CN104112131B - Method and device for generating training samples used for face detection - Google Patents
Method and device for generating training samples used for face detection Download PDFInfo
- Publication number
- CN104112131B CN104112131B CN201310137893.XA CN201310137893A CN104112131B CN 104112131 B CN104112131 B CN 104112131B CN 201310137893 A CN201310137893 A CN 201310137893A CN 104112131 B CN104112131 B CN 104112131B
- Authority
- CN
- China
- Prior art keywords
- face
- positive sample
- sample set
- face positive
- represent
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Landscapes
- Image Processing (AREA)
Abstract
The invention discloses a method and device for generating training samples used for face detection. The method includes the following steps that: an original face positive sample set and an original face negative sample set are obtained; image processing is performed on face positive samples in the original face positive sample set, and the processed face positive samples are added into the original face positive sample set, so that an intermediate face positive sample set can be obtained; random extraction and random weighting are performed on face positive samples in the intermediate face positive sample set, and the randomly extracted and weighted face positive samples are added into the intermediate face positive sample set, so that a final face positive sample set can be obtained; random extraction and bitwise AND logical operation are performed on face negative samples in the original face negative sample set, and the processed face negative samples are added into the original face negative sample set, and a final face negative sample set can be obtained; and the final face positive sample set and the final face negative sample set are adopted as training samples used for face detection. With the method and device provided by the technical scheme of the invention adopted, labor sources can be decreased, and the diversity of the training samples can be improved, and the accuracy of detection results can be ensured.
Description
Technical field
The present invention relates to field of computer technology, the generation method of espespecially a kind of training sample for Face datection and dress
Put.
Background technology
Face datection is mainly used in the pretreatment of recognition of face in actual applications, i.e., accurate calibration goes out face in the picture
Position and size.Human face detection tech has obtained good application in gate control system, intelligent monitor system at present.Separately
Outward, in notebook computer, also begin to use face recognition technology to log in as computer successively voucher.In recent years, in number
Also Face datection algorithm is integrated with camera and mobile phone, is supplied to user to use as new function.In such applications, face
Detection is all to play vital effect.
Due to facial image easily by ambient light, express one's feelings, block, the various factors such as age and attitude is affected, this allows for
Face datection becomes a complicated, challenging research topic.Through a large amount of scholar's years of researches, based on level link
The Adaboost algorithm of structure is considered as detection speed most fast and effect the best way.Adaboost algorithm is a kind of based on sample
The method of this study：In the training stage, the common trait of the great amount of samples to being input into is learnt and is concluded, and generates training sample
This, including face positive sample and face negative sample, face positive sample is the image of face, and face negative sample is background residing for face
Image；It is in detection-phase, whether consistent with training sample by the feature of analysis of the image, determine whether image is face
Image.It can be seen that, impact of the training sample to Face datection result is very big.
In general, training sample multiformity is better, and Detection results are better.At present, generally by artificial collection image
Mode obtain training sample, as training sample is required with preferable multiformity, then be accomplished by gathering substantial amounts of image,
Therefore substantial amounts of human resourcess can be wasted；And due to there are many features unavoidably very close to this is just in the image of collection
Reduce the multiformity of training sample, it is impossible to ensure the accuracy of face testing result.
The content of the invention
The embodiment of the present invention provides a kind of generation method of the training sample for Face datection and device, existing to solve
Somebody's work collection training sample mode caused by human resourcess' serious waste, training sample multiformity reduce and cannot
Ensure the problem of the accuracy of face testing result.
A kind of generation method of the training sample for Face datection, including：
Obtain first and the face positive sample of quantity is set as original face positive sample set, and obtain second set number
The face negative sample of amount is used as original face negative sample set；
Image procossing is carried out to the face positive sample in the original face positive sample set, and is added to the primitive man
In face positive sample set, middle face positive sample set is obtained；To the face positive sample in the middle face positive sample set
Randomly selected and random weighting, and be added in the middle face positive sample set, obtained final face positive sample collection
Close；And
Face negative sample in the original face negative sample set is randomly selected and step-by-step logical operationss, and is added
It is added in the original face negative sample set, obtains final face negative sample set；
Using final face positive sample set and final face negative sample set as the instruction for Face datection
Practice sample.
A kind of generating means of the training sample for Face datection, including：
Acquiring unit, sets the face positive sample of quantity as original face positive sample set for obtaining first, and
Obtain second the face negative sample of quantity is set as original face negative sample set；
Processing unit, for carrying out image procossing to the face positive sample in the original face positive sample set, and adds
It is added in the original face positive sample set, obtains middle face positive sample set；To the middle face positive sample set
In face positive sample randomly selected and random weighting, and be added in the middle face positive sample set, obtain most
Whole face positive sample set；And the face negative sample in the original face negative sample set is randomly selected and step-by-step
Logical operationss, and be added in the original face negative sample set, obtain final face negative sample set；
Signal generating unit, for using final face positive sample set and final face negative sample set as being used for
The training sample of Face datection.
