grader
- 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.)
- Active
Links
Abstract
The present invention provides a kind of grader adaptive training method and systems.This method includes:The multiple image continuously inputted using current class device online recognition, and it is directly appended to using identified image as the positive negative sample of high confidence level the sample database of grader;And the image output for being difficult to current class device is difficult sample, and for the context of the difficult sample, is identified as positive and negative authentic specimen, and assign the sample database that grader is added to after the identified certain weight of positive and negative authentic specimen.
Description
Technical field
The present invention relates to the fields for identifying specific object in the picture, in particular to the adaptive of Image Classifier
Answer training method and the system using this method.
Background technology
Widely available with digital technology, the especially extensive use of image digitazation, people are identified using image
The demand of specific object is also more and more.A large amount of generations of digital picture bring hard work for manual identified and appoint
Business.For this purpose, it is intended that it is a kind of can in automatic identification image specific object technology.Therefore, object automatic identification technology is answered
It transports and gives birth to.
Target identification is mainly used in field of video monitoring.In order to identify the objectives in image, on the one hand to extract
On the other hand some specific features of target will obtain some positive and negative image patterns to train classification based on these specific features
Device is to identify in current input image to being detected to the image currently obtained by trained listening group
No includes the position of target and target.Specifically, monitoring identifying system is by acquiring the object in various illumination conditions
Positive negative sample, feature is extracted to them, and is clustered, grader is respectively trained for each subclass being polymerized to, is then adopted
Object detection is carried out with trained grader.When carrying out object detection, the image of current scene is obtained first, then
It generates there may be the candidate region of object, feature is extracted to each candidate region, and draw candidate region according to these features
Corresponding subclass is assigned to, finally differentiates that the candidate region whether there is object using the grader of corresponding subclass.Therefore, in order to carry
The accuracy of identification of object in hi-vision, on the one hand will be improved in terms of feature extraction, further aspect is that in grader side
Face is improved.The improvement of feature extraction usually has directive property to the identification of specific object, but the improvement of grader is then right
All identifying systems are of universal significance.
Shown in FIG. 1 is the flow chart of trained subclass grader in the prior art.As shown in Figure 1, at step S11, point
Class device training unit reads sample image from the sample database of the memory of system.Include that the sample image of object pattern is known as
Positive sample is known as negative sample without the spurious edition image comprising object pattern.At step S12, feature extraction unit is from every width
Feature vector is extracted in sample image.Then, it at step S13, based on the feature vector that every width figure is extracted, carries out at cluster
Reason.Common k-nearest neighbor, K mean cluster method may be used in cluster.It is exactly specifically to be carried out to the feature vector of each image
Similarity-rough set, two width sample images corresponding to the feature vector of similarity between any two within certain threshold range are returned
For one kind.It is of course also possible to use existing other modes are clustered.In fact, cluster is a kind of routine techniques, contain
Various cluster modes, therefore be not specifically described herein.Later, at step S14, the institute's directed quantity for being classified as a kind of is corresponded to
Sample image be classified as a subclass, and each subclass is numbered.Mean value and variance of these subclasses in feature space
It will be used when will be stored for dividing candidate image when detection.Finally at step S15, the methods of existing machine learning pair is utilized
Each subclass training is the separated grader of positive negative sample.Support vector machines, AdaBoost etc. can be used in training subclass grader.
After having trained subclass grader for each subclass, the image that scene is obtained can be detected using subclass
In whether there is object.Shown in Fig. 2 is to carry out object in candidate image based on the subclass grader being trained in the prior art
The flow chart of detection.As shown in Figure 2, it is necessary first to detect candidate image area as defeated from current scene at step S21
Enter image.Using some simple features, symmetry or Haar+AdaBoost frames etc. are such as relied on.Scene image is usually adopted
It is obtained with scene image acquiring unit.Scene image acquiring unit is usually video camera, can object-based shooting photo,
Continuous scene image inputs to computer.Then, at step S22, feature extraction unit is carried from the scene image received
Go out feature vector.Then, the feature vector for calculating corresponding scene image at step S23 and every height in sample image set
The distance between class.Specifically, aiming at each subclass, the corresponding feature vector of all sample images in the subclass is calculated
Average value, to obtain the averaged feature vector of the subclass, the i.e. center of the subclass, the then spy of relatively corresponding scene image
The averaged feature vector of sign vector and the subclass, obtains distance between the two, which is then the feature of corresponding scene image
Vector is at a distance from the subclass.Then, at step S24, by nearest son at a distance from the feature vector of corresponding scene image
The number of the subclass is assigned the feature vector of the correspondence scene image by class as the subclass belonging to corresponding scene image.It connects
It, at step S25, the number of the subclass of the feature vector based on corresponding scene image calls the classification of the subclass of reference numeral
Device.Finally, it at step S26, is detected using the subclass grader called and whether there is object in corresponding scene image.Tool
How body using training aids belongs to the prior art to detect in candidate image with the presence or absence of object, therefore does not repeat herein.
It will be apparent that the robustness of grader is the main aspect of the variability in face of application scenarios.In the art, most
What is be often used is the grader of off-line training.For angle of statistics, the training sample type being collected into is more, recognition result
Better.However, it is in practical application that the scene faced is varied, therefore, it can not be collected into complete training sample set at all,
This causes clarification of objective in application environment also to be difficult to be fully set forth.In order to make identifying system can adapt to current application field
Scape, common method are that sample, the final off-line training of mark under a large amount of current scenes of collected offline are suitable for current scene
Grader.This is one and takes cumbersome work, and grader is very sensitive for scene changes, it is difficult to apply in video monitoring
In.Therefore, the system for having adaptive learning is very necessary.
Existing adaptive learning method is typically to collect high confidence level sample to update grader, the classification of these samples
The result is that determined by original off-line training grader.This more new algorithm can lead to grader overfitting and still cannot be just
It really distinguishes original classification device under current scene and is difficult to the target correctly identified.
Invention content
For this purpose, the present invention proposes a kind of classifier training method and system of adaptive learning, this method and system pass through
On-line automatic acquisition is come real-time based on the credible positive negative sample of time domain, spatial domain and semantic context information and other useful informations
Update original object recognition classifier.This method and system can be widely used in field of video monitoring, have stronger adaptation
Property and practicability.
According to an aspect of the invention, there is provided a kind of adaptive classifier training method, including:Using current class
The multiple image that device online recognition continuously inputs, and be directly appended to point using identified image as the positive negative sample of high confidence level
The sample database of class device;And the image output for being difficult to current class device is difficult sample, and for the difficult sample
Context, be identified as positive and negative authentic specimen, and be added to after assigning the identified certain weight of positive and negative authentic specimen
The sample database of grader.
According to adaptive classifier training method of the present invention, the context for the difficult sample is identified
For credible positive negative sample, and assign the step for the sample database that grader is added to after the identified certain weight of positive and negative authentic specimen
Suddenly include:Based on the time domain context of specific difficult sample, the positive and negative authentic specimen of the time domain context of all difficult samples is acquired,
And the positive and negative authentic specimen for calculating the time domain context includes the probability of target to be identified under time-domain information;It is doubted based on specific
The spatial domain context of difficult sample, acquires the positive and negative authentic specimen of the spatial domain context of all difficult samples, and calculates on the spatial domain
Positive and negative authentic specimen hereafter includes the probability of target to be identified under spatial information (si);Based on specific difficult sample semantically under
Text, acquires the positive and negative authentic specimen of the semantic context of all difficult samples, and calculates the positive and negative credible sample of the semantic context
This includes the probability of target to be identified under semantic information;And it the influence based on current class device to difficult sample and is calculated
Difficult sample include the probability of target to be identified under time-domain information, spatial information (si) and semantic information, calculate the difficulty sample
This joint probability as upper positive and negative authentic specimen, and be each selected as according to the joint probability imparting calculated positive and negative credible
One weighted value of difficult sample of sample.
According to adaptive classifier training method of the present invention, further include:Before initially use current class device, collect just
Existing positive and negative authentic specimen is to create the training sample database of target to be identified;To each sample extraction in training sample database
Feature;Original target identification grader is trained to obtain initial current class device based on the feature set extracted.
According to adaptive classifier training method of the present invention, the feature that is extracted includes:Histogram of gradients feature or office
Portion's binary pattern feature.
According to adaptive classifier training method of the present invention, the grader includes:Support vector machines, Adaboost classification
Device or neural network.
According to another aspect of the present invention, a kind of grader adaptive training system is provided, including:Simple sample point
It distinguishes unit, by the multiple image Direct Recognition inputted is figure comprising or not comprising target to be identified based on current class device
Picture is directly appended to the sample of grader using the image comprising or not comprising target to be identified as the positive negative sample of high confidence level
Library;And difficult sample resolution cell, reception current class device fails the image of Direct Recognition as difficult sample, and is directed to institute
The context for stating difficult sample is identified as positive and negative authentic specimen, and after assigning the identified certain weight of positive and negative authentic specimen
It is added to the sample database of grader.
Grader adaptive training system according to the present invention, the difficulty sample resolution cell include:Time domain is differentiated single
Member acquires the positive and negative authentic specimen of the time domain context of all difficult samples based on the time domain context of specific difficult sample, and counts
The positive and negative authentic specimen for calculating the time domain context includes the probability of target to be identified under time-domain information;Spatial domain resolution cell,
Based on the spatial domain context of specific difficult sample, the positive and negative authentic specimen of the spatial domain context of all difficult samples is acquired, and is calculated
The positive and negative authentic specimen of the spatial domain context includes the probability of target to be identified under spatial information (si);Semantic resolution cell, base
In the semantic context of specific difficult sample, the positive and negative authentic specimen of the semantic context of all difficult samples is acquired, and calculates institute
The positive and negative authentic specimen for stating semantic context includes the probability of target to be identified under semantic information;And integrated treatment unit,
Influence based on current class device to difficult sample is believed with the difficult sample calculated in time-domain information, spatial information (si) and semanteme
The lower probability for including target to be identified of breath, calculates joint probability of the difficulty sample as upper positive and negative authentic specimen, and according to institute
The joint probability of calculating assigns one weighted value of difficult sample for being each selected as positive and negative authentic specimen.
It, being capable of the believable positive and negative target sample of online acquisition using the above-mentioned grader adaptive training method and system of the present invention
Sheet and other useful information.The sample collection method of the present invention is a kind of intelligent sample automatic acquiring method, in conjunction with
Time domain, spatial domain and semantic context information can be added the sample that original classification device is difficult to and work as front court to promote grader
The classification accuracy of scape, while over-fitting can be avoided.And by assigning different authentic specimens different confidence levels
Decline to reduce the performance that error sample selection is brought.Real-time update is carried out to original classification device based on authentic specimen, can be made
System has stronger adaptivity and practicability, makes every effort to obtain in the field of video monitoring of such as automobile assistant driving more extensive
Use.
Description of the drawings
By reading the detailed description of preferred embodiment of the invention below being considered in conjunction with the accompanying, it is better understood with this
The above and other target, feature, advantage and the technology and industrial significance of invention.
Shown in FIG. 1 is the flow chart of trained original classification device in the prior art.
Shown in Fig. 2 is the stream for being carried out in the prior art based on the subclass grader being trained to object detection in candidate image
Cheng Tu.
Shown in Fig. 3 is the flow chart of online adaptive classification device training method according to the ... of the embodiment of the present invention.
Shown in Fig. 4 is the flow chart that authentic specimen according to the ... of the embodiment of the present invention collects process.
The schematic diagram of the example shown in fig. 5 for being spatial domain context according to the ... of the embodiment of the present invention.
The schematic diagram of the strength shown in fig. 6 for being semantic context according to the ... of the embodiment of the present invention.
Shown in Fig. 7 is the concrete configuration figure of online adaptive classification device training system according to the ... of the embodiment of the present invention.
Specific implementation mode
In order to make those skilled in the art more fully understand the present invention, with reference to the accompanying drawings and detailed description to this hair
It is bright to be described in further detail.
According to an embodiment of the invention, the invention reside on the basis of above-mentioned Fig. 1 and Fig. 2, to point in identifying system
The training of class device is improved.The totality of the method shown in Fig. 3 for being adaptive training grader according to the ... of the embodiment of the present invention
Flow chart.
As shown in figure 3, first, at step S31, using current class device acquiring unit 710, for calling currently
The grader 30 used.Current class device may be to obtain original classification device using any prior art, can also be using this
The adaptive classifier that the classifier training method of invention obtains.Then, at step S32, sample collection unit 720 is from
Positive and negative authentic specimen is collected in identified input picture, and collected positive and negative authentic specimen is input to self study list
In member 730.Finally, in step S33, sample and received from sample that self study unit 730 is possessed based on current class device
Positive and negative authentic specimen is obtained in collection unit 720, executes self-learning strategy to update current class device, so that current class device
It is adaptive to currently identified environment.
Under current scene, after the authentic specimen of system automatic collection to certain amount, adaptive classifier will be real-time
Update is primary.That is, after the authentic specimen for collecting certain data, the S31-S33 that repeats the above steps is primary.
Here, SlearnAnd SotherRepresent the influence value of useful information under current scene.SlearnIt represents and is collected under current scene
Credible positive negative sample influence value.SotherRepresent the influence value of other useful informations under current scene.Here " other useful
Information " can also be the useful information obtained from other channels, such as target not just from the useful information in video frame
GPS positioning result.SlearnAnd SotherComplementation influences.WithIt is the influence value of credible positive negative sample respectively.It is as follows
Formula gives the output recognition result of adaptive classifier:
S=(1- γ) Sclassifier+γ(Slearn+Sother)
γ is the weight of study.Final categorised decision result is as follows:
T is the threshold value of target identification.
It is application example of the adaptive targets sorting algorithm on visual dictionary model below.Visual dictionary model can be compared with
Target signature is described well, there is higher recognition accuracy.View-based access control model dictionary model algorithm specific steps include:Feature carries
It takes, visual dictionary generates, visual dictionary matches, the model of view-based access control model dictionary generates.It is credible being obtained according to the method for the present invention
After sample, characteristic point and its description of these samples are extracted first.Find the high frequency entry C of positive and negative authentic specimenk, and to each
The high frequency entry of a selection is with weights omega (Ck), the weight is directly proportional to the frequency that corresponding entry occurs in authentic specimen.?
Under the model, S is expressed by the similarity of characteristic point and the high frequency visual entry selected in region to be identifiedlearn.Tool
Body calculation is as follows:
X=+or-
Here, CkIt is k-th of high frequency entry in x classes (x indicates that positive sample target or negative sample are non-targeted).fpIt is to wait knowing
Characteristic point in other region.Similarity is common similarity measurement strategy.
Fig. 4 is the flow chart of sample collection unit acquisition authentic specimen according to the ... of the embodiment of the present invention.As shown in figure 4, first
First at step S3210, the sample reception unit 7205 in sample collection unit 720 is received from current class device by current
The image data that grader was identified.These include three classes image by current class device identification image data
Data:The first kind be identified as include identified target image, be referred to as the positive sample with high confidence level herein;
Second class is the identified image without containing identified target, is referred to as the negative sample with high execution degree herein;Third class
For using current class device be difficult to correctly identify whether include identified target image, referred to herein as " difficult sample ".?
In existing online adaptive classification device, usually by the sample database of the first kind and the second class sample collection to grader, to more
Its new sample database and training grader.Grader after training in this way will be more difficult to identification third class image, from
And leading to over-fitting, the robustness for the grader for being also is deteriorated.For this purpose, the present invention is in adaptive online collection grader
Third class sample is utilized when sample, rather than abandons third class sample.
For this purpose, the present invention is at step S3220,7205 simple sample of sample reception unit in sample collection unit 720
Classification results of the resolution cell 7210 based on current class device differentiate the positive negative sample with high confidence level, i.e., above-described
The first kind and the second class sample.These positive negative samples with high confidence level are added directly to the sample database of grader.
Then, at step S3230, the difficult sample of 7205 simple sample resolution cell of sample reception unit 7210 is differentiated single
Time domain resolution cell 7230 in member 7220 executes the difficult sample that all current class devices are not told in the time domain to be divided
It distinguishes.Specifically, whether sentencing difficult disconnected sample based on the context authentic specimen in time of the difficult sample inputted
It is credible, that is, if it includes identified target to have multiframe in the frame number of the predetermined quantity of current class device whithin a period of time,
These be identified as between the frame number comprising target it is unrecognized go out target frame image include necessarily then identified mesh
It marks (the positive authentic specimen of high confidence level).Therefore, if certain difficult sample frame is identified as the multiframe comprising target at these and just may be used
Between believing sample, then these images being difficult to out by current class device (difficult sample) are then based on successive frame where it
Time domain specification (time domain context) and think to belong to positive authentic specimen by time domain resolution cell 7230 and (include the sample of identified target
This), and assign its certain time domain context probability.Equally, if the predetermined quantity of current class device whithin a period of time
There is multiframe not include identified target (high confidence level bears authentic specimen) in frame number, then these frames for being identified as including target
The frame image for being identified target between number, which does not include necessarily then, yet identified target.Therefore, if certain difficult sample
Frame is between these are not recognized as the negative authentic specimen of the multiframe comprising target, then these are difficult to out by current class device
Time domain specification (time domain context) of the image (difficult sample) then based on successive frame where it and by time domain resolution cell 7230
Think to belong to negative authentic specimen (sample for not including identified target), and assigns its certain time domain context probability.
As described above, in object recognition system, time domain contextual information is indicated with movable information.Due to visual angle, posture
Deng variation, certain target individuals can be correctly validated during being continuously tracked in partial frame image, but cannot be by always
Identification.By tracking result it can be assumed that these recognition results belong to a target.Therefore, movable information can be used for solving
The difficult target of this type.
For example, a target has correctly been tracked N frames, calculate its based on time domain contextual information Probability p (o=x | o,
Temporal) it is:
Wherein, otIt is observation of the target in t frames,It is N number of observation of the target in t-N to t frames.The observation is just
It is state existing for target, ot=x indicates that target present existence or feature are judged as classification x.
Meanwhile at step S3240, the spatial domain resolution cell 7240 in difficult sample resolution cell 7220 is to all current
The difficult sample that grader is not told executes resolution on spatial domain.Spatial domain contextual information is indicated with parallax information.Specifically
For, whether the context spatially based on the difficult sample inputted is credible come judgement sample, that is, due to passing through parallax
Figure can obtain the real space size of some target, for example, the information such as its true altitude, width.If the spatial information with
Identified target has high similarity, it will usually be considered the frame image being difficult to by current class device in its spatial domain
There is certain confidence level in terms of context.Size according to target can calculate its Probability p based on spatial domain contextual information
(o=x | o, Spatial).
Equally, at step S3250, the semantic resolution cell 7250 in difficult sample resolution cell 7220 is to all current
The difficult sample that grader is not told executes identification in semantic context.Compared to time domain, spatial information (si), semantically under
Literary information is a kind of understanding information of high level.It is often a kind of common sense or rule, and this information is for identifying current class
Device not can recognize that the difficult negative sample of target is helpful.Semantic context information usually with occur simultaneously in scene
The characteristic of other targets (non-target to be identified) is related, for example, automobile assistant driving system is needed from the road picture of shooting
Detection identification pedestrian target, but it is pedestrian that original classification device, which often accidentally knows railing etc. like personage's body, and these objects often have
Like the target of people in the features, such as road shown in Fig. 5 such as standby its unique texture, shape, color.Semantic context information
Whether the occasion also comprising target to be identified appearance is reasonable, for example, pedestrian will not swim in the air.As Fig. 6 thick dashed line is searched for
Shown in frame, the humanoid mark wherein in crossing traffic light may be identified as humanoid subject.But the humanoid subject
It is that floating is skyborne.Therefore, according to the semantic context information, it is object to be identified that can exclude the humanoid subject not, because
The specimen discerning is the negative sample in difficult sample by this.Therefore, such difficult negative sample can be acquired by semantic understanding.This
The result that kind is differentiated simply indicates probability by 0 and 1 result.
Then, at step S3260, the integrated treatment unit 7260 in difficult sample resolution cell 7220, which is directed to, to be resolved
Difficult sample, consider current class device, time domain context, spatial domain context and semantic context to its confidence level
It influences, obtains the final probability value of each difficult sample.Integrated treatment unit 7260 be combine above-mentioned three classes contextual information and
Grader recognition result obtains the authentic specimen based on context, need to meet following joint probability formula:
p(x|o,Temporal,Spatial,Semantic,Classifier)
Wherein, x is recognition result, be entirely in the case where observing o, in conjunction with Temporal, Spatial, Semantic,
What the result of Classifier obtained, to be a probability;O is the observation of sample.Temporal,Spatial,Semantic,
Classifier is respectively the influence of time domain, spatial domain, semanteme and grader, i.e., final differentiation probability.
The probability value calculated by above-mentioned joint probability formula assigns each authentic specimen one in all difficult samples
Determine weight.For the authentic specimen in be output to grader, weighted value is directly proportional to the joint probability calculated.
Shown in Fig. 7 is the concrete configuration figure of online adaptive classification device training system according to the ... of the embodiment of the present invention.Such as
Shown in Fig. 7, the figure online adaptive classification device training system 700 of the invention includes:Current class device acquiring unit
710, sample collection unit 720 and self study unit 730.The sample collection unit 720 include sample reception unit 7205,
Simple sample resolution cell 7210 and difficult sample resolution cell 7220.The difficulty sample resolution cell 7220 includes time domain
Resolution cell, 7230 spatial domain resolution cells 7240, semantic resolution cell 7250 and integrated treatment unit 7260.
The further identification process of difficult sample through the above steps, is collected into more comprehensively authentic specimen, makes
The sample database of grader is obtained more close to the truth of grader institute application environment.Same algorithm can be promoted and working as front court
The accuracy rate of identification target in scape.Moreover, because high confidence level sample is not only added in the sample database of grader, also add
Authentic specimen in difficult sample, therefore can effectively solve the problem that common over-fitting during self study.Meanwhile passing through root
The different weight of authentic specimen of selection is assigned according to different contextual informations and is applied in grader update, can effectively be reduced
The influence that error sample brings system performance.
The basic principle that the present invention is described above in association with specific embodiment, however, it is desirable to, it is noted that this field
For those of ordinary skill, it is to be understood that the whole either any steps or component of methods and apparatus of the present invention, Ke Yi
Any computing device (including processor, storage medium etc.) either in the network of computing device with hardware, firmware, software or
Combination thereof is realized that this is that those of ordinary skill in the art use them in the case where having read the explanation of the present invention
Basic programming skill can be achieved with.
Therefore, the purpose of the present invention can also by run on any computing device a program or batch processing come
It realizes.The computing device can be well known fexible unit.Therefore, the purpose of the present invention can also include only by offer
The program product of the program code of the method or device is realized to realize.That is, such program product is also constituted
The present invention, and the storage medium for being stored with such program product also constitutes the present invention.Obviously, the storage medium can be
Any well known storage medium or any storage medium that developed in the future.
It may also be noted that in apparatus and method of the present invention, it is clear that each component or each step are can to decompose
And/or reconfigure.These decompose and/or reconfigure the equivalent scheme that should be regarded as the present invention.Also, execute above-mentioned series
The step of processing, can execute according to the sequence of explanation in chronological order naturally, but not need to centainly sequentially in time
It executes.Certain steps can execute parallel or independently of one another.
Above-mentioned specific implementation mode, does not constitute limiting the scope of the invention.Those skilled in the art should be bright
It is white, design requirement and other factors are depended on, various modifications, combination, sub-portfolio and replacement can occur.It is any
Modifications, equivalent substitutions and improvements made by within the spirit and principles in the present invention etc., should be included in the scope of the present invention
Within.
Claims (8)
1. a kind of grader adaptive training method, including:
The multiple image continuously inputted using current class device online recognition, and identified image is positive and negative as high confidence level
Sample is directly appended to the sample database of grader;And
The image output that current class device is difficult to is difficult sample, and for the context of the difficult sample, by it
It is identified as positive and negative authentic specimen, and assigns the sample for being added to grader after the identified certain weight of positive and negative authentic specimen
Library, wherein the weight is time domain context, spatial domain context and semantic context based on specific difficult sample and current
Influence of the grader to difficult sample and it is calculated.
2. grader adaptive training method according to claim 1, the context for the difficult sample will
It is identified as credible positive negative sample, and assigns the sample that grader is added to after the identified certain weight of positive and negative authentic specimen
The step of library includes:
Based on the time domain context of specific difficult sample, the positive and negative authentic specimen of the time domain context of all difficult samples is acquired, and
The positive and negative authentic specimen for calculating the time domain context includes the probability of target to be identified under time-domain information;
Based on the spatial domain context of specific difficult sample, the positive and negative authentic specimen of the spatial domain context of all difficult samples is acquired, and
The positive and negative authentic specimen for calculating the spatial domain context includes the probability of target to be identified under spatial information (si);
Based on the semantic context of specific difficult sample, the positive and negative authentic specimen of the semantic context of all difficult samples is acquired, and
The positive and negative authentic specimen for calculating the semantic context includes the probability of target to be identified under semantic information;And
Based on current class device to the difficult sample for influencing and being calculated of difficult sample in time-domain information, spatial information (si) and language
The probability for including target to be identified under adopted information, calculates joint probability of the difficulty sample as upper positive and negative authentic specimen, and according to
One weighted value of difficult sample for being each selected as positive and negative authentic specimen is assigned according to the joint probability calculated.
3. grader adaptive training method according to claim 1 or 2, further includes:
Before initially use current class device, just existing positive and negative authentic specimen is collected to create the training sample of target to be identified
This library;
To each sample extraction feature in training sample database;
Original target identification grader is trained to obtain initial current class device based on the feature set extracted.
4. grader adaptive training method according to claim 3, the feature that is extracted include:Histogram of gradients is special
Sign or local binary patterns feature.
5. grader adaptive training method according to claim 3, the grader include:Support vector machines,
Adaboost graders or neural network.
6. a kind of grader adaptive training system, including:
The multiple image Direct Recognition inputted is comprising or not comprising waiting for based on current class device by simple sample resolution cell
The image for identifying target is directly appended to point using the image comprising or not comprising target to be identified as the positive negative sample of high confidence level
The sample database of class device;And
Difficult sample resolution cell, reception current class device fail the image of Direct Recognition as difficult sample, and for described
The context of difficult sample is identified as positive and negative authentic specimen, and is incited somebody to action after assigning the identified certain weight of positive and negative authentic specimen
It is added to the sample database of grader, wherein the weight is time domain context, spatial domain context based on specific difficult sample
Influence with semantic context and current class device to difficult sample and it is calculated.
7. grader adaptive training system according to claim 6, the difficulty sample resolution cell include:
Time domain resolution cell is acquiring the time domain context of all difficult samples just based on the time domain context of specific difficult sample
Negative authentic specimen, and the positive and negative authentic specimen for calculating the time domain context is general comprising target to be identified under time-domain information
Rate;
Spatial domain resolution cell is acquiring the spatial domain context of all difficult samples just based on the spatial domain context of specific difficult sample
Negative authentic specimen, and the positive and negative authentic specimen for calculating the spatial domain context is general comprising target to be identified under spatial information (si)
Rate;
Semantic resolution cell is acquiring the semantic context of all difficult samples just based on the semantic context of specific difficult sample
Negative authentic specimen, and the positive and negative authentic specimen for calculating the semantic context is general comprising target to be identified under semantic information
Rate;And
Integrated treatment unit, influence and the difficult sample that is calculated based on current class device to difficult sample time-domain information,
The probability for including target to be identified under spatial information (si) and semantic information calculates the difficulty sample as upper positive and negative authentic specimen
Joint probability, and one weight of difficult sample for being each selected as positive and negative authentic specimen is assigned according to the joint probability calculated
Value.
8. the grader adaptive training system described according to claim 6 or 7, the grader include:Support vector machines,
Adaboost graders or neural network.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201410250540.5A CN105224947B (en) | 2014-06-06 | 2014-06-06 | classifier training method and system |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201410250540.5A CN105224947B (en) | 2014-06-06 | 2014-06-06 | classifier training method and system |
Publications (2)
Publication Number | Publication Date |
---|---|
CN105224947A CN105224947A (en) | 2016-01-06 |
CN105224947B true CN105224947B (en) | 2018-11-13 |
Family
ID=54993907
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201410250540.5A Active CN105224947B (en) | 2014-06-06 | 2014-06-06 | classifier training method and system |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN105224947B (en) |
Families Citing this family (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105974219B (en) * | 2016-04-11 | 2021-01-15 | 中国电力科学研究院 | Classification and identification method for energy-saving electrical appliance load types |
CN106022392B (en) * | 2016-06-02 | 2019-09-13 | 华南理工大学 | A kind of training method that deep neural network sample is accepted or rejected automatically |
CN106920248A (en) * | 2017-01-19 | 2017-07-04 | 博康智能信息技术有限公司上海分公司 | A kind of method for tracking target and device |
CN106909885A (en) * | 2017-01-19 | 2017-06-30 | 博康智能信息技术有限公司上海分公司 | A kind of method for tracking target and device based on target candidate |
CN106920247A (en) * | 2017-01-19 | 2017-07-04 | 博康智能信息技术有限公司上海分公司 | A kind of method for tracking target and device based on comparison network |
CN107832780B (en) * | 2017-10-17 | 2020-04-10 | 北京木业邦科技有限公司 | Artificial intelligence-based wood board sorting low-confidence sample processing method and system |
CN108647731A (en) * | 2018-05-14 | 2018-10-12 | 宁波江丰生物信息技术有限公司 | Cervical carcinoma identification model training method based on Active Learning |
CN108874900A (en) * | 2018-05-24 | 2018-11-23 | 四川斐讯信息技术有限公司 | A kind of acquisition methods and system of samples pictures data acquisition system |
CN110288007B (en) * | 2019-06-05 | 2021-02-02 | 北京三快在线科技有限公司 | Data labeling method and device and electronic equipment |
JP2021008747A (en) * | 2019-07-01 | 2021-01-28 | 株式会社小松製作所 | Method of manufacturing trained work classification estimation model, training data, method performed by computer, and system including work machine |
CN111368926B (en) * | 2020-03-06 | 2021-07-06 | 腾讯科技(深圳)有限公司 | Image screening method, device and computer readable storage medium |
CN111461243B (en) * | 2020-04-08 | 2023-06-20 | 中国医学科学院肿瘤医院 | Classification method, classification device, electronic equipment and computer-readable storage medium |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103150578A (en) * | 2013-04-09 | 2013-06-12 | 山东师范大学 | Training method of SVM (Support Vector Machine) classifier based on semi-supervised learning |
CN103426008A (en) * | 2013-08-29 | 2013-12-04 | 北京大学深圳研究生院 | Vision human hand tracking method and system based on on-line machine learning |
CN103473560A (en) * | 2013-09-08 | 2013-12-25 | 西安电子科技大学 | Target tracking method based on context constraint codes |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US8442279B2 (en) * | 2008-06-27 | 2013-05-14 | Lockheed Martin Corporation | Assessing biometric sample quality using wavelets and a boosted classifier |
-
2014
- 2014-06-06 CN CN201410250540.5A patent/CN105224947B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103150578A (en) * | 2013-04-09 | 2013-06-12 | 山东师范大学 | Training method of SVM (Support Vector Machine) classifier based on semi-supervised learning |
CN103426008A (en) * | 2013-08-29 | 2013-12-04 | 北京大学深圳研究生院 | Vision human hand tracking method and system based on on-line machine learning |
CN103473560A (en) * | 2013-09-08 | 2013-12-25 | 西安电子科技大学 | Target tracking method based on context constraint codes |
Also Published As
Publication number | Publication date |
---|---|
CN105224947A (en) | 2016-01-06 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN105224947B (en) | classifier training method and system | |
CN108596211B (en) | Shielded pedestrian re-identification method based on centralized learning and deep network learning | |
US10445602B2 (en) | Apparatus and method for recognizing traffic signs | |
CN104615986B (en) | The method that pedestrian detection is carried out to the video image of scene changes using multi-detector | |
CN111144490A (en) | Fine granularity identification method based on alternative knowledge distillation strategy | |
CN107633226B (en) | Human body motion tracking feature processing method | |
CN105718866B (en) | A kind of detection of sensation target and recognition methods | |
CN103136504A (en) | Face recognition method and device | |
CN110298297A (en) | Flame identification method and device | |
KR20170006355A (en) | Method of motion vector and feature vector based fake face detection and apparatus for the same | |
JP2015215877A (en) | Object detection method from stereo image pair | |
CN110781785A (en) | Traffic scene pedestrian detection method improved based on fast RCNN algorithm | |
CN111191535B (en) | Pedestrian detection model construction method based on deep learning and pedestrian detection method | |
CN112257799A (en) | Method, system and device for detecting household garbage target | |
CN108446613A (en) | A kind of pedestrian's recognition methods again based on distance centerization and projection vector study | |
WO2015146113A1 (en) | Identification dictionary learning system, identification dictionary learning method, and recording medium | |
CN111274964B (en) | Detection method for analyzing water surface pollutants based on visual saliency of unmanned aerial vehicle | |
CN109697727A (en) | Method for tracking target, system and storage medium based on correlation filtering and metric learning | |
CN111444816A (en) | Multi-scale dense pedestrian detection method based on fast RCNN | |
CN115100497A (en) | Robot-based method, device, equipment and medium for routing inspection of abnormal objects in channel | |
CN108509825A (en) | A kind of Face tracking and recognition method based on video flowing | |
CN113870254A (en) | Target object detection method and device, electronic equipment and storage medium | |
CN111814760B (en) | Face recognition method and system | |
US20200034649A1 (en) | Object tracking system, intelligent imaging device, object feature extraction device, and object feature extraction method | |
CN109711232A (en) | Deep learning pedestrian recognition methods again based on multiple objective function |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |