CN102208009A - Classifier and classification method - Google Patents
Classifier and classification method Download PDFInfo
- Publication number
- CN102208009A CN102208009A CN2010101376907A CN201010137690A CN102208009A CN 102208009 A CN102208009 A CN 102208009A CN 2010101376907 A CN2010101376907 A CN 2010101376907A CN 201010137690 A CN201010137690 A CN 201010137690A CN 102208009 A CN102208009 A CN 102208009A
- Authority
- CN
- China
- Prior art keywords
- classifier
- sub
- confidence
- sample
- degree
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Abstract
The invention relates to a classifier and a classification method, wherein a plurality of sub-classifiers are used, a sample acquires a corresponding confidence through any one of the sub-classifiers and the confidence gradually increases along with the rank of the sub-classifier from the initial confidence of the sample. According to the invention, the classifier can have a better differentiation of confidence towards different samples.
Description
Technical field
The application relates generally to the target classification technology, i.e. sorter and sorting technique.Particularly, the application relates to a kind of cascade classifier, and a kind of cascade sort method.
Background technology
P.Viola and M.Jones have proposed the sorter of cascade system in article Rapid Object Detection using a BoostedCascade of Simple Features (CVPR 2001), admitted widely and use.Produced the cascade classifier of many mutation afterwards again, the Cluster Boosted Tree Classifier for Multi-view of B.Wu and R.Nevatia for example, Multi-PoseObject Detection (ICCV 2007), and C.Huang, H.Ai, the Vector Boosting for Rotation Invariant Multi-View Face Detection (ICCV2005) of Y.Li and S.Lao.More than three pieces of documents all the incorporated by here merge in the present specification.
As shown in Figure 1, no matter be which kind of cascade classifier, all possess following characteristics: a) be divided into the level of a plurality of series connection, for example C
1To C
N B) input sample 102 is verified step by step, as long as by arbitrary number of level refusal (arrow that makes progress from sub-classifier among Fig. 1), then this sample is differentiated and is negative sample (output among Fig. 1 " 0 "); C) have only by whole level sorters differentiate for positive sample by final differentiate be positive sample (output among Fig. 1 "+1 ").Therefore cascade classifier has only two kinds of output+1 and 0 (perhaps-1, being expressed as 0 among Fig. 1).
Summary of the invention
Provided hereinafter about brief overview of the present invention, so that basic comprehension about some aspect of the present invention is provided.Should be appreciated that this general introduction is not about exhaustive general introduction of the present invention.It is not that intention is determined key of the present invention or pith, neither be intended to limit scope of the present invention.Its purpose only is to provide some notion with the form of simplifying, with this as the preorder in greater detail of argumentation after a while.
Generally, the application carries out the progression of degree of confidence at having only sample just to be differentiated for positive problem has designed by making up a plurality of decision level by whole grades in traditional cascade classifier structure, makes sorter can have better degree of confidence to distinguish to different samples.
According to a kind of embodiment that the application provided, a kind of cascade classifier is provided, the sub-classifier that comprises a plurality of cascades, wherein, sample is by wherein obtaining corresponding degree of confidence after arbitrary sub-classifier, this degree of confidence increases with the sub-classifier progression that passes through gradually from the initial degree of confidence of sample.
According to the another kind of embodiment that the application provided, a kind of sorting technique is provided, comprise: make the sub-classifier of sample by a plurality of cascades, wherein, sample is by wherein obtaining corresponding degree of confidence after arbitrary sub-classifier, this degree of confidence increases with the sub-classifier progression that passes through gradually from the initial degree of confidence of sample.
According to the foregoing description, can make cascade classifier finer differentiation be arranged by progression formula confidence calculations to different samples, can obtain finer degree of confidence output, rather than simple " right and wrong ".
In addition, the application's embodiment also provides the computer program that can carry out on computers with realization said method or device, and the computer-readable medium that stores or transmit described computer program.
Description of drawings
With reference to below in conjunction with the explanation of accompanying drawing, can understand above and other purpose of the present invention, characteristics and advantage more easily to the embodiment of the invention.In the accompanying drawings, technical characterictic or parts identical or correspondence will adopt identical or corresponding Reference numeral to represent.Described accompanying drawing comprises in this manual and forms the part of this instructions together with following detailed description, and is used for further illustrating the preferred embodiments of the present invention and explains principle and advantage of the present invention.In the accompanying drawings:
Fig. 1 is the synoptic diagram at the existing cascade classifier that illustrated of " background technology " part;
Fig. 2 is the synoptic diagram that can be used in the example of the computing equipment 200 of realizing various embodiments of the present invention;
Fig. 3 is the synoptic diagram according to the cascade classifier of one embodiment of the present invention;
Fig. 4 is the synoptic diagram according to the cascade classifier of another embodiment of the invention;
Fig. 5 is the synoptic diagram according to the cascade classifier of another embodiment of the invention;
Fig. 6 is the synoptic diagram according to the cascade classifier of another embodiment of the invention;
Fig. 7 is the synoptic diagram according to the cascade classifier of another embodiment of the invention;
Fig. 8 is the synoptic diagram according to the cascade classifier of another embodiment of the invention;
Fig. 9 is the process flow diagram according to the sorting technique of one embodiment of the present invention;
Figure 10 is the process flow diagram according to the sorting technique of another embodiment of the invention.
Embodiment
To be described one exemplary embodiment of the present invention in conjunction with the accompanying drawings hereinafter.For clarity and conciseness, all features of actual embodiment are not described in instructions.Yet, should understand, in the process of any this practical embodiments of exploitation, must make a lot of decisions specific to embodiment, so that realize developer's objectives, for example, meet and system and professional those relevant restrictive conditions, and these restrictive conditions may change to some extent along with the difference of embodiment.In addition, might be very complicated and time-consuming though will also be appreciated that development, concerning the those skilled in the art that have benefited from present disclosure, this development only is customary task.
At this, what also need to illustrate a bit is, for fear of having blured the present invention because of unnecessary details, only show in the accompanying drawings with according to closely-related apparatus structure of the solution of the present invention and/or treatment step, and omitted other details little with relation of the present invention.