The present invention has the beneficial effect that：
The generation method of the training sample for Face datection provided in an embodiment of the present invention and device, by obtaining first
The face positive sample of setting quantity is used as original face positive sample set, and obtains the second face negative sample work for setting quantity
For original face negative sample set；Image procossing is carried out to the face positive sample in the original face positive sample set, and is added
It is added in the original face positive sample set, obtains middle face positive sample set；To the middle face positive sample set
In face positive sample randomly selected and random weighting, and be added in the middle face positive sample set, obtain most
Whole face positive sample set；And the face negative sample in the original face negative sample set is randomly selected and step-by-step
Logical operationss, and be added in the original face negative sample set, obtain final face negative sample set；By the final people
The set of face positive sample and final face negative sample set are used as the training sample for Face datection.The program obtains setting
Then these face positive samples and face negative sample are carried out certain process by the face positive sample of quantity and face negative sample,
More face positive samples and face negative sample are can be obtained by so, without the need for can just obtain by way of artificial collection image
To multifarious training sample, that is, human resourcess are saved, improve the multiformity of training sample again, and can ensure that face
The accuracy of testing result.
Description of the drawings
Fig. 1 be the embodiment of the present invention in for Face datection training sample generation method flow chart；
Fig. 2 is to obtain the method flow of final face positive sample set in the embodiment of the present invention；
Fig. 3 be the embodiment of the present invention in for Face datection training sample generating means structural representation.
Specific embodiment
For human resourcess' serious waste caused by the mode of existing artificial collection training sample, training sample it is various
Property reduce and cannot ensure face testing result accuracy problem, it is provided in an embodiment of the present invention for Face datection
The generation method of training sample, the flow process of the method is as shown in figure 1, execution step is as follows：
S10：Obtain first and the face positive sample of quantity is set as original face positive sample set, and obtain second set
The face negative sample of fixed number amount is used as original face negative sample set.
Can be obtained by way of artificial collection, it is also possible to the existing image of direct access.Generic face positive sample
Quantity is greater than 3000 width, including the various ages, express one's feelings, certain angle of vacillating now to the left, now to the right（Less than 45 degree）, pitch certain angle up and down
Degree（Less than 25 degree）, it is not agnate, these face positive samples can serve as original face positive sample set；Face negative sample will
More than 10000 width, as far as possible including various indoor scenes（Computer, desk, TV background, kitchen, bedroom, office, window etc.）
With outdoor different scenes（Trees, meadow, building, road, sky, sea, HUAMU, farmland, food market etc.）, these face negative samples
Can serve as original face negative sample set.
S11：Image procossing is carried out to the face positive sample in original face positive sample set, and is just added to original face
In sample set, middle face positive sample set is obtained.
S12：Face positive sample in middle face positive sample set is randomly selected and random weighting, and is added to
In middle face positive sample set, final face positive sample set is obtained.
S13：Face negative sample in original face negative sample set is randomly selected and step-by-step logical operationss, and added
It is added in original face negative sample set, obtains final face negative sample set.
S11 can be first carried out, then performs S13；S13 can also be first carried out, then performs S11；Can certainly perform simultaneously
S11 and S13.
S14：Using the set of final face positive sample and the set of final face negative sample as the training sample for Face datection
This.
The program obtains the face positive sample of setting quantity and face negative sample, then to these face positive samples and face
Negative sample carries out certain process, so can be obtained by more face positive samples and face negative sample, need not be by artificial
The mode of collection image can be obtained by multifarious training sample, that is, save human resourcess, improve training sample again
Multiformity, and can ensure that the accuracy of Face datection result.
More preferably, the face positive sample of the setting quantity of acquisition first in above-mentioned S10 is used as original face positive sample set
Afterwards, also include：The similarity of face positive sample two-by-two is calculated in original face positive sample set according to the first selected characteristic；From
Similarity is deleted in original face positive sample set more than in two face positive samples of given threshold.
After the face negative sample of the setting quantity of acquisition second in above-mentioned S10 is as original face negative sample set, also
Including：The similarity of face negative sample two-by-two is calculated in original face negative sample set according to the second selected characteristic；From primitive man
Similarity is deleted in face negative sample set more than in two face negative samples of given threshold.
In order to improve detection efficiency, can get rid of first similar in the set of face positive sample and face negative sample set
Sample, after similarity is calculated to the sample two-by-two in the set of face positive sample and face negative sample set respectively, can be by phase
Like a deletion in sample of the degree more than given threshold.
Specifically, it is above-mentioned that the phase of face positive sample two-by-two is calculated in original face positive sample set according to the first selected characteristic
Like spending, and the similarity of face negative sample two-by-two is calculated in original face negative sample set according to the second selected characteristic, specifically
Including：Face positive sample two-by-two or similarity Score of face negative sample two-by-two are calculated by following equation：
Wherein, N is the number of the first selected characteristic or the second selected characteristic, f1i、f2iFor the first of face positive sample two-by-two
The value of the second selected characteristic i of selected characteristic i or two-by-two face negative sample.
