Suspect's recognition of face identificationTechnical field
The present invention relates to a kind of microcomputer data processings, the comparison especially by computer to video, pictureIt calculates, realizes the system and method for suspect's crime fact identification.
Background technique
With the maturation of existing video monitoring network system construction and going deep into for informatization, public safety departmentIt is increasing to emergency event video intelligentization tracking demand.Such as: when needing through city video monitoring net to suspicion of crimeWhen people carries out tracking determining, at present or the multitude of video data as captured by each monitor camera of artificial screening is needed, neededIt goes to screen comparison frame by frame from the video data of certain period of each shot by camera.Traditional data processing methodIt is to be concentrated in together whole data centered on picture auditor, is screened frame by frame by picture auditor, picture at this moment is examinedCentral processing unit of the person of looking into as computer；Platform is either stored by picture auditor to each associated camera informationUpper to be screened frame by frame, such the time it takes is longer, and personnel are easy fatigue, and details is not easy to find, so success rate is notIt is high.After the research and development by many years, development completes a variety of " face alignment identification engines " for engineering department, comes at presentSubstantive service stage, " face alignment identification engine " can be learnt by the data to suspect's picture, be stored from engineFound out in a large amount of passerby's pictures with the higher picture of suspect's similarity, be lined up according to similarity height, listing canThe suspect of energy.But found from practice, " the face alignment identification engine " of each company's production respectively has excellent length at present, also respectivelyThere is deficiency, so accuracy rate is not high so far.
Summary of the invention
The purpose of the present invention is the results by analyzing various " face alignment identification engines " directly to be utilized.It is logicalAssociation composite algorism is crossed, raising judges speed and accuracy of judgement degree.
Data storage unit:
A) panorama sketch valut: panoramic pictures carry out being stored in panorama sketch valut after label determines；Label has: shooting seat in the plane is clappedThe time is taken the photograph, is automatically credited in picture storage；
B) local picture library: mainly face picture has carried out association structure data and has been stored in local picture library together；KnotStructure data include: personnel's household register, it is illegal record etc., system has been called identity personnel；Structural data is for the first time by peopleWork mark, it is subsequent picture enter engine matching in automatically generate；
C) pay close attention to personnel's small video library: concern personnel include identity personnel and interim supplement mark personnel；According to faceThe when and where of picture transfers one section of small video clips in panorama sketch valut, is stored in concern personnel small video library.
Multi engine scheduler module:
Include Data reduction model；
Image data I push is distributed to each " face alignment identification engine " input interface, when " face alignment identification is drawnHold up " when receiving image data I, image data I passback is carried out, Data reduction model will scheme in multi engine scheduler moduleSheet data I and artwork data I carries out subtraction and compares, and passes through when total is zero；Total is not zero, and retransmits.
" face alignment identification engine " screening:
Screening technique: having had more identification engines to have practicability at present, by calculating sifting, determines and knows to severalOther engine is preferentially associated with, and can be improved to face alignment recognition correct rate.
Implementation method is as follows:
Enabling " the face alignment identification engine " quantity for participating in screening is n, 20 >=n >=2；
Same picture is inputted, n platform " face alignment identification engine " exports N number of most like record, recommendation N=20；
It is the similarity value range for comparing picture is unified are as follows: 0~100；Setting system thresholds A is 80-70.
To the evaluation method of " face alignment identification engine ",
Method 1: ballot mode is taken:
It chooses n family " face alignment identification engine " and participates in screening, screening set alarm threshold value T=n/2 uses samePicture, by picture portrait with identity personnel are compared in database when, and if only if in the output of at least T engineIt is hit in the most like record of top n, which is added alarm logging.
It calculates step in two steps:
Step 1: the mean value and variance of " face alignment identification engine " output result are calculated；
Step 2: it calculates with the normal distribution cumulative distribution function of mean value and variance in the value of sm, the as value of wm:The comprehensive output result finally presented sorts from high to low according to above-mentioned comprehensive similarity.
" the face alignment identification engine " system of choosing, with the descending selected and sorted of wm value.
2: one ticket trigger-type of method:
As long as there is more than one " face alignment identification engine " alarm, final alarm list is added in alarm entryIn.The calculating step of the comprehensive score of final alarm list and sortord with method 1.
After the ranking achievement that one " face alignment identification engine " obtains in method 1 and method 2 is added, still useNumerical values recited determines, descending sequencing selection.
Secondary comprehension analysis and processing module:
Comprising hitting statistical model, weighted model, comprehensive similarity computation model.
On the basis of the recognition result of " face alignment identification engine " output that different vendor provides, the purpose is to apply" face alignment identification engine " output of multiple commercial vendors uses rule as a result, designing, and by mutually confirming, carries out secondary comprehensiveAnalysis is closed, final identification comparison result is exported；The accuracy and stability for improving recognition of face are realized, to be subsequentThe various profound applications such as historical record analysis provide accurately result data more than single engine.
Setting principle:
Principle 1: it is required that " face alignment identification engine " provides practical identification Top-1~Top- to certain face picture20 hit results screenshot and hit rate data.
Top indicates hit probability sequence.
Principle 2: the unified human face similarity degree value for assigning each " face alignment identification engine " hit results is 0~100；Setting system thresholds A is human face similarity degree 80.
Design judgment rule；Unified similarity score section:
95 points or more: extremely similar, it is believed that substantially completely determine, it is accurate to hit；
90~95 points: much like, greater probability is possible, correct to identify；
80~90 points: it is somewhat like, there is certain correctness, it may identification；
70~80 points: being somewhat like, accuracy is low, and identification possibility is lower；
< 70 divides: substantially mismatching, does not charge to.
Principle 3: unified final result only takes Top-1~Top-10.
Implementation method:
By taking n family " face alignment identification engine " application as an example:
Numbered to each setting: n=1,2,3 ..., n；
Step 1: inputting the face picture of same people to be identified to " face alignment identification engine " n, each " face alignmentIdentify engine " the independent engine identification of n progress；
Step 2: each " face alignment identification engine " n is calculated by similarity calculation, hit probability order models；
To n platform " face alignment identification engine " unified setting threshold value A, threshold value A is arranged between 70-80, by operator's rootIt is set according to instruction；Each " face alignment identification engine " obtains preceding 20 Top sequence conduct of screenshot and hit rate dataTop-1~Top-20 list；If taking determining numerical value when numerical value satisfaction >=A of Top；If when the numerical value < A of Top,To the Top record of discontented threshold value, it is assigned a value of 0；Top sort when comprising be assigned a value of 0 record, Top still maintain Top-1~Top-20 list.
The identification card number of one personnel in n platform " face alignment identification engine " by any one, " draw by face alignment identificationHold up " gained Top-1~Top-20 list in hit, then be denoted as Listm；
Step 3: statistical model all lists the picture of the 20*n in whole n platforms " face alignment identification engine ",List1, List2, List3 ... ..., Listm；Then Listm picture is screened, filters out the identity of all personnelDemonstrate,prove statistical result.According to the practical carry out statistic of classification hit in Top-20 by how many " face alignment identification engines ": such asWhen fruit n >=7, an identity card is hit in Top-20 by 2 " face alignment identification engines ", then is not counted；If a bodyPart card is hit in Top-20 by 3 " face alignment identification engines ", is not also counted；If ... an identity card is by catwalk" face alignment identification engine " hit, start recording, i.e., the record ability typing at least hit in Top-20 by n/2 family's engineRecord；If the record of the number that an identity card is hit >=T times is concentrated, i.e., at least ordered in Top-20 by T family's engineIn record；..., if an identity card is hit m times, m≤n；As m=n, that is, indicate by n family's engine in Top-20Interior all hits；The record that we filter out ListT-Listm list again is denoted as primary election the results list F0；
Step 4: weighted model records every of primary election the results list F0, according to its corresponding each single engine similarityValue is weighted and averaged.
Assuming that the record appears in List1 and List2, corresponding list engine similarity is respectively s1 and s2, secondary pointAnalysis will determine corresponding weight w1 and w2 according to the similarity distribution situation of List1 and List2；
Step 5:
Calculate the mean value muT and variance sigmaT of the distribution of ListT similarity；
Calculate the mean value muT+1 and variance sigmaT+1 of the distribution of ListT+1 similarity；
Calculate the mean value muT+2 and variance sigmaT+2 of the distribution of ListT+2 similarity；
……；
Calculate the mean value mum and variance sigmam of the distribution of Listm similarity；
Step 6:
Using mean value muT, variance sigmaT Gaussian Profile sT cumulative probability as weight wT；
Using mean value muT+1, variance sigmaT+1 Gaussian Profile sT+1 cumulative probability as weight wT+1；
Using mean value muT+2, variance sigmaT+2 Gaussian Profile sT+2 cumulative probability as weight wT+2；
……；
Using mean value mum, variance sigmam Gaussian Profile sm cumulative probability as weight wm；
Note: single engine similarity s1, s2, s3 ..., sm value it is higher, w1, w2, w3 ..., wm weight also accordingly gets overIt is high；
Step 7: comprehensive similarity computation model is to comprehensive similarity value are as follows:
(w1*s1+w2*s2+w3*s3+……+wm*sm)/(w1+w2+w3+……+wm)；
Step 8: it exports final secondary analysis result: primary dcreening operation the results list F0 being ranked up according to similarity, equallyRetain the sequence that comprehensive similarity is more than certain threshold value, threshold value 70-80 is adjustable, program default 80；The sequence since peakThe Top list of threshold value or more is arranged out, top 10 is as comprehensive Top-10 output as a result, inadequate 10 then all outputs.
It is proved by testing and practicing, suspect's recognition of face after secondary analysis assert that system passes through using existingThe analysis of " face alignment identification engine " is as a result, by " face alignment identification engine " individually engine identification；Export Top list；T times and T+1, T+2 record again list primary election result F0 are hit in screening；Every record of primary election result F0 list is addedWeight average；Obtain comprehensive similarity value (wT*sT+w (T+1) * s (T+1)+w (T+2) * s (T+2)+...+wm*sm)/(wT+w(T+1)+w(T+2)+……+wm)；Export the comprehensive Top-10 ranking of final secondary analysis result.It facts have proved larger improveAccuracy at target.Rule is used as a result, designing by what " the face alignment identification engine " of application multiple commercial vendors exported, passes through phaseMutually confirmation carries out secondary comprehension analysis, exports final identification comparison result；Realize the accuracy for improving recognition of face and steadyIt is qualitative, to provide accurately number of results more than single engine for the various profound applications such as subsequent historical record analysisAccording to.
Detailed description of the invention
One of Fig. 1 face comparison identification example；
System schematic is assert in Fig. 2 suspect's recognition of face；
Fig. 3 suspect's recognition of face, which is assert, compares identification example schematic diagram after system carries out face comparison identification.
Specific embodiment
One, the alarm of identity staff list is searched,
Method 1: ballot mode is taken:
It chooses n family " face alignment identification engine " and participates in and screen, screening set alarm threshold value T=n/2, when with object diagramWhen piece is compared with personnel in identity staff list, and if only if the most like record of top n exported at least T engineAlarm logging is added in the object by middle hit.
Same alarm logging appears in t engine, and t >=T calculates step in two steps:
Step 1: the mean value and variance of " face alignment identification engine " output result are calculated；
Step 2: it calculates with the normal distribution cumulative distribution function of mean value and variance in the value of s, the as value of w: mostThe comprehensive output result presented eventually sorts from high to low according to above-mentioned comprehensive similarity.
2: one ticket trigger-type of method:
As long as there is more than one " face alignment identification engine " alarm, final alarm list is added in alarm entryIn.The comprehensive score and sortord of final alarm list are the same as method 1.
Two, ordinary person's list is alarmed,
System provides following two integration scenario and carries out the identification of ordinary person's list；
Method 1: ballot mode is taken:
It chooses n family " face alignment identification engine " and participates in and screen, screening set alarm threshold value T=n/2, when with object diagramWhen piece is compared with personnel in identity staff list, and if only if in the most like record of top n that at least T engine exportsAlarm logging is added in the object by hit.
Same alarm logging appears in t engine, and t >=T calculates step in two steps:
Step 1: the mean value and variance of " face alignment identification engine " output result are calculated；
Step 2: it calculates with the normal distribution cumulative distribution function of mean value and variance in the value of s, the as value of w: mostThe comprehensive output result presented eventually sorts from high to low according to above-mentioned comprehensive similarity.
Ordinary person is required just to think to compare to pass through when multi engine synthesis recognized list non-empty.
Method 2: veto by one vote formula:
As long as there is more than one " face alignment identification engine " alarm, final alarm list is added in alarm entryIn.The comprehensive score and sortord of final alarm list are the same as method 1.
Overall merit is carried out by the score obtained to two methods, thus selected " face alignment identification engine ".
Embodiment 1: by taking n family's " face alignment identification engine " applies simultaneously as an example.
Step 1: inputting the face picture of same people to be identified to " face alignment identification engine " n, each " face alignmentIdentify engine " the independent engine identification of n progress；
Step 2: each " face alignment identification engine " n is calculated by similarity calculation, hit probability order models,If A is 70-80 threshold value；Preceding 20 sequences are obtained as Top-1~Top-20 list；If satisfaction >=A similarity threshold numberWhen measuring < Top-20, then the Top of discontented threshold value is recorded, be assigned a value of 0.
Top-1~Top-20 list obtained by " face alignment identification engine " n, is denoted as Listm；
Step 3: statistical model will filter out List1, List2, List3 ... ..., and >=2 records are hit in ListmIt concentrates, i.e., the record at least hit in Top-20 by two engines, and by the record filtered out again list, is denoted as primary electionThe results list F0；
Step 4: weighted model records every of primary election the results list F0, according to its corresponding each single engine similarityValue is weighted and averaged.
Assuming that the record appears in List1 and List2, corresponding list engine similarity is respectively s1 and s2, secondary pointAnalysis will determine corresponding weight w1 and w2 according to the similarity distribution situation of List1 and List2；
Step 5:
Calculate the mean value mu1 and variance sigma1 of the distribution of List1 similarity；
Calculate the mean value mu2 and variance sigma2 of the distribution of List2 similarity；
Calculate the mean value mu3 and variance sigma3 of the distribution of List3 similarity；
……；
Calculate the mean value mum and variance sigmam of the distribution of Listm similarity；
Step 6:
Using mean value mu1, variance sigma1 Gaussian Profile s1 cumulative probability as weight w1；
Using mean value mu2, variance sigma2 Gaussian Profile s2 cumulative probability as weight w2；
Using mean value mu3, variance sigma3 Gaussian Profile s3 cumulative probability as weight w3；
……；
Using mean value mum, variance sigmam Gaussian Profile sm cumulative probability as weight wm；
Note: single engine similarity s1, s2, s3 ..., sm value it is higher, w1, w2, w3 ..., wm weight also accordingly gets overIt is high；
Step 7: comprehensive similarity computation model is to comprehensive similarity value are as follows:
(w1*s1+w2*s2+w3*s3+……+wm*sm)/(w1+w2+w3+……+wm)；
Step 8: it exports final secondary analysis result: primary dcreening operation the results list F0 being ranked up according to similarity, equallyRetain the sequence that comprehensive similarity is more than certain threshold value, threshold value 70-80 is adjustable, program default 80；The sequence since peakThe Top list of threshold value or more is arranged out, top 10 is as comprehensive Top-10 output as a result, inadequate 10 then all outputs.
Embodiment 2: by taking n family's " face alignment identification engine " applies simultaneously as an example.
Step 1: inputting the face picture of same people to be identified to " face alignment identification engine " n, each " face alignmentIdentify engine " the independent engine identification of n progress；
Step 2: each " face alignment identification engine " n is calculated by similarity calculation, hit probability order models,If A is 70-80 threshold value；Preceding 20 sequences are obtained as Top-1~Top-20 list；If satisfaction >=A similarity threshold peopleWhen member quantity < Top-20, then 0 is assigned a value of to the Top record of discontented threshold value.
Top-1~Top-20 list obtained by " face alignment identification engine " n, is denoted as Listm；
Step 3: statistical model will filter out List1, List2, List3 ... ..., and >=2 records are hit in ListmIt concentrates, i.e., the record at least hit in Top-20 by two engines, and by the record filtered out again list, is denoted as primary electionThe results list F0；
Step 4: weighted model records every of primary election the results list F0, according to its corresponding each single engine similarityValue is weighted and averaged.
Assuming that the record appears in List1 and List2, corresponding list engine similarity is respectively s1 and s2, secondary pointAnalysis will determine corresponding weight w1 and w2 according to the similarity distribution situation of List1 and List2；
Step 5:
Calculate the mean value mu1 and variance sigma1 of the distribution of List1 similarity；
Calculate the mean value mu2 and variance sigma2 of the distribution of List2 similarity；
Calculate the mean value mu3 and variance sigma3 of the distribution of List3 similarity；
……；
Calculate the mean value mum and variance sigmam of the distribution of Listm similarity；
Step 6:
Using mean value mu1, variance sigma1 Gaussian Profile s1 cumulative probability as weight w1；
Using mean value mu2, variance sigma2 Gaussian Profile s2 cumulative probability as weight w2；
Using mean value mu3, variance sigma3 Gaussian Profile s3 cumulative probability as weight w3；
……；
Using mean value mum, variance sigmam Gaussian Profile sm cumulative probability as weight wm；
Note: single engine similarity s1, s2, s3 ..., sm value it is higher, w1, w2, w3 ..., wm weight also accordingly gets overIt is high；
Step 7: comprehensive similarity computation model is to comprehensive similarity value are as follows:
(w1*s1+w2*s2+w3*s3+……+wm*sm)/(w1+w2+w3+……+wm)；
Step 8: it exports final secondary analysis result: primary dcreening operation the results list F0 being ranked up according to similarity, equallyRetain the sequence that comprehensive similarity is more than certain threshold value, the threshold value is adjustable, program default 80；Sequence arranges since peakTop list more than threshold value out, top 10 is as comprehensive Top-10 output as a result, inadequate 10 then all outputs.
Embodiment 3: by taking n family's " face alignment identification engine " applies simultaneously as an example.
Step 1: inputting the face picture of same people to be identified to " face alignment identification engine " n, each " face alignmentIdentify engine " the independent engine identification of n progress；
Step 2: each " face alignment identification engine " n is calculated by similarity calculation, hit probability order models,If A is 70-80 threshold value；Preceding 20 sequences are obtained as Top-1~Top-20 list；If satisfaction >=A similarity threshold numberWhen measuring < Top-20, then the Top of discontented threshold value is recorded, be assigned a value of 0.
Top-1~Top-20 list obtained by " face alignment identification engine " n, is denoted as Listm；
Step 3: statistical model will filter out List1, List2, List3 ... ..., and >=2 records are hit in ListmIt concentrates, i.e., the record at least hit in Top-20 by two engines, and by the record filtered out again list, is denoted as primary electionThe results list F0；
Step 4: weighted model records every of primary election the results list F0, according to its corresponding each single engine similarityValue is weighted and averaged.
Assuming that the record appears in List1 and List2, corresponding list engine similarity is respectively s1 and s2, secondary pointAnalysis will determine corresponding weight w1 and w2 according to the similarity distribution situation of List1 and List2；
Step 5:
Calculate the mean value mu1 and variance sigma1 of the distribution of List1 similarity；
Calculate the mean value mu2 and variance sigma2 of the distribution of List2 similarity；
Calculate the mean value mu3 and variance sigma3 of the distribution of List3 similarity；
……；
Calculate the mean value mum and variance sigmam of the distribution of Listm similarity；
Step 6:
Using mean value mu1, variance sigma1 Gaussian Profile s1 cumulative probability as weight w1；
Using mean value mu2, variance sigma2 Gaussian Profile s2 cumulative probability as weight w2；
Using mean value mu3, variance sigma3 Gaussian Profile s3 cumulative probability as weight w3；
……；
Using mean value mum, variance sigmam Gaussian Profile sm cumulative probability as weight wm；
Note: single engine similarity s1, s2, s3 ..., sm value it is higher, w1, w2, w3 ..., wm weight also accordingly gets overIt is high；
Step 7: comprehensive similarity computation model is to comprehensive similarity value are as follows:
(w1*s1+w2*s2+w3*s3+……+wm*sm)/(w1+w2+w3+……+wm)；
Step 8: it exports final secondary analysis result: primary dcreening operation the results list F0 being ranked up according to similarity, equallyRetain the sequence that comprehensive similarity is more than certain threshold value, the threshold value is adjustable, program default 80；Sequence arranges since peakTop list more than threshold value out, top 10 is as comprehensive Top-10 output as a result, inadequate 10 then all outputs.