CN105224947A - Sorter training method and system - Google Patents
Sorter training method and system Download PDFInfo
- Publication number
- CN105224947A CN105224947A CN201410250540.5A CN201410250540A CN105224947A CN 105224947 A CN105224947 A CN 105224947A CN 201410250540 A CN201410250540 A CN 201410250540A CN 105224947 A CN105224947 A CN 105224947A
- Authority
- CN
- China
- Prior art keywords
- sample
- negative
- positive
- sorter
- difficult
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Landscapes
- Image Analysis (AREA)
Abstract
The invention provides a kind of sorter adaptive training method and system.The method comprises: adopt the multiple image that current class device ONLINE RECOGNITION inputs continuously, and the image be identified directly is added to the Sample Storehouse of sorter as the positive negative sample of high confidence level; And impalpable for current class device image is exported as difficult sample, and for the context of described difficult sample, be identified as positive and negative authentic specimen, and give identify the certain weight of positive and negative authentic specimen after added to the Sample Storehouse of sorter.
Description
Technical field
The present invention relates to the field identifying concrete object in the picture, in particular to the adaptive training method of Image Classifier and the system using the method.
Background technology
Along with extensively popularizing of digital technology, the widespread use of especially image digitazation, people utilize image to identify that the demand of concrete object also gets more and more.A large amount of generations of digital picture bring hard work task concerning artificial cognition.For this reason, people expect a kind of can the technology of concrete object in recognition image automatically.Therefore, object automatic identification technology arises at the historic moment.
Target identification is mainly used in field of video monitoring.In order to the objectives in recognition image, some specific features of target will be extracted on the one hand, will obtain some positive and negative image patterns based on these specific features carrys out training classifier on the other hand, thus detected by the image of trained listening group to current acquisition, thus identify the position whether including target and target in current input image.Specifically, monitoring recognition system, by being captured in the positive negative sample of the object of various illumination condition, being extracted feature to them, and is carried out cluster, for each subclass be polymerized to training classifier respectively, then adopts trained sorter to carry out object detection.When carrying out object detection, first the image of current scene is obtained, then the candidate region that may there is object is generated, feature is extracted to each candidate region, and according to these features, candidate region is divided into corresponding subclass, whether this candidate region exists object finally to utilize the sorter of corresponding subclass to differentiate.Therefore, in order to improve the accuracy of identification of objects in images, to improve in feature extraction on the one hand, being to improve in sorter on the other hand.The improvement of feature extraction has directive property to the identification of concrete object usually, but the improvement of sorter is then of universal significance to all recognition systems.
Shown in Fig. 1 is the process flow diagram of training subclass sorter in prior art.As shown in Figure 1, in step S11 place, sorter training unit reads sample image from the Sample Storehouse of the storer of system.The sample image including object pattern is called positive sample, and the spurious edition image not comprising object pattern is called negative sample.In step S12 place, feature extraction unit extracts proper vector from every width sample image.Subsequently, in step S13 place, based on the proper vector that every width figure extracts, carry out clustering processing.Cluster can adopt common k-nearest neighbor, K means Method.Specifically carry out similarity-rough set to the proper vector of every width image exactly, two width sample images corresponding to the proper vector of similarity between any two within certain threshold range are classified as a class.Certainly, other modes existing also can be adopted to carry out cluster.In fact, cluster is a kind of routine techniques, contains various cluster mode, is not therefore specifically described at this.Afterwards, in step S14 place, sample image corresponding for the institute's directed quantity being classified as a class is classified as a subclass, and each subclass is numbered.The average of these subclasses in feature space and variance will be stored for when dividing candidate image during detection.Last in step S15 place, utilize the methods such as existing machine learning each subclass to be trained to the sorter that positive negative sample is separated.Training subclass sorter can adopt support vector machine, AdaBoost etc.
After trained subclass sorter for each subclass, in subclass can be utilized image that Test Field obtains, whether there is object.Shown in Fig. 2 is based on the process flow diagram being carried out object detection in candidate image by the subclass sorter of training in prior art.As shown in Figure 2, first need from current scene, to detect candidate image area as input picture in step S21 place.Utilize some simple features, as relied on symmetry, or Haar+AdaBoost framework etc.Scene image adopts scene image acquiring unit to obtain usually.Scene image acquiring unit is generally video camera, and it can object-basedly be taken pictures, and continuous scene image inputs to computing machine.Subsequently, in step S22 place, feature extraction unit proposes proper vector from received scene image.Then, in step S23 place, the distance between each subclass in the proper vector of corresponding scene image and sample image set is calculated.Specifically, be exactly for each subclass, calculate the mean value of all sample image characteristic of correspondence vectors in this subclass, thus obtain the averaged feature vector of this subclass, the i.e. center of this subclass, then the proper vector of more corresponding scene image and the averaged feature vector of this subclass, obtain distance between the two, this distance is then the proper vector of corresponding scene image and the distance of this subclass.Then, in step S24 place, using the nearest subclass of the proper vector with corresponding scene image as the subclass belonging to corresponding scene image, and the numbering of this subclass is given the proper vector of this corresponding scene image.Then, in step S25 place, the numbering based on the subclass of the proper vector of corresponding scene image calls the sorter of the subclass of reference numeral.Finally, in step S26 place, adopt the subclass sorter called to detect in corresponding scene image whether there is object.Whether the concrete training aids that how to adopt exists object belong to prior art to detect in candidate image, does not therefore repeat at this.
Obviously, the robustness of sorter is the main aspect of the polytrope in the face of application scenarios.In the art, the sorter of off-line training is the most often used.From angle of statistics, the training sample kind collected is more, and recognition result is better.But by varied for the scene faced in practical application, therefore, cannot collect complete training sample set, this causes clarification of objective in applied environment to be also difficult to be fully set forth at all.In order to enable recognition system adapt to current application scene, conventional method is sample under a large amount of current scene of collected offline, marks the sorter that final off-line training is applicable to current scene.This is a loaded down with trivial details job consuming time, and sorter is very responsive for scene changes, is difficult to be applied in video monitoring.Therefore, the system possessing adaptive learning is very necessary.
Existing adaptive learning method normally collects high confidence level sample to upgrade sorter, and the classification results of these samples is determined by original off-line training sorter.This update algorithm can cause sorter overfitting and original classification device is difficult to the correct target identified under still correctly can not distinguishing current scene.
Summary of the invention
For this reason, the present invention proposes a kind of sorter training method and system of adaptive learning, the method and system based on the credible positive negative sample of time domain, spatial domain and semantic context information and other useful information, carry out real-time update original object recognition classifier by on-line automatic collection.The method and system can be widely used in field of video monitoring, possess stronger adaptability and practicality.
According to an aspect of the present invention, provide a kind of adaptive classifier training method, comprise: adopt the multiple image that current class device ONLINE RECOGNITION inputs continuously, and the image be identified directly is added to the Sample Storehouse of sorter as the positive negative sample of high confidence level; And impalpable for current class device image is exported as difficult sample, and for the context of described difficult sample, be identified as positive and negative authentic specimen, and give identify the certain weight of positive and negative authentic specimen after added to the Sample Storehouse of sorter.
According to adaptive classifier training method of the present invention, the described context for described difficult sample, be identified as credible positive negative sample, and give identify the certain weight of positive and negative authentic specimen after added to the Sample Storehouse of sorter step comprise: based on the time domain context of concrete difficult sample, gather the contextual positive and negative authentic specimen of time domain of all difficult samples, and calculate the probability that the contextual positive and negative authentic specimen of described time domain comprises target to be identified under time-domain information; Based on the spatial domain context of concrete difficult sample, gather the contextual positive and negative authentic specimen in spatial domain of all difficult samples, and calculate the probability that the contextual positive and negative authentic specimen in described spatial domain comprises target to be identified under spatial information (si); Based on the semantic context of concrete difficult sample, gather the positive and negative authentic specimen of the semantic context of all difficult samples, and the positive and negative authentic specimen calculating described semantic context comprises the probability of target to be identified under semantic information; And based on current class device, the impact of difficult sample and the difficult sample calculated are comprised under time-domain information, spatial information (si) and semantic information to the probability of target to be identified, calculate the joint probability of this difficult sample as upper positive and negative authentic specimen, and give each difficult sample weighted value being selected as positive and negative authentic specimen according to the joint probability calculated.
According to adaptive classifier training method of the present invention, also comprise: before initial use current class device, collect just existing positive and negative authentic specimen to create the training sample database of target to be identified; To each the sample extraction feature in training sample database; Train original target recognition classifier based on the feature set extracted thus obtain initial current class device.
According to adaptive classifier training method of the present invention, described in be extracted feature and comprise: histogram of gradients feature or local binary patterns feature.
According to adaptive classifier training method of the present invention, described sorter comprises: support vector machine, Adaboost sorter or neural network.
According to another aspect of the present invention, provide a kind of sorter adaptive training system, comprise: simple sample resolution element, be the image comprising or do not comprise target to be identified based on current class device by inputted multiple image Direct Recognition, the image comprising or do not comprise target to be identified directly added to the Sample Storehouse of sorter as the positive negative sample of high confidence level; And difficult sample resolution element, receive current class device and fail the image of Direct Recognition as difficult sample, and for the context of described difficult sample, be identified as positive and negative authentic specimen, and give identify the certain weight of positive and negative authentic specimen after added to the Sample Storehouse of sorter.
According to sorter adaptive training system of the present invention, described difficult sample resolution element comprises: time domain resolution element, based on the time domain context of concrete difficult sample, gather the contextual positive and negative authentic specimen of time domain of all difficult samples, and calculate the probability that the contextual positive and negative authentic specimen of described time domain comprises target to be identified under time-domain information; Spatial domain resolution element, based on the spatial domain context of concrete difficult sample, gathers the contextual positive and negative authentic specimen in spatial domain of all difficult samples, and calculates the probability that the contextual positive and negative authentic specimen in described spatial domain comprises target to be identified under spatial information (si); Semantic resolution element, based on the semantic context of concrete difficult sample, gather the positive and negative authentic specimen of the semantic context of all difficult samples, and the positive and negative authentic specimen calculating described semantic context comprises the probability of target to be identified under semantic information; And overall treatment unit, based on current class device, the impact of difficult sample and the difficult sample calculated are comprised under time-domain information, spatial information (si) and semantic information to the probability of target to be identified, calculate the joint probability of this difficult sample as upper positive and negative authentic specimen, and give each difficult sample weighted value being selected as positive and negative authentic specimen according to the joint probability calculated.
Adopt the present invention's above-mentioned sorter adaptive training method and system, can the believable positive and negative target sample of online acquisition and other useful information.Sample collection method of the present invention is a kind of sample automatic acquiring method of intelligence, it can add the impalpable sample of original classification device to promote the classification accuracy of sorter in current scene in conjunction with time domain, spatial domain and semantic context information, can avoid Expired Drugs simultaneously.And reduce by giving the different confidence level of different authentic specimens the hydraulic performance decline that error sample selects to bring.Based on authentic specimen, real-time update is carried out to original classification device, system can be made to possess stronger adaptivity and practicality, make every effort to obtain in the field of video monitoring of such as automobile assistant driving use widely.
Accompanying drawing explanation
By reading the detailed description of the following the preferred embodiments of the present invention considered by reference to the accompanying drawings, above and other target of the present invention, feature, advantage and technology and industrial significance will be understood better.
Shown in Fig. 1 is the process flow diagram of training original classification device in prior art.
Shown in Fig. 2 is based on the process flow diagram being carried out object detection in candidate image by the subclass sorter of training in prior art.
Shown in Fig. 3 is the process flow diagram of online adaptive classification device training method according to the embodiment of the present invention.
Shown in Fig. 4 is the process flow diagram of authentic specimen collection process according to the embodiment of the present invention.
Shown in Fig. 5 is schematic diagram according to the contextual example in the spatial domain of the embodiment of the present invention.
Shown in Fig. 6 is the schematic diagram of the strength of semantic context according to 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 embodiment of the present invention.
Embodiment
In order to make those skilled in the art understand the present invention better, below in conjunction with the drawings and specific embodiments, the present invention is described in further detail.
According to embodiments of the invention, the invention reside on the basis of above-mentioned Fig. 1 and Fig. 2, the training of the sorter in recognition system is improved.Shown in Fig. 3 is the overview flow chart of the method for adaptive training sorter according to the embodiment of the present invention.
As shown in Figure 3, first, in step S31 place, current class device acquiring unit 710 is adopted, for calling the current sorter 30 used.Current class device may be adopt any prior art to obtain original classification device, also can be the adaptive classifier adopting sorter training method of the present invention to obtain.Subsequently, in step S32 place, sample collection unit 720 collects positive and negative authentic specimen from the input picture be identified, and is input in self study unit 730 by collected positive and negative authentic specimen.Finally, in step S33, the sample that self study unit 730 has based on current class device and obtain positive and negative authentic specimen from sample collection unit 720, performs self-learning strategy and upgrades current class device, thus current class device is adaptive to currently be identified environment.
Under current scene, after system collects the authentic specimen of some automatically, adaptive classifier will real-time update once.That is, after the authentic specimen collecting a given data, repeat above-mentioned steps S31-S33 once.
Here, S
learnand S
otherrepresent the influence value of useful information under current scene.S
learnthe influence value of the credible positive negative sample collected under representing current scene.S
otherrepresent the influence value of other useful information under current scene.Here " other useful information " is not only the useful information coming from frame of video can also be the useful information obtained from other channel, the GPS positioning result of such as target.S
learnand S
othercomplementary impact.
with
the influence value of credible positive negative sample respectively.Following formula gives the output recognition result of adaptive classifier:
S＝(1-γ)S
classifier+γ(S
learn+S
other)
γ is the weight of study.Final categorised decision result is as follows:
T is the threshold value of target identification.
Below the application example of adaptive targets sorting algorithm on visual dictionary model.Visual dictionary model can describe target signature preferably, has higher recognition accuracy.View-based access control model dictionary model algorithm concrete steps comprise: feature extraction, visual dictionary generate, the model generation of visual dictionary coupling, view-based access control model dictionary.After obtaining authentic specimen according to the inventive method, first extract unique point and the description thereof of these samples.Find the high frequency entry C of positive and negative authentic specimen
k, and to each high frequency entry selected with weights omega (C
k), the frequency that this weight occurs in authentic specimen to corresponding entry is directly proportional.Under the model, S is expressed by the unique point in region to be identified and the similarity of high frequency visual entry selected
learn.Concrete account form is as follows:
x＝+or-
Here, C
kit is the kth high frequency entry in x class (x represents positive sample object or negative sample non-targeted).F
pit is the unique point in region to be identified.Similarity is common similarity measurement strategy.
Fig. 4 is the process flow diagram gathering authentic specimen according to the sample collection unit of the embodiment of the present invention.As shown in Figure 4, first in step S3210 place, the sample reception unit 7205 in sample collection unit 720 to receive from current class device through current class device carry out the view data that identified.These include three class view data through current class device identification view data: the first kind is identified as including the image being identified target, are referred to as the positive sample with high confidence level at this; Equations of The Second Kind is identified not containing the image being identified target, is referred to as the negative sample with high execution degree at this; 3rd class is adopt current class device to be difficult to correctly identify whether to include the image being identified target, referred to herein as " difficult sample ".In existing online adaptive classification device, usually by the first kind and the Equations of The Second Kind sample collection Sample Storehouse to sorter, thus upgrade its Sample Storehouse and training classifier.The sorter adopted after training in this way more will be difficult to identification the 3rd class image, thus cause Expired Drugs, and the robustness of the sorter being also is deteriorated.For this reason, the present invention make use of the 3rd class sample when self-adaptation collects the sample of sorter online, instead of abandons the 3rd class sample.
For this reason, the present invention is in step S3220 place, sample reception unit 7205 simple sample resolution element 7210 in sample collection unit 720, based on the classification results of current class device, differentiates the positive negative sample with high confidence level, i.e. the above-described first kind and Equations of The Second Kind sample.These positive negative samples with high confidence level are added directly to the Sample Storehouse of sorter.
Subsequently, in step S3230 place, the time domain resolution element 7230 in the difficult sample resolution element 7220 of sample reception unit 7205 simple sample resolution element 7210 performs resolution to the difficult sample that all current class devices are not told in time domain.Specifically, whether the context authentic specimen in time based on inputted difficult sample sentences difficult disconnected sample credible, namely, if have multiframe to comprise in the frame number of the predetermined quantity of current class device within a period of time be identified target, so between these frame numbers being identified as comprising target unrecognized go out the two field picture of target then must include the target (the positive authentic specimen of high confidence level) be identified.Therefore, if certain difficult sample frame is identified as comprising between the positive authentic specimen of multiframe of target at these, then these by current class device be difficult to the image (difficult sample) that identifies then based on its place successive frame time domain specification (time domain context) and thought by time domain resolution element 7230 and belong to positive authentic specimen (comprising the sample being identified target), and give its certain time domain context probability.Equally, if have multiframe not comprise in the frame number of the predetermined quantity of current class device within a period of time be identified target (high confidence level bears authentic specimen), so these two field pictures being identified as the identified target comprised between the frame number of target then must not include the target be identified yet.Therefore, bear between authentic specimen if certain difficult sample frame is not recognized as at these multiframe comprising target, then these by current class device be difficult to the image (difficult sample) that identifies then based on its place successive frame time domain specification (time domain context) and thought by time domain resolution element 7230 and belong to negative authentic specimen (not comprising the sample being identified target), and give its certain time domain context probability.
As mentioned above, in object recognition system, time domain contextual information movable information represents.Due to the change such as visual angle, attitude, some target individual, in Continuous Tracking process, can be correctly validated in partial frame image, but can not be identified always.Can assert that these recognition results belong to a target by tracking results.Therefore, movable information can by the difficult target solving this type.
Such as, a target has correctly been followed the tracks of N frame, calculates its Probability p (o=x|o, Temporal) based on time domain contextual information to be:
Wherein, o
tthe observation of target at t frame,
the N number of observation of target in t-N to t frame.Described observation is exactly the state that target exists, o
t=x represents that the existence that target is present or feature are judged as classification x.
Meanwhile, in step S3240 place, the spatial domain resolution element 7240 in difficult sample resolution element 7220 performs resolution to the difficult sample that all current class devices are not told on spatial domain.Spatial domain contextual information parallax information represents.Specifically, whether the context spatially based on inputted difficult sample carrys out judgement sample credible, that is, owing to can be obtained the real space size of certain target by disparity map, such as, and its information such as true altitude, width.If this spatial information be identified target there is high similarity, usually can be considered to this and in its spatial domain context, be there is certain confidence level by the impalpable two field picture of current class device.Its Probability p (o=x|o, Spatial) based on spatial domain contextual information can be calculated according to the size of target.
Equally, in step S3250 place, the semantic resolution element 7250 in difficult sample resolution element 7220 performs identification to the difficult sample that all current class devices are not told in semantic context.Compared to time domain, spatial information (si), semantic context information is a kind of understanding information of high level.It is usually a kind of general knowledge or rule, and the difficult negative sample that this information fails to identify target for identification current class device is helpful.Semantic context information is usually relevant with the characteristic of other target occurred in scene (non-target to be identified) simultaneously, such as, automobile assistant driving system needs to detect from the road picture of shooting to identify pedestrian target, but railing etc. is often known for pedestrian like people's object by original classification device by mistake, and these objects often possess the feature such as texture, shape, color of its uniqueness, such as, like the target of people in the road shown in Fig. 5.Whether the occasion that semantic context information also comprises target to be identified appearance is reasonable, and such as, pedestrian can not swim in the air.As shown in Fig. 6 thick dashed line search box, the humanoid mark wherein on crossing stop-light may be identified as humanoid subject.But this humanoid subject is floating skyborne.Therefore, according to this semantic context information, can get rid of this humanoid subject is not object to be identified, therefore by negative sample that this specimen discerning is in difficult sample.Therefore, this type of difficult negative sample can be gathered by semantic understanding.The result of this resolution simply represents probability by the result of 0 and 1.
Subsequently, in step S3260 place, overall treatment unit 7260 in difficult sample resolution element 7220 is for the difficult sample be resolved, consider current class device, time domain context, spatial domain context and semantic context to the impact of its confidence level, obtain the final probable value of each difficult sample.Overall treatment unit 7260, for obtaining based on contextual authentic specimen in conjunction with above-mentioned three class contextual informations and sorter recognition result, need meet following joint probability formula:
p(x|o,Temporal,Spatial,Semantic,Classifier)
Wherein, x is recognition result, and be whole under observation o, the result in conjunction with Temporal, Spatial, Semantic, Classifier obtains, for being a probability; O is the observation of sample.Temporal, Spatial, Semantic, Classifier are respectively the impact of time domain, spatial domain, semanteme and sorter, namely final differentiation probability.
By the probable value that above-mentioned joint probability formula calculates, give the certain weight of each authentic specimen in all difficult samples.For the authentic specimen in outputted to sorter, its weighted value is directly proportional to calculated joint probability.
Shown in Fig. 7 is the concrete configuration figure of online adaptive classification device training system according to the embodiment of the present invention.As shown in Figure 7, described figure online adaptive classification device training system 700 of the present invention comprises: current class device acquiring unit 710, sample collection unit 720 and self study unit 730.Described sample collection unit 720 comprises sample reception unit 7205, simple sample resolution element 7210 and difficult sample resolution element 7220.Described difficult sample resolution element 7220 comprises time domain resolution element, 7230 spatial domain resolution elements 7240, semantic resolution element 7250 and overall treatment unit 7260.
By the further identifying of difficult sample described in above-mentioned steps, collect authentic specimen more comprehensively, make the Sample Storehouse of sorter more close to the truth of sorter institute applied environment.The accuracy rate of the identification target of same algorithm in current scene can be promoted.And, owing to not only adding high confidence level sample in the Sample Storehouse of sorter, also add the authentic specimen in difficult sample, therefore, it is possible to effectively solve Expired Drugs common in self study process.Meanwhile, by giving the different weight of authentic specimen selected according to different contextual informations and being applied to during sorter upgrades, the impact that error sample is brought system performance can effectively be reduced.
Below ultimate principle of the present invention is described in conjunction with specific embodiments, but, it is to be noted, for those of ordinary skill in the art, whole or any step or the parts of method and apparatus of the present invention can be understood, can in the network of any calculation element (comprising processor, storage medium etc.) or calculation element, realized with hardware, firmware, software or their combination, this is that those of ordinary skill in the art use their basic programming skill just can realize when having read explanation of the present invention.
Therefore, object of the present invention can also be realized by an operation program or batch processing on any calculation element.Described calculation element can be known fexible unit.Therefore, object of the present invention also can realize only by the program product of providing package containing the program code realizing described method or device.That is, such program product also forms the present invention, and the storage medium storing such program product also forms the present invention.Obviously, described storage medium can be any storage medium developed in any known storage medium or future.
Also it is pointed out that in apparatus and method of the present invention, obviously, each parts or each step can decompose and/or reconfigure.These decompose and/or reconfigure and should be considered as equivalents of the present invention.Further, the step performing above-mentioned series of processes can order naturally following the instructions perform in chronological order, but does not need necessarily to perform according to time sequencing.Some step can walk abreast or perform independently of one another.
Above-mentioned embodiment, does not form limiting the scope of the invention.It is to be understood that depend on designing requirement and other factors, various amendment, combination, sub-portfolio can be there is and substitute in those skilled in the art.Any amendment done within the spirit and principles in the present invention, equivalent replacement and improvement etc., all should be included within scope.
Claims (8)
1. a sorter adaptive training method, comprising:
Adopt the multiple image that current class device ONLINE RECOGNITION inputs continuously, and the image be identified directly is added to the Sample Storehouse of sorter as the positive negative sample of high confidence level; And
Impalpable for current class device image is exported as difficult sample, and for the context of described difficult sample, is identified as positive and negative authentic specimen, and give identify the certain weight of positive and negative authentic specimen after added to the Sample Storehouse of sorter.
2. sorter adaptive training method according to claim 1, the described context for described difficult sample, be identified as credible positive negative sample, and give identify the certain weight of positive and negative authentic specimen after added to the Sample Storehouse of sorter step comprise:
Based on the time domain context of concrete difficult sample, gather the contextual positive and negative authentic specimen of time domain of all difficult samples, and calculate the probability that the contextual positive and negative authentic specimen of described time domain comprises target to be identified under time-domain information;
Based on the spatial domain context of concrete difficult sample, gather the contextual positive and negative authentic specimen in spatial domain of all difficult samples, and calculate the probability that the contextual positive and negative authentic specimen in described spatial domain comprises target to be identified under spatial information (si);
Based on the semantic context of concrete difficult sample, gather the positive and negative authentic specimen of the semantic context of all difficult samples, and the positive and negative authentic specimen calculating described semantic context comprises the probability of target to be identified under semantic information; And
Based on current class device, the impact of difficult sample and the difficult sample calculated are comprised under time-domain information, spatial information (si) and semantic information to the probability of target to be identified, calculate the joint probability of this difficult sample as upper positive and negative authentic specimen, and give each difficult sample weighted value being selected as positive and negative authentic specimen according to the joint probability calculated.
3. sorter adaptive training method according to claim 1 and 2, also comprises:
Before initial use current class device, collect just existing positive and negative authentic specimen to create the training sample database of target to be identified;
To each the sample extraction feature in training sample database;
Train original target recognition classifier based on the feature set extracted thus obtain initial current class device.
4. sorter adaptive training method according to claim 3, described in be extracted feature and comprise: histogram of gradients feature or local binary patterns feature.
5. sorter adaptive training method according to claim 3, described sorter comprises: support vector machine, Adaboost sorter or neural network.
6. a sorter adaptive training system, comprising:
Simple sample resolution element, be the image comprising or do not comprise target to be identified based on current class device by inputted multiple image Direct Recognition, the image comprising or do not comprise target to be identified directly added to the Sample Storehouse of sorter as the positive negative sample of high confidence level; And
Difficult sample resolution element, receive current class device and fail the image of Direct Recognition as difficult sample, and for the context of described difficult sample, be identified as positive and negative authentic specimen, and give identify the certain weight of positive and negative authentic specimen after added to the Sample Storehouse of sorter.
7. sorter adaptive training system according to claim 6, described difficult sample resolution element comprises:
Time domain resolution element, based on the time domain context of concrete difficult sample, gathers the contextual positive and negative authentic specimen of time domain of all difficult samples, and calculates the probability that the contextual positive and negative authentic specimen of described time domain comprises target to be identified under time-domain information;
Spatial domain resolution element, based on the spatial domain context of concrete difficult sample, gathers the contextual positive and negative authentic specimen in spatial domain of all difficult samples, and calculates the probability that the contextual positive and negative authentic specimen in described spatial domain comprises target to be identified under spatial information (si);
Semantic resolution element, based on the semantic context of concrete difficult sample, gather the positive and negative authentic specimen of the semantic context of all difficult samples, and the positive and negative authentic specimen calculating described semantic context comprises the probability of target to be identified under semantic information; And
Overall treatment unit, based on current class device, the impact of difficult sample and the difficult sample calculated are comprised under time-domain information, spatial information (si) and semantic information to the probability of target to be identified, calculate the joint probability of this difficult sample as upper positive and negative authentic specimen, and give each difficult sample weighted value being selected as positive and negative authentic specimen according to the joint probability calculated.
8. the sorter adaptive training system according to claim 6 or 7, described sorter comprises: support vector machine, Adaboost sorter 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 true CN105224947A (en) | 2016-01-06 |
| CN105224947B 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) |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105974219A (en) * | 2016-04-11 | 2016-09-28 | 中国电力科学研究院 | Energy-saving electrical appliance load type classification and identification method |
| CN106022392A (en) * | 2016-06-02 | 2016-10-12 | 华南理工大学 | Deep neural network sample automatic accepting and rejecting training method |
| 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 |
| 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 |
| 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 |
| CN107832780A (en) * | 2017-10-17 | 2018-03-23 | 北京木业邦科技有限公司 | Low confidence sample processing method and system are sorted based on artificial intelligence plank |
| 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 |
| CN110288007A (en) * | 2019-06-05 | 2019-09-27 | 北京三快在线科技有限公司 | The method, apparatus and electronic equipment of data mark |
| CN111368926A (en) * | 2020-03-06 | 2020-07-03 | 腾讯科技（深圳）有限公司 | Image screening method, device and computer readable storage medium |
| CN111461243A (en) * | 2020-04-08 | 2020-07-28 | 中国医学科学院肿瘤医院 | Classification method, classification device, electronic equipment and computer-readable storage medium |
| CN113825879A (en) * | 2019-07-01 | 2021-12-21 | 株式会社小松制作所 | Method for manufacturing learned work classification estimation model, data for learning, method executed by computer, and system including work machine |
Citations (4)
| 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 |
| US20130148866A1 (en) * | 2008-06-27 | 2013-06-13 | Lockheed Martin Corporation | Assessing biometric sample quality using wavelets and a boosted classifier |
| 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 |
-
2014
- 2014-06-06 CN CN201410250540.5A patent/CN105224947B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20130148866A1 (en) * | 2008-06-27 | 2013-06-13 | Lockheed Martin Corporation | Assessing biometric sample quality using wavelets and a boosted classifier |
| 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 |
Cited By (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105974219A (en) * | 2016-04-11 | 2016-09-28 | 中国电力科学研究院 | Energy-saving electrical appliance load type classification and identification method |
| 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 |
| CN106022392A (en) * | 2016-06-02 | 2016-10-12 | 华南理工大学 | Deep neural network sample automatic accepting and rejecting training method |
| 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 |
| 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 |
| 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 |
| CN107832780A (en) * | 2017-10-17 | 2018-03-23 | 北京木业邦科技有限公司 | Low confidence sample processing method and system are sorted based on artificial intelligence plank |
| 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 |
| CN110288007A (en) * | 2019-06-05 | 2019-09-27 | 北京三快在线科技有限公司 | The method, apparatus and electronic equipment of data mark |
| 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 |
| CN113825879A (en) * | 2019-07-01 | 2021-12-21 | 株式会社小松制作所 | Method for manufacturing learned work classification estimation model, data for learning, method executed by computer, and system including work machine |
| CN111368926A (en) * | 2020-03-06 | 2020-07-03 | 腾讯科技（深圳）有限公司 | Image screening method, device and computer readable storage medium |
| CN111461243A (en) * | 2020-04-08 | 2020-07-28 | 中国医学科学院肿瘤医院 | Classification method, classification device, electronic equipment and computer-readable storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105224947B (en) | 2018-11-13 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105224947A (en) | Sorter training method and system | |
| CN112883819B (en) | Multi-target tracking method, device, system and computer readable storage medium | |
| Cap et al. | A deep learning approach for on-site plant leaf detection | |
| CN109558823B (en) | Vehicle identification method and system for searching images by images | |
| Meuter et al. | A decision fusion and reasoning module for a traffic sign recognition system | |
| CN104200495B (en) | A kind of multi-object tracking method in video monitoring | |
| CN104615986B (en) | The method that pedestrian detection is carried out to the video image of scene changes using multi-detector | |
| CN111161315B (en) | Multi-target tracking method and system based on graph neural network | |
| CN103530600B (en) | Licence plate recognition method under complex illumination and system | |
| CN104200237A (en) | High speed automatic multi-target tracking method based on coring relevant filtering | |
| CN109800624A (en) | A kind of multi-object tracking method identified again based on pedestrian | |
| JP2017156886A (en) | Device of tracking object taking similarity degree between images into consideration, program thereof and method thereof | |
| CN110334703B (en) | Ship detection and identification method in day and night image | |
| CN109190544B (en) | Human identity recognition method based on sequence depth image | |
| KR20200061118A (en) | Tracking method and system multi-object in video | |
| CN112257799A (en) | Method, system and device for detecting household garbage target | |
| CN103593679A (en) | Visual human-hand tracking method based on online machine learning | |
| CN107133970A (en) | Online multi-object tracking method and device based on movable information | |
| CN110688883A (en) | Vehicle and pedestrian detection method and device | |
| CN110781785A (en) | Traffic scene pedestrian detection method improved based on fast RCNN algorithm | |
| CN112613668A (en) | Scenic spot dangerous area management and control method based on artificial intelligence | |
| CN109636834A (en) | Video frequency vehicle target tracking algorism based on TLD innovatory algorithm | |
| CN104778699A (en) | Adaptive object feature tracking method | |
| CN110309785B (en) | Blind guiding robot control method based on image recognition technology | |
| Zhang et al. | Vehicle detection and tracking in remote sensing satellite vidio based on dynamic association |
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 |