Specifically, image procossing is carried out to the face positive sample in original face positive sample set in above-mentioned S11, specifically
Including：One below or combination are included to the image procossing that the face positive sample in original face positive sample set is carried out：At random
Noise superposition, illumination variation, block process.
Specifically, the above-mentioned face positive sample in original face positive sample set carries out random noise superposition, by under
Row formula carries out random noise superposition：Wherein, Inoise（I, j）
After representing random noise superposition（I, j）The brightness of point, I（I, j）Before representing random noise superposition（I, j）The brightness of point, k is random
Number.
Specifically, the above-mentioned face positive sample in original face positive sample set carries out illumination variation process, by under
Row formula carries out vertical progressive conversion：Wherein, I1（I, j）Represent vertical progressive
After conversion（I, j）The brightness of point, I（I, j）Before representing vertical progressive conversion（I, j）The brightness of point, k is random number, and H represents face
The height of positive sample.
Vertical strip conversion is carried out by following equation：
Wherein, I2（I, j）After representing vertical strip conversion（I, j）The brightness of point, I（I, j）Before representing vertical strip conversion（I, j）Point
Brightness.
The progressive conversion of level is carried out by following equation：Wherein, I3（I,
j）After the progressive conversion of expression level（I, j）The brightness of point, I（I, j）Before the progressive conversion of expression level（I, j）The brightness of point, W are represented
The width of face negative sample.
Horizontal strip conversion is carried out by following equation：
Wherein, I4（I, j）After representing horizontal strip conversion（I, j）The brightness of point, I（I, j）Before representing horizontal strip conversion（I, j）Point
Brightness.
Specifically, the face positive sample in original face positive sample set is carried out blocking process, is entered by following equation
Row left eye blocks process：Wherein, IL（I, j）Represent left
After eye-shade gear（I, j）The brightness of point, I（I, j）Before representing that left eye is blocked（I, j）The brightness of point, H represent the height of face positive sample
Degree, W represent the width of face positive sample, iLRepresent the abscissa of left eye, jLThe vertical coordinate of left eye is represented, k is random number.
Right eye is carried out by following equation and blocks process：
Wherein, IR（I, j）After representing that right eye is blocked（I, j）The brightness of point, I（I, j）Before representing that right eye is blocked（I, j）The brightness of point, iRTable
Show the abscissa of right eye, jLRepresent the vertical coordinate of right eye.
Eyes are carried out by following equation and blocks process：
Wherein, ILR（I, j）After representing that eyes are blocked（I, j）The brightness of point, I（I, j）Before representing that eyes are blocked（I, j）Point
Brightness.
Mouth is carried out by following equation and blocks process：
Wherein, MR（I, j）After representing that mouth is blocked（I, j）The brightness of point, I（I, j）Before representing that mouth is blocked（I, j）The brightness of point, iMTable
Show the abscissa of mouth, jMRepresent the vertical coordinate of mouth.
Specifically, as shown in Fig. 2 in above-mentioned S12 the face positive sample in middle face positive sample set is carried out with
Machine is extracted and random weighting, and is added in middle face positive sample set, obtains final face positive sample set, concrete to wrap
Include：
S120：Random number is obtained, by the number of face positive sample of the random number to including in middle face positive sample set
Take the remainder.
Random number can be generated using prior art, use k=random（）To represent, if middle face positive sample set
The number of the face positive sample included in town is Np, then can take kmodNp, as Ns.
S121：Face positive sample, the number of the face positive sample of extraction is randomly selected from middle face positive sample set
For remainder.
Ns face positive sample is randomly selected from middle face positive sample set.
S122：Random to generate weighted value, wherein, the number of the random weighted value for generating is remainder, and and is 1.
Ns weighted value, W is generated at random1、W2……WNs, wherein, W1+W2+……+WNs=1。
S123：Using face positive sample weighted sum of the random weighted value for generating to remainder, the positive sample of newly-increased face is obtained
This, and be added in middle face positive sample set, obtain final face positive sample set.
Specifically, in above-mentioned S123 face positive sample weighted sum of the weighted value that use is generated at random to remainder, leads to
Cross following equation and obtain newly-increased face positive sample Snew:Wherein, SiRepresent the bright of i-th face positive sample
Degree matrix, wiRepresent i-th weighted value, NsRepresent remainder.
Specifically, the face negative sample in original face negative sample set is randomly selected in above-mentioned S13 and by
Position logical operationss, and be added in original face negative sample set, final face negative sample set is obtained, is specifically included：At random
Extract two face negative samples in original face negative sample set；By randomly select two face negative samples carry out with, or
Three newly-increased face negative samples are obtained with XOR, and is added in original face negative sample set, obtained final face and bear
Sample set.
Wherein, randomly select two face samples are carried out into step-by-step logic and operation and can passes through formula Snew1=S1＆ S2
Realize, randomly select two face samples are carried out into step-by-step logic or computing can pass through formula Snew1=S1|S2Realize, will be with
Two face samples that machine is extracted carry out step-by-step logic XOR and can pass through formula Snew1=S1^S2Realize.
Based on same inventive concept, the embodiment of the present invention provides a kind of generation dress of training sample for Face datection
Put, the structure of the device as shown in figure 3, including：
Acquiring unit 30, sets the face positive sample of quantity as original face positive sample set for obtaining first, with
And the face negative sample of the second setting quantity is obtained as original face negative sample set.
Processing unit 31, for carrying out image procossing to the face positive sample in original face positive sample set, and adds
To in original face positive sample set, middle face positive sample set is obtained；To the face in middle face positive sample set just
Sample is randomly selected and random weighting, and is added in middle face positive sample set, obtains final face positive sample collection
Close；And the face negative sample in original face negative sample set is randomly selected and step-by-step logical operationss, and be added to
In original face negative sample set, final face negative sample set is obtained.
Signal generating unit 32, for using the set of final face positive sample and the set of final face negative sample as examining for face
The training sample of survey.
Preferably, above-mentioned acquiring unit 30, is additionally operable to obtaining the face positive sample of the first setting quantity as primitive man
After face positive sample set, the similar of face positive sample two-by-two is calculated in original face positive sample set according to the first selected characteristic
Degree；Similarity is deleted from original face positive sample set more than in two face positive samples of given threshold；Obtain
After the face negative sample of the second setting quantity is as original face negative sample set, also include：According to the second selected characteristic meter
Calculate in original face negative sample set the similarity of face negative sample two-by-two；Similarity is deleted from original face negative sample set
More than one in two face negative samples of given threshold.
Specifically, above-mentioned acquiring unit 30, specifically for：Face positive sample two-by-two or people two-by-two are calculated by following equation
Similarity Score of face negative sample：Wherein, N is the first selected characteristic or the second selected characteristic
Number, f1i、f2iThe first selected characteristic i for face positive sample two-by-two or two-by-two the second selected characteristic i of face negative sample
Value.
Specifically, above-mentioned processing unit 31, specifically includes：Face positive sample in original face positive sample set is carried out
Image procossing include one below or combination：Random noise superposition, illumination variation, block process.
Specifically, above-mentioned processing unit 31, specifically for：Random noise superposition is carried out by following equation：Wherein, Inoise（I, j）After representing random noise superposition（I, j）
The brightness of point, I（I, j）Before representing random noise superposition（I, j）The brightness of point, k is random number.
Specifically, above-mentioned processing unit 31, specifically for：
Vertical progressive conversion is carried out by following equation：Wherein, I1（I,
j）After representing vertical progressive conversion（I, j）The brightness of point, I（I, j）Before representing vertical progressive conversion（I, j）Point brightness, k be with
Machine number, H represent the height of face positive sample；
Vertical strip conversion is carried out by following equation：
Wherein, I2（I, j）After representing vertical strip conversion（I, j）The brightness of point, I（I, j）Before representing vertical strip conversion（I, j）Point
Brightness；
The progressive conversion of level is carried out by following equation：Wherein, I3（I,
j）After the progressive conversion of expression level（I, j）The brightness of point, I（I, j）Before the progressive conversion of expression level（I, j）The brightness of point, W are represented
The width of face negative sample；
Horizontal strip conversion is carried out by following equation：
Wherein, I4（I, j）After representing horizontal strip conversion（I, j）The brightness of point, I（I, j）Before representing horizontal strip conversion（I, j）Point
Brightness.
Specifically, above-mentioned processing unit 31, specifically for：
Left eye is carried out by following equation and blocks process：
Wherein, IL（I, j）After representing that left eye is blocked（I, j）The brightness of point, I（I, j）Before representing that left eye is blocked（I, j）The brightness of point, H tables
The height of face positive sample of leting others have a look at, W represent the width of face positive sample, iLRepresent the abscissa of left eye, jLRepresent the vertical seat of left eye
Mark, k is random number；
Right eye is carried out by following equation and blocks process：
Wherein, IR（I, j）After representing that right eye is blocked（I, j）The brightness of point, I（I, j）Before representing that right eye is blocked（I, j）The brightness of point, iRTable
Show the abscissa of right eye, jLRepresent the vertical coordinate of right eye；
Eyes are carried out by following equation and blocks process：Wherein, ILR
（I, j）After representing that eyes are blocked（I, j）The brightness of point, I（I, j）Before representing that eyes are blocked（I, j）The brightness of point；
Mouth is carried out by following equation and blocks process：
Wherein, MR（I, j）After representing that mouth is blocked（I, j）The brightness of point, I（I, j）Before representing that mouth is blocked（I, j）The brightness of point, iMTable
Show the abscissa of mouth, jMRepresent the vertical coordinate of mouth.
Specifically, above-mentioned processing unit 31, specifically for：Random number is obtained, by random number to middle face positive sample collection
The number of the face positive sample included in conjunction takes the remainder；Face positive sample is randomly selected from middle face positive sample set, is taken out
The number of the face positive sample for taking is remainder；Random to generate weighted value, wherein, the number of the random weighted value for generating is remainder,
And and be 1；Using face positive sample weighted sum of the random weighted value for generating to remainder, newly-increased face positive sample is obtained, and
It is added in middle face positive sample set, obtains final face positive sample set.
Specifically, above-mentioned processing unit 31, specifically for：Newly-increased face positive sample S is obtained by following equationnew：Wherein, SiRepresent the luminance matrix of i-th face positive sample, wiRepresent i-th weighted value, NsRepresent remaining
Number.
Specifically, above-mentioned processing unit 31, specifically for：Randomly select two faces in original face negative sample set
Negative sample；By randomly select two face negative samples carry out with, or obtain three newly-increased face negative samples with XOR, and
It is added in original face negative sample set, obtains final face negative sample set.
Obviously, those skilled in the art can carry out the essence of various changes and modification without deviating from the present invention to the present invention
God and scope.So, if these modifications of the present invention and modification belong to the scope of the claims in the present invention and its equivalent technologies
Within, then the present invention is also intended to comprising these changes and modification.
Claims (20)
1. a kind of generation method of the training sample for Face datection, it is characterised in that include：
Obtain first and the face positive sample of quantity is set as original face positive sample set, and obtain second set quantity
Face negative sample is used as original face negative sample set；
Image procossing is carried out to the face positive sample in the original face positive sample set, and is just added to the original face
In sample set, middle face positive sample set is obtained；Face positive sample in the middle face positive sample set is carried out
Randomly select, the face positive sample to extracting carries out random weighting；And based on the weighted value in random weighting result, to what is extracted
Face positive sample carries out the computing based on weighted value, and is added in the middle face positive sample set, obtains final face
Positive sample set；And
Face negative sample in the original face negative sample set is randomly selected, the face negative sample to extracting is carried out
Step-by-step logical operationss, and be added in the original face negative sample set, obtain final face negative sample set；
Using final face positive sample set and final face negative sample set as the training sample for Face datection
This.
2. the method for claim 1, it is characterised in that obtain first and set the face positive sample of quantity as primitive man
After face positive sample set, also include：
The similarity of face positive sample two-by-two is calculated in the original face positive sample set according to the first selected characteristic；
Similarity is deleted from the original face positive sample set more than in two face positive samples of given threshold；
After the face negative sample of acquisition the second setting quantity is as original face negative sample set, also include：
The similarity of face negative sample two-by-two is calculated in the original face negative sample set according to the second selected characteristic；
Similarity is deleted from the original face negative sample set more than in two face negative samples of the given threshold
One.
3. method as claimed in claim 2, it is characterised in that the original face positive sample is calculated according to the first selected characteristic
The similarity of face positive sample two-by-two in set, and calculated in the original face negative sample set according to the second selected characteristic
The similarity of face negative sample, specifically includes two-by-two：
Face positive sample two-by-two or similarity Score of face negative sample two-by-two are calculated by following equation：
Wherein, N is the number of the first selected characteristic or the second selected characteristic, f1i、f2iFor face positive sample two-by-two first is chosen
The value of the second selected characteristic i of feature i or two-by-two face negative sample.
4. the method for claim 1, it is characterised in that to the face positive sample in the original face positive sample set
Image procossing is carried out, is specifically included：
One below or combination are included to the image procossing that the face positive sample in the original face positive sample set is carried out：With
The superposition of machine noise, illumination variation, block process.
5. method as claimed in claim 4, it is characterised in that to the face positive sample in the original face positive sample set
Random noise superposition is carried out, is specifically included：
Random noise superposition is carried out by following equation：
Wherein, InoiseThe brightness that (i, j) puts after (i, j) expression random noise superposition, before I (i, j) represents random noise superposition (i,
J) brightness put, k are random number.
6. method as claimed in claim 4, it is characterised in that to the face positive sample in the original face positive sample set
Illumination variation process is carried out, is specifically included：
Vertical progressive conversion is carried out by following equation：Wherein, I1(i, j) is represented
The brightness that (i, j) puts after vertical progressive conversion, the brightness that (i, j) puts before the vertical progressive conversion of I (i, j) expressions, H represent face just
The height of sample, i and j are position coordinateses；
Vertical strip conversion is carried out by following equation：Its
In, I2(i, j) represent (i, j) brightness for putting after vertical strip conversion, I (i, j) represent vertical strip conversion before (i, j) put it is bright
Degree, i and j are position coordinateses；
The progressive conversion of level is carried out by following equation：Wherein, I3(i, j) table
The brightness that (i, j) puts after showing the progressive conversion of level, the brightness that (i, j) puts before the progressive conversion of I (i, j) expressions level, W represent face
The width of positive sample, i and j are position coordinateses；
Horizontal strip conversion is carried out by following equation：Its
In, I4(i, j) represent (i, j) brightness for putting after horizontal strip conversion, I (i, j) represent horizontal strip conversion before (i, j) put it is bright
Degree, i and j are position coordinateses.
7. method as claimed in claim 4, it is characterised in that to the face positive sample in the original face positive sample set
Carry out blocking process, specifically include：
Left eye is carried out by following equation and blocks process：
Wherein, IL(i, j) represent left eye block after (i, j) brightness for putting, I (i, j) represent left eye block before (i, j) brightness for putting, H tables
The height of face positive sample of leting others have a look at, W represent the width of face positive sample, iLRepresent the abscissa of left eye, jLRepresent the vertical seat of left eye
Mark, k are random number, and i and j is position coordinateses；
Right eye is carried out by following equation and blocks process：
Wherein, IR(i, j) represent right eye block after (i, j) brightness for putting, I (i, j) represent right eye block before (i, j) brightness for putting, iRTable
Show the abscissa of right eye, jLThe vertical coordinate of right eye is represented, i and j is position coordinateses；
Eyes are carried out by following equation and blocks process：Wherein, ILR
(i, j) represent eyes block after (i, j) brightness for putting, I (i, j) represent eyes block before (i, j) brightness for putting, i and j is position
Coordinate；
Mouth is carried out by following equation and blocks process：
Wherein, IM(i, j) represent mouth block after (i, j) brightness for putting, I (i, j) represent mouth block before (i, j) brightness for putting, iMTable
Show the abscissa of mouth, jMThe vertical coordinate of mouth is represented, i and j is position coordinateses.
8. the method for claim 1, it is characterised in that to the face positive sample in the middle face positive sample set
Randomly selected and random weighting, and be added in the middle face positive sample set, obtained final face positive sample collection
Close, specifically include：
Random number is obtained, the number of face positive sample of the random number to including in the middle face positive sample set is taken
Remainder；
Face positive sample is randomly selected from the middle face positive sample set, and the number of the face positive sample of extraction is described
Remainder；
Random to generate weighted value, wherein, the number of the random weighted value for generating is the remainder, and and is 1；
Using face positive sample weighted sum of the random weighted value for generating to the remainder, newly-increased face positive sample is obtained, and
It is added in the middle face positive sample set, obtains final face positive sample set.
9. method as claimed in claim 8, it is characterised in that using the random weighted value for generating to the face of the remainder just
Sample weighting is sued for peace, and specifically includes：
Newly-increased face positive sample S is obtained by following equationnew：
Wherein, SiRepresent the luminance matrix of i-th face positive sample, wiRepresent i-th weighted value, NsRepresent remainder.
10. the method for claim 1, it is characterised in that to the negative sample of face in the original face negative sample set
Originally randomly selected and step-by-step logical operationss, and be added in the original face negative sample set, obtained final face and bear
Sample set, specifically includes：
Randomly select two face negative samples in the original face negative sample set；
By randomly select two face negative samples carry out with, or obtain three newly-increased face negative samples with XOR, and add
It is added in the original face negative sample set, obtains final face negative sample set.
11. a kind of generating means of the training sample for Face datection, it is characterised in that include：
Acquiring unit, sets the face positive sample of quantity as original face positive sample set for obtaining first, and obtains
The face negative sample of the second setting quantity is used as original face negative sample set；
Processing unit, for carrying out image procossing to the face positive sample in the original face positive sample set, and is added to
In the original face positive sample set, middle face positive sample set is obtained；To in the middle face positive sample set
Face positive sample is randomly selected, and the face positive sample to extracting carries out random weighting；And based in random weighting result
Weighted value, the face positive sample to extracting carry out the computing based on weighted value, and are added to the middle face positive sample set
In, obtain final face positive sample set；And the face negative sample in the original face negative sample set is carried out at random
Extract, the face negative sample to extracting carries out step-by-step logical operationss, and is added in the original face negative sample set, obtains
Final face negative sample set；
Signal generating unit, for using final face positive sample set and final face negative sample set as face
The training sample of detection.
12. devices as claimed in claim 11, it is characterised in that acquiring unit, are additionally operable to obtaining the first setting quantity
After face positive sample is as original face positive sample set, the original face positive sample collection is calculated according to the first selected characteristic
The similarity of face positive sample two-by-two in conjunction；
Similarity is deleted from the original face positive sample set more than in two face positive samples of given threshold；
After the face negative sample of acquisition the second setting quantity is as original face negative sample set, also include：
The similarity of face negative sample two-by-two is calculated in the original face negative sample set according to the second selected characteristic；
Similarity is deleted from the original face negative sample set more than in two face negative samples of the given threshold
One.
13. devices as claimed in claim 12, it is characterised in that the acquiring unit, specifically for：
Face positive sample two-by-two or similarity Score of face negative sample two-by-two are calculated by following equation：
Wherein, N is the number of the first selected characteristic or the second selected characteristic, f1i、f2iFor face positive sample two-by-two first is chosen
The value of the second selected characteristic i of feature i or two-by-two face negative sample.
14. devices as claimed in claim 11, it is characterised in that the processing unit, specifically include：
One below or combination are included to the image procossing that the face positive sample in the original face positive sample set is carried out：With
The superposition of machine noise, illumination variation, block process.
15. devices as claimed in claim 14, it is characterised in that the processing unit, specifically for：
Random noise superposition is carried out by following equation：
Wherein, InoiseThe brightness that (i, j) puts after (i, j) expression random noise superposition, before I (i, j) represents random noise superposition (i,
J) brightness put, k are random number.
16. devices as claimed in claim 14, it is characterised in that the processing unit, specifically for：
Vertical progressive conversion is carried out by following equation：Wherein, I1(i, j) is represented
The brightness that (i, j) puts after vertical progressive conversion, the brightness that (i, j) puts before the vertical progressive conversion of I (i, j) expressions, H represent face just
The height of sample, i and j are position coordinateses；
Vertical strip conversion is carried out by following equation：Its
In, I2(i, j) represent (i, j) brightness for putting after vertical strip conversion, I (i, j) represent vertical strip conversion before (i, j) put it is bright
Degree, i and j are position coordinateses；
The progressive conversion of level is carried out by following equation：Wherein, I3(i, j) table
The brightness that (i, j) puts after showing the progressive conversion of level, the brightness that (i, j) puts before the progressive conversion of I (i, j) expressions level, W represent face
The width of positive sample, i and j are position coordinateses；
Horizontal strip conversion is carried out by following equation：Its
In, I4(i, j) represent (i, j) brightness for putting after horizontal strip conversion, I (i, j) represent horizontal strip conversion before (i, j) put it is bright
Degree, i and j are position coordinateses.
17. devices as claimed in claim 14, it is characterised in that the processing unit, specifically for：
Left eye is carried out by following equation and blocks process：
Wherein, IL(i, j) represent left eye block after (i, j) brightness for putting, I (i, j) represent left eye block before (i, j) brightness for putting, H tables
The height of face positive sample of leting others have a look at, W represent the width of face positive sample, iLRepresent the abscissa of left eye, jLRepresent the vertical seat of left eye
Mark, k are random number, and i and j is position coordinateses；
Right eye is carried out by following equation and blocks process：
Wherein, IR(i, j) represent right eye block after (i, j) brightness for putting, I (i, j) represent right eye block before (i, j) brightness for putting, iRTable
Show the abscissa of right eye, jLThe vertical coordinate of right eye is represented, i and j is position coordinateses；
Eyes are carried out by following equation and blocks process：Wherein, ILR
(i, j) represent eyes block after (i, j) brightness for putting, I (i, j) represent eyes block before (i, j) brightness for putting, i and j is position
Coordinate；
Mouth is carried out by following equation and blocks process：
Wherein, IM(i, j) represent mouth block after (i, j) brightness for putting, I (i, j) represent mouth block before (i, j) brightness for putting, iMTable
Show the abscissa of mouth, jMThe vertical coordinate of mouth is represented, i and j is position coordinateses.
18. devices as claimed in claim 11, it is characterised in that the processing unit, specifically for：
Random number is obtained, the number of face positive sample of the random number to including in the middle face positive sample set is taken
Remainder；
Face positive sample is randomly selected from the middle face positive sample set, and the number of the face positive sample of extraction is described
Remainder；
Random to generate weighted value, wherein, the number of the random weighted value for generating is the remainder, and and is 1；
Using face positive sample weighted sum of the random weighted value for generating to the remainder, newly-increased face positive sample is obtained, and
It is added in the middle face positive sample set, obtains final face positive sample set.
19. devices as claimed in claim 18, it is characterised in that the processing unit, specifically for：
Newly-increased face positive sample S is obtained by following equationnew：
Wherein, SiRepresent the luminance matrix of i-th face positive sample, wiRepresent i-th weighted value, NsRepresent remainder.
20. devices as claimed in claim 11, it is characterised in that the processing unit, specifically for：
Randomly select two face negative samples in the original face negative sample set；
By randomly select two face negative samples carry out with, or obtain three newly-increased face negative samples with XOR, and add
It is added in the original face negative sample set, obtains final face negative sample set.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310137893.XA CN104112131B (en) | 2013-04-19 | 2013-04-19 | Method and device for generating training samples used for face detection |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310137893.XA CN104112131B (en) | 2013-04-19 | 2013-04-19 | Method and device for generating training samples used for face detection |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104112131A CN104112131A (en) | 2014-10-22 |
| CN104112131B true CN104112131B (en) | 2017-03-22 |
Family
ID=51708915
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201310137893.XA Active CN104112131B (en) | 2013-04-19 | 2013-04-19 | Method and device for generating training samples used for face detection |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104112131B (en) |
Families Citing this family (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP3158498A4 (en) * | 2014-11-15 | 2017-08-16 | Beijing Kuangshi Technology Co. Ltd. | Face detection using machine learning |
| CN107729877B (en) | 2017-11-14 | 2020-09-29 | 浙江大华技术股份有限公司 | Face detection method and device based on cascade classifier |
| CN109344877B (en) * | 2018-08-31 | 2020-12-11 | 深圳先进技术研究院 | Sample data processing method, sample data processing device and electronic equipment |
| CN112819173A (en) * | 2019-11-18 | 2021-05-18 | 上海光启智城网络科技有限公司 | Sample generation method and device and computer readable storage medium |
| CN114742171A (en) * | 2022-04-24 | 2022-07-12 | 中山大学 | Compression method, device and storage medium for intrinsic orthogonal decomposition sample |
| CN115050086B (en) * | 2022-08-15 | 2022-11-04 | 北京百度网讯科技有限公司 | Sample image generation method, model training method, image processing method and device |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1317677C (en) * | 2004-03-19 | 2007-05-23 | 中国科学院计算技术研究所 | Genetic algorithm based human face sample generating method |
| CN101408943A (en) * | 2007-10-09 | 2009-04-15 | 三星电子株式会社 | Method for generating a training set for human face detection |
| CN102103695A (en) * | 2009-12-21 | 2011-06-22 | 北京中星微电子有限公司 | Method and device for generating image sample |
| WO2012131149A1 (en) * | 2011-03-25 | 2012-10-04 | Nokia Corporation | Method apparatus and computer program product for detection of facial expressions |
-
2013
- 2013-04-19 CN CN201310137893.XA patent/CN104112131B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1317677C (en) * | 2004-03-19 | 2007-05-23 | 中国科学院计算技术研究所 | Genetic algorithm based human face sample generating method |
| CN101408943A (en) * | 2007-10-09 | 2009-04-15 | 三星电子株式会社 | Method for generating a training set for human face detection |
| CN102103695A (en) * | 2009-12-21 | 2011-06-22 | 北京中星微电子有限公司 | Method and device for generating image sample |
| WO2012131149A1 (en) * | 2011-03-25 | 2012-10-04 | Nokia Corporation | Method apparatus and computer program product for detection of facial expressions |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104112131A (en) | 2014-10-22 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104112131B (en) | Method and device for generating training samples used for face detection | |
| CN108875633B (en) | Expression detection and expression driving method, device and system and storage medium | |
| CN103810504B (en) | Image processing method and device | |
| CN107766842B (en) | Gesture recognition method and application thereof | |
| CN108805068A (en) | A kind of motion assistant system, method, apparatus and medium based on student movement | |
| CN106874826A (en) | Face key point-tracking method and device | |
| CN103198330B (en) | Real-time human face attitude estimation method based on deep video stream | |
| WO2007120558A3 (en) | Image classification based on a mixture of elliptical color models | |
| Jiang et al. | Online robust action recognition based on a hierarchical model | |
| CN104268514A (en) | Gesture detection method based on multi-feature fusion | |
| CN107301200A (en) | A kind of article appraisal procedure and system analyzed based on Sentiment orientation | |
| CN110163567A (en) | Classroom roll calling system based on multitask concatenated convolutional neural network | |
| WO2022178833A1 (en) | Target detection network training method, target detection method, and apparatus | |
| CN107491543A (en) | A kind of client-based calligraphy auxiliary exercise method and system | |
| CN107066979A (en) | A kind of human motion recognition method based on depth information and various dimensions convolutional neural networks | |
| CN109388725A (en) | The method and device scanned for by video content | |
| JP2010128633A (en) | Information presentation device, method and program | |
| Wang et al. | Distortion recognition for image quality assessment with convolutional neural network | |
| Liu et al. | Trampoline motion decomposition method based on deep learning image recognition | |
| Hsieh et al. | A distance computer vision assisted yoga learning system | |
| CN110298345A (en) | A kind of area-of-interest automatic marking method of medical images data sets | |
| Boychuk et al. | An exploratory sentiment and facial expressions analysis of data from photo-sharing on social media: The case of football violence | |
| CN106485226A (en) | A kind of video pedestrian detection method based on neutral net | |
| CN112801092B (en) | Method for detecting character elements in natural scene image | |
| CN103093467A (en) | Shot boundary detection method based on double detection model |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant |