CN110503031A - A method of improving face recognition accuracy rate and passage speed - Google Patents

Abstract

The invention discloses a kind of raising face recognition accuracy rate and the methods of passage speed, include the following steps: that the registration of user and bottom library all personnel are shone, in the alignment score maximum value that identification model A is obtained, i.e. top1 score meets: Score_top1 >=Th, Th is default recognition threshold, indicates that user is identified by；If at this time, Score_top1≤Th1, Th1 > Th, Th1 is by user setting, indicate that the user is identified by while the bottom library photo of the user differs greatly with current true man's similarity, the scene for increasing or updating the period, which is shone to shine as replacement, to be appended in user's face database.

Description

A method of improving face recognition accuracy rate and passage speed

Technical field

The present invention relates to field of face identification, specifically a kind of method for improving face recognition accuracy rate and passage speed.

Background technique

Recognition of face will receive the influence of different factors in the case where actually using scene, including light, registration shine, camera peace Dress height etc..These uncontrolled factors will affect recognition of face passage speed and accuracy rate.When actual use, these Disturbing factor can generally reduce face recognition accuracy rate, and tested personnel is caused to identify difficult problem.

It is full marks every time that recognition of face, which verifies score not being, this is by practical service environment and to register the difference according to style Property (such as registration is according to by p figure, U.S. face, beautification), the registration of tested personnel shine and there are age gap, light are different when passing through equipment Sample (light and shade is different, or even has other colour light sources to influence face tone), micro- expression changes, facial angle has the factors such as deviation Caused by.When in use, by these disturbing factors, a people may need repeatedly to identify to be passed through reality, significantly Reduce recognition accuracy and passage speed.And in order to quickly obtain recognition result, neural network instruction on embedded device Identification model size, the characteristic dimension practised are limited.

Summary of the invention

In order to solve the above technical problems existing in the prior art, the present invention provides a kind of raising recognition of face is accurate The method of rate and passage speed, includes the following steps:

The registration of user and bottom library all personnel are shone, and in the alignment score maximum value that identification model A is obtained, i.e. top1 score is full Foot: Score_top1 >=Th, Th are default recognition threshold, indicate that user is identified by；

If at this point, Score_top1≤Th1, Th1 > Th, Th1 by user setting, indicate that the user is identified by while the user Bottom library photo differ greatly with current true man's similarity, increase or update the scene of the period according to as replacement according to being appended to this In user's face database.

Further, the scene, which is shone, can be used as replacement according to it is necessary to meet following condition:

1, scene, which is shone, meets face quality indicator, including face size, angle, fuzziness, illumination；

2, secondary verification is done with identification model B higher than identification model A accuracy rate, larger, by all registrations of the user It shines according to the replacement with the non-period and is compared respectively with what the scene was shone, maximum alignment score >=Th2, and minimum comparison point Number >=Th3, Th2 and Th3 are by user's self-setting, for constraining replacement according to the confidence level whether to come into force.

Further, decide whether to enable bigger identification model B by monitoring cpu usage amount, i.e., when heavy traffic, Only candidate is replaced according to storage, when the business free time, identification model B is enabled and carrys out secondary verification.

Further, if the constraint condition of replacement is as follows:

1, period replacement is shone if it does not exist, which is registered in user's face database according to as replacement according to addition；

If 2, being shone before in the presence of the replacement of the period, compare the period is new, old replacement according to and all registrations photograph and the non-period For the similarity score that registration is shone by the score after mapping function, acquirement divides the higher person to shine as period replacement.

Further, the mapping function is as follows:

F(Scores)= (1-(1-s1)*(1-s2)*...*(1-sn))

The score that wherein Scores is s1, s2 ..., sn combines, and n indicates the quantity that registration is shone and the registration of the non-period is shone, and s1 is arrived Sn respectively indicates new, old replacement and takes the similarity for being compared picture with n.

The present invention can reduce the interference of external interference factor, mention in the case where guaranteeing even to improve recognition accuracy High traffic rate and reduce the recognition of face verification time.

Detailed description of the invention

Fig. 1 is flow chart of the method for the present invention.

Specific embodiment

The present invention will be further explained below with reference to the attached drawings.

As shown in Figure 1, user, when using face recognition technology verifying identity, ideally, user and bottom library are all The registration of personnel is shone, and in the alignment score maximum value that identification model A is obtained, i.e. top1 score meets: Score_top1 >=Th, Th To preset recognition threshold, indicate that user is identified by, while showing that the identity of the user is the people that the corresponding registration of top1 score is shone Member's identity.If simultaneously should at this point, Score_top1≤Th1, Th1 > Th, Th1 by user setting, indicate that the user is identified by The bottom library photo of user with current true man's similarity be not it is especially high, increase at this time or update the scene of the period according to as replacement It according to being appended in user's face database, in this way when comparing next time, replaces and shines there are the user in the library of bottom, increase user knowledge Not successful probability.

Scene, which is shone, can be used as replacement according to it is necessary to meet following condition:

1, scene is shone and needs to meet the conditions such as face quality indicator, including face size, angle, fuzziness, illumination.

2, secondary verification is done with higher accuracy, bigger identification model B, all registrations of the user is shone and the non-period Replacement according to being compared respectively with what the identification scene was shone, the replacement of the non-period is according to referring to on-site identification using this method Photo addition is registered in user's face database, and the replacement that each user to be measured can generate different periods is shone, can according to it is early, In, evening is distinguish.If maximum alignment score >=Th2, and minimum comparison score >=Th3, then indicating identification scene It is required according to replacement is met, Th2 and Th3 are by user's self-setting, for constraining replacement according to the confidence level whether to come into force.

Because of identification model B, to calculating, power consumption is bigger, and equipment decides whether to enable more by monitoring cpu usage amount I.e. when heavy traffic (cpu consumption is higher) candidate's replacement is only shone and is stored, when the business free time by identification model B greatly (cpu consumption is lower) enables identification model B to do secondary verification.

The constraint condition whether replaced: 1, if there is no the period replace shine, just by the scene according to as replacement shine chase after Add and is registered in user's face database；2, if shone before in the presence of the replacement of the period, compare the period is new, old replacement according to According to the similarity score shone with the registration of the non-period by the score after mapping function, acquirement divides the higher person to make for all registrations It replaces and shines for the period.

Mapping function is as follows:

The score of F (Scores)=(1- (1-s1) * (1-s2) * ... * (1-sn)), Scores s1, s2 ..., sn combine, n Registration is indicated according to the quantity shone with the registration of the non-period, s1 to sn respectively indicates new, old replacement and takes the phase for being compared picture with n Like degree.

The present invention shines Exchange rings by registration, so that equipment possesses the ability of " half self study ", very good solution registration There are the disturbing factors such as deviation accurate to recognition of face according to style disunity, age differences, light, micro- expression shape change, facial angle The influence of rate, passage speed can increase recognition accuracy and passage speed.

Claims (5)

1. a kind of method for improving face recognition accuracy rate and passage speed, includes the following steps:

The registration of user and bottom library all personnel are shone, and in the alignment score maximum value that identification model A is obtained, i.e. top1 score is full Foot: Score_top1 >=Th, Th are default recognition threshold, indicate that user is identified by；

If at this point, Score_top1≤Th1, Th1 > Th, Th1 by user setting, indicate that the user is identified by while the user Bottom library photo differ greatly with current true man's similarity, increase or update the scene of the period according to as replacement according to being appended to this In user's face database.

2. improving the method for face recognition accuracy rate and passage speed as described in claim 1, it is characterised in that:

The scene, which is shone, can be used as replacement according to it is necessary to meet following condition:

1, scene, which is shone, meets face quality indicator, including face size, angle, fuzziness, illumination；

2, secondary verification is done with identification model B higher than identification model A accuracy rate, larger, by all registrations of the user It shines according to the replacement with the non-period and is compared respectively with what the scene was shone, maximum alignment score >=Th2, and minimum comparison point Number >=Th3, Th2 and Th3 are by user's self-setting, for constraining replacement according to the confidence level whether to come into force.

3. improving the method for face recognition accuracy rate and passage speed as claimed in claim 2, it is characterised in that:

Decide whether to enable bigger identification model B by monitoring cpu usage amount, i.e., when heavy traffic, only replaces candidate According to storage, when the business free time, enables identification model B and carry out secondary verification.

4. improving the method for face recognition accuracy rate and passage speed as claimed in claim 1 or 2, it is characterised in that: whether The constraint condition of replacement is as follows:

1, period replacement is shone if it does not exist, which is registered in user's face database according to as replacement according to addition；

If 2, being shone before in the presence of the replacement of the period, compare the period is new, old replacement according to and all registrations photograph and the non-period For the similarity score that registration is shone by the score after mapping function, acquirement divides the higher person to shine as period replacement.

5. improving the method for face recognition accuracy rate and passage speed as claimed in claim 4, it is characterised in that:

The mapping function is as follows:

F(Scores)= (1-(1-s1)*(1-s2)*...*(1-sn))

The score that wherein Scores is s1, s2 ..., sn combines, and n indicates the quantity that registration is shone and the registration of the non-period is shone, and s1 is arrived Sn respectively indicates new, old replacement and takes the similarity for being compared picture with n.