At first see Fig. 2, illustrate can be used in and realize that various embodiments of the present invention comprise the structural representation of example of the computing equipment 200 of cascade classifier and sorting technique.
In Fig. 2, CPU (central processing unit) (CPU) 201 carries out various processing according to program stored among ROM (read-only memory) (ROM) 202 or from the program that storage area 208 is loaded into random-access memory (ram) 203.In RAM 203, also store data required when CPU 201 carries out various processing or the like as required.
Following parts are connected to input/output interface 205: importation 206 comprises keyboard, mouse or the like; Output 207 comprises display, such as cathode ray tube (CRT) display, LCD (LCD) or the like and loudspeaker or the like; Storage area 208 comprises hard disk or the like; With communications portion 209, comprise that network interface unit is such as LAN card, modulator-demodular unit or the like.Communications portion 209 is handled such as the Internet executive communication via network.
As required, driver 210 also is connected to input/output interface 205.Detachable media 211 is installed on the driver 210 as required such as disk, CD, magneto-optic disk, semiconductor memory or the like, makes the computer program of therefrom reading be installed to as required in the storage area 208.
Can from network such as the Internet or storage medium such as detachable media 211 installation procedure to computing equipment.
It will be understood by those of skill in the art that this storage medium is not limited to shown in Figure 2 wherein having program stored therein, distribute separately so that the detachable media 211 of program to be provided to the user with equipment.The example of detachable media 211 comprises disk (comprising floppy disk (registered trademark)), CD (comprising compact disc read-only memory (CD-ROM) and digital universal disc (DVD)), magneto-optic disk (comprising mini-disk (MD) (registered trademark)) and semiconductor memory.Perhaps, storage medium can be hard disk that comprises in ROM 202, the storage area 208 or the like, computer program stored wherein, and be distributed to the user with the equipment that comprises them.
First embodiment
According to first kind of embodiment that the application provided, a kind of cascade classifier is provided, the sub-classifier that comprises a plurality of cascades, wherein, sample is by wherein obtaining corresponding degree of confidence after arbitrary sub-classifier, this degree of confidence increases with the sub-classifier progression that passes through gradually from the initial degree of confidence of sample.Like this, cascade classifier can provide a real number confidence value to sample, rather than refusal or acceptance simply, thereby can be distinguished different samples better by different confidence values.
Fig. 3 illustrates a kind of concrete topology example of this embodiment, and wherein cascade classifier has multistage (be designated as the N level, N is the natural number greater than 1) sub-classifier C of cascade
i(i is for satisfying the natural number of 1＜=i＜=N).Among the figure, represent that from each sub-classifier arrow to the right sample passes through to export corresponding degree of confidence V behind this sub-classifier
iAnd sample enters the next stage sub-classifier.The arrow that makes progress from each sub-classifier represents that sample fails by this grade sub-classifier and export corresponding degree of confidence V
I-1In the present embodiment, these sub-classifier levels are configured to make input sample 102 by i level sub-classifier C
iThe degree of confidence V of Huo Deing afterwards
iSatisfy V
i＞V
I-1, V wherein
0Initial degree of confidence for sample.Obviously, if input sample 102 does not pass through certain one-level sub-classifier V
i, then import the degree of confidence V that sample obtains
I-1Be previous stage sub-classifier C
I-1The degree of confidence of being given.
Specifically, for example, if input sample 102 has passed through C
2And do not pass through C
3, then obtain degree of confidence V
2If input sample 102 has further passed through C
3, then obtain degree of confidence V
3, and V
3＞V
2Like this, by the cascade classifier of this structure, make it possible to give different degree of confidence to sample according to the input sub-classifier progression that sample passed through, thereby help more careful differentiation, help to realize more exactly for example purpose such as target detection sample.
According to present embodiment, the degree of confidence of sub-classifiers at different levels increases progressively step by step.For incremental manner step by step, can be any way.For example, according to a kind of modification, can relevant with the performance of sub-classifier or irrespectively directly be set in confidence value by sample obtained behind certain sub-classifier.
According to another kind of modification, can be set to by the confidence value that sample obtained behind certain sub-classifier relevant with the performance of sub-classifier or irrespectively arithmetic increase progressively, for example whenever just increase by one greater than zero fixed value by a sub-classifier stage, perhaps increase by one be associated with the degree of confidence of sub-classifier itself or other index (for example false drop rate etc.) greater than zero value.
According to another modification, can be set to relevant by the confidence value that sample obtained behind certain sub-classifier or exponential increasing irrespectively with the performance of sub-classifier, for example whenever just increase a fixed value, perhaps increase a value that is associated with degree of confidence or other index (for example false drop rate etc.) of sub-classifier itself by a sub-classifier stage index.
According to another modification, can be set to by the confidence value that sample obtained behind certain sub-classifier relevant with the performance of sub-classifier or irrespectively multiplier increase progressively, for example whenever just multiply by one greater than 1 fixed value by a sub-classifier stage degree of confidence, perhaps multiply by one be associated with the degree of confidence of sub-classifier itself or other index (for example false drop rate etc.) greater than 1 value.
Above-mentioned various modification can be selected for use arbitrarily.But in order to obtain better effect, inner structure that can the zygote sorter determines suitably the incremental manner of degree of confidence.For example,, then can directly set confidence value, otherwise just calculate confidence value according to degree of confidence or other index of sub-classifier if sub-classifier does not provide confidence value or other index accurately.
For example, can determine described increment value according to the degree of confidence of the determined sample of sub-classifiers at different levels.For example, if being the merging that comprises a plurality of basic sorters, certain sub-classifier level strengthens (Boosting) sorter (article " Boosting Histograms ofOriented Gradients for Human Detection " (http://lear.inrialpes.fr/people/triggs/pubs/Dalal-cvpr05.pdf of people such as Marco Pedersoli for example, last visit on August 11st, 2009) disclosing a kind of such merging strengthens sorter, the content of the document merges among the application by quoting here), this sub-classifier can be relevant with the absolute value of difference between the threshold value of the output sum of described a plurality of basic sorters and this Boosting sorter to the increment value of degree of confidence.Particularly, the Boosting sorter is made of a plurality of more weak basic sorters, and each basic sorter has the output of oneself.Have only when the output sum of all basic sorters surpasses the threshold value of this Boosting sorter, just calculate by this Boosting sorter.Therefore, the absolute value of the difference of described basic sorter output sum and described threshold value has just reflected the degree of confidence of this Boosting sorter to sample.
Again for example, degree of confidence is that arithmetic increases progressively, exponential increasing or multiplier increases progressively, preferably consider when the design sub-classifier, each grade sub-classifier with respect to the lifting of the performance of previous stage sorter be that arithmetic increases progressively, exponential increasing or multiplier increases progressively.
As an example, generally all sub-classifiers at different levels are designed to that multiplier increases progressively.Therefore the degree of confidence of present embodiment also can adopt multiplier to increase progressively.Except direct setting greater than 1 increase progressively the coefficient, the performance of degree of confidence and sub-classifier is increased progressively explicitly.For example, on the principle, the false drop rate of sub-classifiers at different levels should progressively reduce, because the next stage sub-classifier is only handled the sample that has passed through the upper level sub-classifier when training.Therefore, can consider to use the increase progressively coefficient relevant with false drop rate.For example, can make this coefficient is f
I-1/ f
i, f wherein
iBe i level sub-classifier C
iFalse drop rate.Because the false drop rate of sub-classifiers at different levels should progressively reduce, so f
i＜f
I-1, and then this increases progressively coefficient greater than 1, so the degree of confidence of sample can progressively improve along with the increase of the sub-classifier progression that passes through.
Above-mentioned all modification for present embodiment can further improve.For example, the increment value that can increase progressively arithmetic, exponential increasing or multiplier increases progressively or increase progressively coefficient at different sub-classifier level weightings.The definite of weighting coefficient can have various modes.A kind of mode is to determine according to the position of sub-classifier level, and for example, past more downstream means that sample is that the probability of positive sample is high more, therefore can give higher weighting coefficient to the sub-classifier level in downstream.
Perhaps, can determine weighting coefficient to the different manifestations of concrete input sample according to each sub-classifier level itself.For example, can be with the degree of confidence of the determined sample of sub-classifier as weighting coefficient.At one more specifically in the middle of the modification, for example, if being the merging that comprises a plurality of basic sorters, certain sub-classifier level strengthens (Boosting) sorter, it is described then to be similar to preamble, and the absolute value that described weighting coefficient can be defined as the difference between the threshold value with the output sum of described a plurality of basic sorters and this Boosting sorter is relevant.
Obviously, each grade sub-classifier C
iCan be the sorter of any kind, and as is known to the person skilled in the art, each sub-classifier C
iItself can be constructed as the confidence value of direct acquisition sample, that is to say, the cascade classifier of present embodiment can be by being provided with each sub-classifier C in the manner described above
iDegree of confidence output (degree of confidence that comprises each sub-classifier self, and till this sub-classifier the comprehensive degree of confidence of whole cascade classifier) realize.On the other hand, one or more confidence calculations module (not shown) can be set in cascade classifier also, be used for based on each sub-classifier C
iOutput calculate the comprehensive degree of confidence of whole cascade classifier till this sub-classifier.Obviously, the some or all of of described confidence calculations module can be considered as each sub-classifier C respectively
iA part; Conversely, when not having the confidence calculations module, the confidence calculations function of each grade sub-classifier also can be considered as the confidence calculations module of each sub-classifier individually, perhaps synthetically is considered as the confidence calculations module of whole cascade classifier.
Need to prove that in addition the calculating of degree of confidence of input sample is that the waterfall sequence according to sub-classifier calculates, but this and do not mean that sub-classifier must carry out successively to the classification of input sample.On the contrary, sub-classifier can carry out according to any order the classification of same input sample, perhaps carries out concurrently, as long as the calculating of degree of confidence is carried out just passable according to aforementioned manner.That is to say that the waterfall sequence of cascade classifier is an order in logic, is not the order of actual execution.So, can decide according to the actual conditions of computing equipment to serial execution or executed in parallel.For example,, then can consider to carry out concurrently each sub-classifier, carry out successively the spent time serially to shorten if the computation capability of computing equipment is strong; If situation is opposite, then need to carry out serially, to reduce the expense of computing equipment.
Second embodiment
In the above in first embodiment of Miao Shuing, if the input sample by wherein certain one-level sub-classifier refusal, then the degree of confidence that is obtained is the degree of confidence that sub-classifier that previous stage is passed through is given, and no matter the classification results of the sub-classifier after this grade how.
But the sub-classifier of this refusal might be erroneous judgement.For example, might be except this grade sub-classifier refusal, the input sample can by all the other most levels or even all levels accept, that is to say that this sample should be positive sample.In this case, it is lower that the erroneous judgement of some levels just might cause whole cascade classifier to align the degree of confidence of sample, though thereby cause in processing subsequently, producing wrong result---according to the design of this embodiment, for cascade classifier itself, be the problem of degree of confidence height, but not non-like that just promptly negative judgement in the prior art.If especially this grade is earlier, more can have a strong impact on the value of degree of confidence.Under the high situation of the similarity of negative sample and positive sample, the possibility that this thing happens is higher.
Therefore, in the present embodiment, as shown in Figure 4, cascade classifier can be configured to ignore simply the sub-classifier level of not passing through.Particularly, for example, if sample, then makes V by i level sub-classifier
i=V
I-1Like this, the degree of confidence of previous stage can be delivered to next stage, is left in the basket and make when prime.
According to actual conditions, can set arbitrarily and to ignore the sub-classifier level how many levels are not passed through.A kind of extreme case is to ignore any sub-classifier level of not passing through.Can certainly set the progression of ignoring at most.For example cascade classifier can be configured to allow to ignore one-level, just when running into second sub-classifier level of not passing through, then no longer consider the result of sub-classifier level subsequently, output degree of confidence at this moment.Also cascade classifier can be configured to allow to ignore two-stage, just when running into the 3rd the sub-classifier level of not passing through, then no longer consider the result of sub-classifier level subsequently, output degree of confidence at this moment.Or the like.
What in the present embodiment, can realize by disposing each sub-classifier level or cascade classifier rightly the control of ignoring level.In addition, also can expect the control device (not shown) is set in cascade classifier, be used for the unsanctioned sub-classifier level of sample is counted, and control whether stop classification and output according to count value.Obviously, all or part of part that can be considered as each sub-classifier level respectively of described control device.
The 3rd embodiment
The inventor notices, in many application, the size of distinguishing confidence value in detail is only meaningful to positive sample and negative sample that those are judged by accident easily, and for classification accuracy high positive sample and negative sample, the detailed calculated confidence value is nonsensical, wastes computational resource on the contrary.
Therefore for the technical scheme of above-mentioned first embodiment and second embodiment, as shown in Figure 5 and Figure 6, can use according to first cascade classifier 504 of first embodiment or according to second cascade classifier 604 and 502 cascades of rough sort device of second embodiment.Fall positive sample of great majority and negative sample by these rough sort device 502 choosings.The degree of confidence that the rough sort device is given the input sample is the initial degree of confidence V0 of the cascade classifier of first embodiment or second embodiment.
In other words, can think that also present embodiment is in the cascade classifier of first embodiment and second embodiment, first order sub-classifier is configured to screen with higher accuracy rate the rough sort device (not shown) of positive sample of great majority and negative sample.
Further, as shown in Figure 7, described rough sort device also can be a third level connection sorter 702.Particularly, this third level connection sorter 702 can be the cascade classifier of any kind, especially can be traditional cascade classifier, that is, as long as its arbitrary grade of sub-classifier refusal sample, sample promptly by refuse.
It should be noted that the situation that only illustrates first cascade classifier 504 and 702 cascades of third level connection sorter among Fig. 7, in fact also can be second cascade classifier 604 and 702 cascades of third level connection sorter.
From another viewpoint, can think that also present embodiment is the cascade classifier that is made of multistage sub-classifier, wherein preceding M (M is the natural number greater than 1) level sub-classifier B
j(j is for satisfying the natural number of 1＜=j＜=M) promptly needs only sample by anyon sorter B by operate in conventional
jRefusal, then this sample is differentiated and is negative sample (output among Fig. 7 " 0 "); Back N level sub-classifier is by the first or second embodiment work.Specifically, for example, sample is at sub-classifier B
MCan only be in the past or passed through, or be rejected; And at sub-classifier C
1After, then sample obtains step by step with the sub-classifier C that is passed through
iQuantity increase and the degree of confidence that improves.
Pass through present embodiment, can screen positive sample of great majority and negative sample by preceding M level sub-classifier, and give more meticulous degree of confidence to positive sample and the negative sample obscured easily by back N level sub-classifier, so that distinguish positive sample and negative sample more accurately.
How selecting the M value will have tremendous influence to the performance of sorter, if the M value is excessive, then the performance of whole cascade classifier is near traditional cascade classifier; If the M value is very few, then there is a large amount of samples to carry out confidence calculations by back N level sub-classifier, efficient reduces.In concrete the application, the M value will be determined according to the characteristic of concrete application demand and sub-classifier, is that those skilled in the art accomplish after the basic thought of having grasped the application easily.A kind of mode of determining suitable M value is the performance curve according to traditional cascade classifier, and for example recall ratio-accuracy rate curve makes that the recall ratio of described preceding M level sub-classifier is higher, can allow most positive samples all pass through with M level sub-classifier before guaranteeing.
The 4th embodiment
The application's the respective embodiments described above can be applied to the technical field that any needs are classified, and comprise the target detection technique in image, video, the audio frequency.
Detect and sorting technique for image or objects in video, promptly in still image or video image, detect the plurality of target object and whether exist, and/or to distinguish be any in the plurality of target, and determine its position and size in image.For the target detection technique in the audio frequency, then be in audio stream, to detect whether to have target (comprise the speaker dependent, and/or specific content word etc. for example), and the position of target on time shaft, or the like.
When carrying out target detection, people often use window to travel through.For image, be to use spatial window to travel through; For audio frequency, be to use time window to travel through; For video, can travel through single-frame images usage space window, can on time shaft, travel through frame by frame simultaneously.Because the uncertainty of detected object yardstick can also be carried out multiple dimensioned traversal.Multiple dimensioned traversal can be undertaken by changing window size, also can carry out (for example pyramid diagram picture of design of graphics picture) by changing sample resolution.
In ergodic process, (size that for example detects target is greater than window owing to a variety of causes, perhaps the step-length of window traversal is less than the size that detects target, perhaps only cross over window edge just) because detect the position of target itself, may cause detecting target and cross over a plurality of windows, thereby make a plurality of windows positive response be arranged detecting target.In other words, if each window is exactly a sample, sorter just has the sample standard deviation of a plurality of adjacent (adjacent on the time, perhaps adjacent on the space, perhaps the two is all adjacent) and responds in other words.
For this reason, can merge (" fusion " in other words) to the classification results of the sample of adjacent window apertures (below be referred to as " adjacent sample "), obtain the position and the degree of confidence of couple candidate detection object.
For example, the Chinese patent application 200910161669.8 that is entitled as " method and apparatus that detects the target in the video image " that the applicant submitted on July 28th, 2009 has just been put down in writing the technology of multiple dimensioned traversal and fusion, and the full text of this application is herein incorporated by quoting here.
Need recognize that a plurality ofly having in the adjacent sample that is just responding that needs merge, the importance of each adjacent sample obviously is different, even may have false positive sample.
Therefore, aforementioned each embodiment of the application can be applied to the merging of the adjacent window apertures (adjacent sample) under the window traversal situation.In other words, as shown in Figure 8, the cascade classifier of present embodiment can also comprise merging device 806, be used for the classification results of a plurality of adjacent samples is merged, obtain candidate target (sample that big window is comprised that the adjacent sample after promptly merging, for example a plurality of adjacent window apertures constitute) and degree of confidence thereof.
Like this, for certain target in the sample, the application's second cascade classifier 604 can just be responded in a plurality of windows in its vicinity: not only the sample in these windows can be accepted, and the degree of confidence that obtains can be higher.Therefore, the fusion to adjacent a plurality of window results can obtain bigger comprehensive degree of confidence.
And it is difficult for some with target of detecting (may because noise, illumination condition bad etc.), the application's second cascade classifier 604 can obtain a plurality of more weak responses in its vicinity, can obtain certain degree of confidence after process multiwindow degree of confidence merges, thereby make this target to be detected.This is because object exists really, cause a little less than the response even external condition is bad, but adjacent a plurality of windows still all can just have and respond.And, then may directly refuse the most of windows in these adjacent window apertures for the sorter that has only 0/1 output.
Just responding for most falsenesses, is accidental because it majority occurs, so the window that only has minute quantity in its adjacent window apertures also can just access and responds.Therefore, merge by window, its comprehensive degree of confidence is still lower.
That is to say, in traversal and fusion process, because 604 pairs of samples of second cascade classifier provide different confidence values, rather than pass through simply or refusal, thereby can improve recall ratio to the bad object of condition, simultaneously get rid of falseness effectively and just respond, thereby improved recall ratio and accuracy rate simultaneously.
It should be noted that the combination that Fig. 8 only illustrates rough sort device 502, second cascade classifier 604 and merges device 806, but it is evident that aforementioned all embodiments and any modification thereof all can be combined with described merging device 806.
In addition, the degree of confidence of obtaining the candidate target after the merging from the degree of confidence of described a plurality of adjacent samples can be carried out with multiple mode.For example, calculate each adjacent sample degree of confidence and; Sue for peace or average perhaps with each degree of confidence normalization, and to the degree of confidence after the normalization; Perhaps calculate the mean value of each degree of confidence; Or the like.Should be understood that the method that degree of confidence is merged described herein only is exemplary, is not to be intended to the application is limited to this.In the application's scope, those of ordinary skill in the art can utilize various other suitable merging methods (for example compute histograms etc.) that described degree of confidence is merged.
In this application, so-called adjacent window apertures or adjacent sample are meant that detected candidate target (just having the window that is just responding) The corresponding area has adjacent center.Merge (seeing the Chinese patent application 200910161669.8 that is entitled as " method and apparatus that detects the target in the video image " that the applicant submitted on July 28th, 2009) for the interframe in the Video Detection, described zone also should have the close size of size.For example, for image and Video processing, adjacent each center that can refer to, center differs one or more pixels, those skilled in the art will appreciate that the dbjective state (as movement velocity, direction etc.) that pixel count can detect according to actual needs and decides.Here do not enumerate one by one.And for example, the close size that can refer to of size differs one or more pixels.Those skilled in the art will appreciate that the dbjective state (as target sizes, movement velocity etc.) that the pixel count that differs can detect according to actual needs here and decide.
The 5th embodiment
In the explanation of above cascade classifier to first embodiment to the, four embodiments, embodied the sorting technique of utilizing described cascade classifier to realize simultaneously.Therefore there is no need described sorting technique is carried out the detailed description of repetition at this, and only as follows to described sorting technique recapitulaion.
At first, according to aforementioned first embodiment, a kind of sorting technique is disclosed simultaneously, comprise: make the sub-classifier of sample by a plurality of cascades, wherein, sample is by wherein obtaining corresponding degree of confidence after arbitrary sub-classifier, and this degree of confidence increases with the sub-classifier progression that passes through gradually from the initial degree of confidence of sample.Like this, can provide a real number confidence value to sample, rather than refusal or acceptance simply, thereby can be distinguished different samples better by different confidence values.
As a concrete example of this sorting technique, can make the N level sub-classifier C of sample by cascade
1To C
N, N is the natural number greater than 1, wherein, sample is by i level sub-classifier C
iObtain degree of confidence V afterwards
i, wherein i is for satisfying the natural number of 1＜=i＜=N, V
iSatisfy V
i＞V
I-1, V
0Initial degree of confidence for sample.
Wherein, V
iWith respect to V
I-1Can increase progressively by arithmetic, can exponential increasing, also can multiplier increase progressively.For the incremental manner step by step of degree of confidence, can be any way.For example, can be relevant or irrespectively directly be set in confidence value by sample obtained behind certain sub-classifier with the performance (degree of confidence or other index such as false drop rate etc.) of sub-classifier, perhaps, can be relevant with the performance (degree of confidence or other index such as false drop rate etc.) of sub-classifier or irrespectively be set in by sample obtained behind certain sub-classifier confidence value greater than 0 increment value, exponential increasing value or greater than 1 increase progressively coefficient etc.For example, when at least one sub-classifier was the Boosting that comprises a plurality of basic sorters (merge and strengthen) sorter, this sub-classifier increases progressively coefficient to arithmetic increment value, exponential increasing value or the multiplier of degree of confidence can be relevant with the absolute value of difference between the threshold value of the output sum of described a plurality of basic sorters and this Boosting sorter.
As an example, it is also conceivable that and use the increase progressively coefficient relevant with the sub-classifier false drop rate.For example, can make this coefficient is f
I-1/ f
i, f wherein
iBe i level sub-classifier C
iFalse drop rate.Because the false drop rate of sub-classifiers at different levels should progressively reduce, so f
i＜f
I-1, and then this increases progressively coefficient greater than 1, so the degree of confidence of sample can progressively improve along with the increase of the sub-classifier progression that passes through.
Above-mentioned all modification for present embodiment can further improve.For example, the increment value that can increase progressively arithmetic, exponential increasing or multiplier increases progressively or increase progressively coefficient at different sub-classifier level weightings.The definite of weighting coefficient can have various modes.For example can determine weighting coefficient to the different manifestations (for example degree of confidence that sample is determined) of concrete input sample according to position or each sub-classifier level itself of sub-classifier level.In certain sub-classifier level is that the merging that comprises a plurality of basic sorters strengthens under the situation of (Boosting) sorter, it is described to be similar to preamble, and the absolute value that described weighting coefficient can be defined as the difference between the threshold value with the output sum of described a plurality of basic sorters and this Boosting sorter is relevant.
According to said method, can give more meticulous degree of confidence to sample.
According to aforementioned second embodiment, in above-mentioned sorting technique, can ignore the sub-classifier level of not passing through.Particularly, for example, if sample, then makes V by i level sub-classifier
i=V
I-1Like this, the degree of confidence of previous stage can be delivered to next stage, be left in the basket and make when prime, thus only avoid since the erroneous judgement of some levels just to cause whole cascade classifier to align the degree of confidence of sample lower, thereby cause in processing subsequently, producing wrong result.
According to actual conditions, can set arbitrarily and to ignore the sub-classifier level how many levels are not passed through.What can realize by disposing each sub-classifier level or cascade classifier rightly the control of ignoring level.In addition, also can expect the unsanctioned sub-classifier level of sample is counted, and control the classification results of the sub-classifier level after whether stopping to consider according to count value.Obviously, described counting and controlled step can be considered as the part of the sort operation of each sub-classifier level respectively; Perhaps conversely, each sub-classifier level can be considered as a single controlled step generally to the control of ignoring how many levels.
According to aforementioned the 3rd embodiment, the application has also proposed sorting technique that rough sort (step 906) is combined with the classification (step 910) of degree of confidence progression, as shown in Figure 9.Just before the various embodiments or modification of aforesaid sorting technique, can carry out rough sort earlier, fall high positive sample and the negative sample of most of accuracys rate with screening, only the sample of easy generation erroneous judgement is carried out the classification of degree of confidence progression, simultaneously guaranteed efficiency.
Further, described rough sort device also can use cascade classifier, especially can be traditional cascade classifier, that is, as long as its arbitrary grade of sub-classifier refusal sample, sample is promptly refused by whole cascade classifier.
According to aforementioned the 4th embodiment, the various embodiments and the modification of the application's above-mentioned sorting technique can be applied to the technical field that any needs are classified, and comprise the target detection technique in image, video, the audio frequency.Under the situation of using window traversal and integration technology, use the application's above-mentioned sorting technique especially favourable.
That is to say, as shown in figure 10, at first each sample is classified according to aforementioned various sorting techniques, it is the classification (step 910) of rough sort (step 906) and degree of confidence progression, preliminary classification result 1012 to a plurality of adjacent samples merges (step 1016) then, obtains candidate target and degree of confidence thereof.In Figure 10, rough sort step (step 906) dots, and is in order to illustrate that this step can not have yet.
Thereby, merge by multisample, can further widen the degree of confidence gap between bad couple candidate detection target of condition good couple candidate detection target, condition and the false couple candidate detection target, thereby can improve recall ratio to the bad object of condition, simultaneously get rid of falseness effectively and just respond, thereby improved recall ratio and accuracy rate simultaneously.
Above only to having carried out briefly bright with first embodiment to the, four embodiment respective classified methods.Its ins and outs can be further in detail with reference to the explanation of preamble to first embodiment to the, four embodiments itself.
Some embodiments to the application are described in detail above.To understand as those of ordinary skill in the art, whole or any steps or the parts of method and apparatus of the present invention, can be in the network of any computing equipment (comprising processor, storage medium etc.) or computing equipment, realized with hardware, firmware, software or their combination, this is that those of ordinary skills' their basic programming skill of utilization under the situation of understanding content of the present invention just can be realized, does not therefore need to specify at this.
In addition, it is evident that, when relating to possible peripheral operation in the superincumbent explanation, will use any display device and any input equipment, corresponding interface and the control program that link to each other with any computing equipment undoubtedly.Generally speaking, the hardware of the various operations in the related hardware in computing machine, computer system or the computer network, software and the realization preceding method of the present invention, firmware, software or their combination promptly constitute equipment of the present invention and each building block thereof.
Therefore, based on above-mentioned understanding, purpose of the present invention can also realize by program of operation or batch processing on any messaging device.Described messaging device can be known common apparatus.Therefore, purpose of the present invention also can be only by providing the program product that comprises the program code of realizing described method or equipment to realize.That is to say that such program product also constitutes the present invention, and the storage medium that stores such program product also constitutes the present invention.Obviously, described storage medium can be well known by persons skilled in the art, and perhaps the storage medium of any kind that is developed in the future includes but not limited to floppy disk, CD, magneto-optic disk, storage card, memory stick or the like.
In equipment of the present invention and method, obviously, after can decomposing, make up and/or decompose, each parts or each step reconfigure.These decomposition and/or reconfigure and to be considered as equivalents of the present invention.
The step that also it is pointed out that the above-mentioned series of processes of execution can order following the instructions naturally be carried out in chronological order, but does not need necessarily to carry out according to time sequencing.Some step can walk abreast or carry out independently of one another.
In addition, though above be an embodiment of an embodiment be described, be to be understood that each embodiment is not what isolate.Those skilled in the art obviously can understand after having read present specification, and the various technical characterictics that each embodiment comprised can combination in any between various embodiments, as long as not conflict between them.Certainly, all technical characterictics of mentioning in same embodiment each other also can combination in any, as long as their not conflicts each other.
At last, term " comprises ", " comprising " or its any other variant are intended to contain comprising of nonexcludability, thereby make and comprise that process, method, article or the equipment of a series of key elements not only comprise those key elements, but also comprise other key elements of clearly not listing, or also be included as this process, method, article or equipment intrinsic key element.In addition, do not having under the situation of more restrictions, the key element that limits by statement " comprising ... ", and be not precluded within process, method, article or the equipment that comprises described key element and also have other identical element.
Though be described with reference to the accompanying drawings embodiments of the present invention and advantage thereof, should be appreciated that embodiment described above just is used to illustrate the present invention, and be not construed as limiting the invention.For a person skilled in the art, can make various changes and modifications above-mentioned embodiment and do not deviate from the spirit and scope of the invention.Therefore, scope of the present invention is only limited by appended claim and equivalents thereof, can carry out various changes under not exceeding by the situation of the appended the spirit and scope of the present invention that claim limited, substitute and conversion.
Claims (24)
1. cascade classifier comprises the sub-classifier of a plurality of cascades, and wherein, sample is by wherein obtaining corresponding degree of confidence after arbitrary sub-classifier, and this degree of confidence increases with the sub-classifier progression that passes through gradually from the initial degree of confidence of sample.
2. cascade classifier as claimed in claim 1 comprises the N level sub-classifier C of cascade
1To C
N, N is the natural number greater than 1, wherein, sample is by obtaining degree of confidence V after the i level sub-classifier Ci
i, wherein i is for satisfying the natural number of 1＜=i＜=N, V
iSatisfy V
i＞V
I-1, V
0Initial degree of confidence for sample.
3. cascade classifier as claimed in claim 2, wherein, V
iWith respect to V
I-1Arithmetic increases progressively or exponential increasing.
4. cascade classifier as claimed in claim 3, wherein, at least one sub-classifier is that the merging that comprises a plurality of basic sorters strengthens sorter, and this sub-classifier is relevant with this absolute value that merges the difference between the threshold value that strengthens sorter with the output sum of described a plurality of basic sorters to the arithmetic increment value or the exponential increasing value of degree of confidence.
5. cascade classifier as claimed in claim 2, wherein, V
i=V
I-1* g (C
i), g (C
i) be sub-classifier C
iTo degree of confidence greater than 1 the coefficient that increases progressively.
6. cascade classifier as claimed in claim 5, wherein, g (C
i)=f
I-1/ f
i, f wherein
iBe i level sub-classifier C
iFalse drop rate, f
i＜f
I-1
7. cascade classifier as claimed in claim 6, wherein, to g (C
i) multiply by a weighting coefficient.
8. cascade classifier as claimed in claim 7, wherein, at least one sub-classifier is that the merging that comprises a plurality of basic sorters strengthens sorter, and described weighting coefficient is relevant with this absolute value that merges the difference between the threshold value that strengthens sorter with the output sum of described a plurality of basic sorters.
9. as the described cascade classifier of one of claim 2-8, wherein, if sample, then makes V by i level sub-classifier
i=V
I-1, perhaps with V
I-1Final degree of confidence output as sample.
10. as the described cascade classifier of one of claim 1-8, also comprise preposition rough sort device, the degree of confidence of the sample by this rough sort device is described initial degree of confidence.
11. cascade classifier as claimed in claim 10, wherein, described rough sort device is that multistage sub-classifier cascade constitutes, and wherein, as long as arbitrary grade of sub-classifier refusal sample of this rough sort device, sample is promptly by described rough sort device and described cascade classifier refusal.
12. cascade classifier as claimed in claim 11 also comprises the merging device, will the classification results of a plurality of adjacent samples be merged, and obtains candidate target and degree of confidence thereof.
13. a sorting technique comprises:
Make the sub-classifier of sample by a plurality of cascades, wherein, sample is by wherein obtaining corresponding degree of confidence after arbitrary sub-classifier, and this degree of confidence increases with the sub-classifier progression that passes through gradually from the initial degree of confidence of sample.
14. sorting technique as claimed in claim 13 comprises:
Make the N level sub-classifier C of sample by cascade
1To C
N, N is the natural number greater than 1, wherein, sample is by obtaining degree of confidence V after the i level sub-classifier Ci
i, wherein i is for satisfying the natural number of 1＜=i＜=N, V
iSatisfy V
i＞V
I-1, V
0Initial degree of confidence for sample.
15. sorting technique as claimed in claim 14, wherein, V
iWith respect to V
I-1Arithmetic increases progressively or exponential increasing.
16. sorting technique as claimed in claim 15, wherein, at least one sub-classifier is that the merging that comprises a plurality of basic sorters strengthens sorter, and this sub-classifier is relevant with this absolute value that merges the difference between the threshold value that strengthens sorter with the output sum of described a plurality of basic sorters to the arithmetic increment value or the exponential increasing value of degree of confidence.
17. sorting technique as claimed in claim 14, wherein, V
i=V
I-1* g (C
i), g (C
i) be sub-classifier C
iTo degree of confidence greater than 1 the coefficient that increases progressively.
18. sorting technique as claimed in claim 17, wherein, g (C
i)=f
I-1/ f
i, f wherein
iBe i level sub-classifier C
iFalse drop rate, f
i＜f
I-1
19. sorting technique as claimed in claim 18, wherein, to g (C
i) multiply by a weighting coefficient.
20. sorting technique as claimed in claim 19, wherein, at least one sub-classifier is that the merging that comprises a plurality of basic sorters strengthens sorter, and described weighting coefficient is relevant with this absolute value that merges the difference between the threshold value that strengthens sorter with the output sum of described a plurality of basic sorters.
21. as the described sorting technique of one of claim 14-20, wherein, if sample, then makes V by i level sub-classifier
i=V
I-1, perhaps with V
I-1Final degree of confidence output as sample.
22., also comprise as the described sorting technique of one of claim 13-20:
Make sample at first by preposition rough sort device, the degree of confidence of the sample by this rough sort device is described initial degree of confidence.
23. sorting technique as claimed in claim 22, wherein, described rough sort device is that multistage sub-classifier cascade constitutes, and wherein, as long as arbitrary grade of sub-classifier refusal sample of this rough sort device, sample is promptly by described rough sort device and described sorting technique refusal.
24. sorting technique as claimed in claim 23 also comprises: will the classification results of a plurality of adjacent samples be merged, obtain candidate target and degree of confidence thereof.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2010101376907A CN102208009A (en) | 2010-03-31 | 2010-03-31 | Classifier and classification method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2010101376907A CN102208009A (en) | 2010-03-31 | 2010-03-31 | Classifier and classification method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN102208009A true CN102208009A (en) | 2011-10-05 |
Family
ID=44696836
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN2010101376907A Pending CN102208009A (en) | 2010-03-31 | 2010-03-31 | Classifier and classification method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN102208009A (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104281569A (en) * | 2013-07-01 | 2015-01-14 | 富士通株式会社 | Building device and method, classifying device and method and electronic device |
| CN106407878A (en) * | 2016-06-28 | 2017-02-15 | 北京智芯原动科技有限公司 | Face detection method and device based on multiple classifiers |
| CN106778800A (en) * | 2016-11-14 | 2017-05-31 | 天津津航技术物理研究所 | A kind of AdaBoost cascade classifiers method for quick |
| CN108021925A (en) * | 2016-10-31 | 2018-05-11 | 北京君正集成电路股份有限公司 | A kind of detection method and equipment |
| CN109242165A (en) * | 2018-08-24 | 2019-01-18 | 蜜小蜂智慧（北京）科技有限公司 | A kind of model training and prediction technique and device based on model training |
| CN109740637A (en) * | 2018-12-12 | 2019-05-10 | 天津津航技术物理研究所 | The optimization method of training adaboost cascade classifier |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101226591A (en) * | 2008-01-31 | 2008-07-23 | 上海交通大学 | Personal identification method based on mobile phone pick-up head combining with human face recognition technique |
| US20080253664A1 (en) * | 2007-03-21 | 2008-10-16 | Ricoh Company, Ltd. | Object image detection method and object image detection device |
-
2010
- 2010-03-31 CN CN2010101376907A patent/CN102208009A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20080253664A1 (en) * | 2007-03-21 | 2008-10-16 | Ricoh Company, Ltd. | Object image detection method and object image detection device |
| CN101226591A (en) * | 2008-01-31 | 2008-07-23 | 上海交通大学 | Personal identification method based on mobile phone pick-up head combining with human face recognition technique |
Non-Patent Citations (2)
| Title |
|---|
| MARCO PEDERSOLI等: "Boosting Histograms Of Oriented Gradients for Human", 《HTTP://DIGITAL.CSIC.ES/BITSTREAM/10261/30381/BOOSTING%20HISTOGRAMS.PDF》, 31 December 2007 (2007-12-31), pages 1 - 6 * |
| 付强等: "用于手写汉字识别的级联MQDF分类器", 《清华大学学报( 自然科学版)》, vol. 48, no. 10, 31 October 2008 (2008-10-31), pages 1609 - 1612 * |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104281569A (en) * | 2013-07-01 | 2015-01-14 | 富士通株式会社 | Building device and method, classifying device and method and electronic device |
| CN104281569B (en) * | 2013-07-01 | 2017-08-01 | 富士通株式会社 | Construction device and method, sorter and method and electronic equipment |
| CN106407878A (en) * | 2016-06-28 | 2017-02-15 | 北京智芯原动科技有限公司 | Face detection method and device based on multiple classifiers |
| CN106407878B (en) * | 2016-06-28 | 2019-08-13 | 北京智芯原动科技有限公司 | Method for detecting human face and device based on multi-categorizer |
| CN108021925A (en) * | 2016-10-31 | 2018-05-11 | 北京君正集成电路股份有限公司 | A kind of detection method and equipment |
| CN106778800A (en) * | 2016-11-14 | 2017-05-31 | 天津津航技术物理研究所 | A kind of AdaBoost cascade classifiers method for quick |
| CN109242165A (en) * | 2018-08-24 | 2019-01-18 | 蜜小蜂智慧（北京）科技有限公司 | A kind of model training and prediction technique and device based on model training |
| CN109740637A (en) * | 2018-12-12 | 2019-05-10 | 天津津航技术物理研究所 | The optimization method of training adaboost cascade classifier |
| CN109740637B (en) * | 2018-12-12 | 2023-08-15 | 天津津航技术物理研究所 | Optimization method for training adaboost cascade classifier |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US11886799B2 (en) | Determining functional and descriptive elements of application images for intelligent screen automation | |
| Prasad et al. | Edge curvature and convexity based ellipse detection method | |
| CN102208009A (en) | Classifier and classification method | |
| Strandmark et al. | Hep-2 staining pattern classification | |
| US7519201B2 (en) | Detecting humans via their pose | |
| JP4997178B2 (en) | Object detection device | |
| KR20190123790A (en) | Extract data from electronic documents | |
| Zhou et al. | Loan default prediction on large imbalanced data using random forests | |
| CN103679185A (en) | Convolutional neural network classifier system as well as training method, classifying method and application thereof | |
| US20110085728A1 (en) | Detecting near duplicate images | |
| US20060222239A1 (en) | Systems and methods for detecting text | |
| CN101814147A (en) | Method for realizing classification of scene images | |
| CN102737186B (en) | Malicious file identification method, device and storage medium | |
| CN111126482A (en) | Remote sensing image automatic classification method based on multi-classifier cascade model | |
| CN103198303A (en) | Gender identification method based on facial image | |
| CN101968884A (en) | Method and device for detecting target in video image | |
| CN112733936A (en) | Recyclable garbage classification method based on image recognition | |
| CN113922985A (en) | Network intrusion detection method and system based on ensemble learning | |
| CN103366177A (en) | Object detection classifier generating method, object detection classifier generating apparatus, image object detection method and image object detection apparatus | |
| CN113807347A (en) | Kitchen waste impurity identification method based on target detection technology | |
| CN101996326A (en) | Device and method for detecting various classes of targets | |
| CN101923650A (en) | Random forest classification method and classifiers based on comparison mode | |
| CN110458203B (en) | Advertisement image material detection method | |
| WO2023082993A1 (en) | Information recommendation method, apparatus and system | |
| Rusinol et al. | CVC-UAB's Participation in the Flowchart Recognition Task of CLEF-IP 2012. |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C02 | Deemed withdrawal of patent application after publication (patent law 2001) | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20111005 |