CN111177436B - Face feature retrieval method, device and equipment - Google Patents

Abstract

The present invention provides a face feature retrieval method, device and equipment, wherein the method includes: obtaining a face image to be retrieved; obtaining N groups of face feature data of the face image to be retrieved, and the N groups of face feature data are obtained from N Extract from the face images to be retrieved of different face models; search N groups of face feature data in the face feature database to obtain N groups of retrieval results; according to the pre-acquired similarity conversion relationship, N-1 The similarities of the N−1 groups of search results of the face models are respectively converted into similarities corresponding to the target face models; and the converted search results corresponding to the target face models are output. The embodiment of the present invention can improve the performance of the face retrieval system, and can also improve its algorithm accuracy and data compatibility.

Description

A face feature retrieval method, device and equipment

technical field

The present invention relates to the technical field of image recognition, in particular to a face feature retrieval method, device and equipment.

Background technique

Face recognition is a biometric technology for identification based on human facial feature information. It has the advantages of high efficiency and strong reliability in security, monitoring and other fields, so it has been widely used.

Face retrieval usually includes face image acquisition and detection, face image preprocessing, face image feature extraction, matching and recognition. Among them, the process of face image matching and recognition is to search and match the feature information of the extracted face image with the face feature data stored in the database to obtain the similarity between the two, and by setting a similarity Degree threshold to determine whether it is the same face. For example, when the similarity exceeds the threshold, it can be determined that the extracted face image and the face image stored in the database belong to the same face.

In related technologies, the face retrieval system usually adopts a single-model configuration for face recognition, that is, only one face analysis server is configured, and only one face model is stored in the server, for example: an NVR (Network Video Recorder, Network hard disk video recorder machine), a set of intelligent analysis system.

However, the configuration of a single model does not support the iterative upgrade of the face model, and cannot be expanded, resulting in poor performance of the face retrieval system.

Contents of the invention

Embodiments of the present invention provide a face feature retrieval method, device, and equipment to solve the problem of relatively poor performance in face retrieval systems in the related art.

In order to solve the above technical problems, the present invention adopts the following technical solutions:

In a first aspect, an embodiment of the present invention provides a face feature retrieval method, including:

Obtain the face image to be retrieved;

Obtaining N groups of face feature data of the face images to be retrieved, wherein the N groups of face feature data are respectively extracted from the face images to be retrieved of N types of face models;

Retrieve the N groups of face feature data respectively in the face feature database to obtain N groups of retrieval results;

According to the pre-acquired similarity conversion relationship, the similarity of N-1 groups of retrieval results of N-1 face models are respectively converted into similarities corresponding to the target face model;

Outputting the retrieval results corresponding to the converted N target face models.

In a second aspect, the embodiment of the present invention provides a human face feature retrieval device, including:

The first obtaining module is used to obtain the face image to be retrieved;

The second acquisition module is used to acquire N sets of face feature data of the face images to be retrieved, wherein the N sets of face feature data are respectively obtained from the face images to be retrieved of N types of face models extract;

A retrieval module, configured to retrieve the N groups of face feature data in the face feature database respectively to obtain N groups of retrieval results;

The conversion module is used to convert the similarity of N-1 groups of retrieval results of N-1 face models into similarities corresponding to the target face model according to the similarity conversion relationship obtained in advance;

The output module is used to output the retrieval result corresponding to the converted target face model.

In the third aspect, the embodiment of the present invention also provides a face feature retrieval device, including a processor, a memory, and a computer program stored on the memory and operable on the processor, the computer program being controlled by the When the processor executes, it realizes the steps of the human face feature retrieval method described above.

In a fourth aspect, an embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the face feature retrieval method as described above is implemented A step of.

In the embodiment of the present invention, the face image to be retrieved is obtained; N groups of face feature data of the face image to be retrieved are obtained, wherein the N groups of face feature data are obtained from N types of face models respectively. Extract from the face images to be retrieved; search the N groups of face feature data in the face feature database to obtain N groups of retrieval results; The similarities of the retrieval results of the N−1 groups of models are respectively converted into similarities corresponding to the target face model; and the converted retrieval results corresponding to the target face model are output. In this way, multiple sets of retrieval results corresponding to various face models can be obtained, and the similarity in multiple sets of retrieval results can be converted into similarity corresponding to the same human face model, and the converted similarity retrieval results can be output, so that These retrieval results are comparable, thereby improving the performance of the face retrieval system; in addition, N types of face models are compatible with different face models, thereby improving the data compatibility of the face retrieval system, and using multiple face models can also Improve the accuracy of the algorithm.

Description of drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the accompanying drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only some of the present invention. Embodiments, for those of ordinary skill in the art, other drawings can also be obtained based on these drawings without any creative effort.

Fig. 1 is a structural diagram of a heterogeneous cluster system;

Fig. 2 is the flow chart of a kind of face feature retrieval method that the embodiment of the present invention provides;

Fig. 3 is a structural diagram of a human face feature retrieval device provided by an embodiment of the present invention;

4 is a structural diagram of another face feature retrieval device provided by an embodiment of the present invention;

Fig. 5 is a structural diagram of a face feature retrieval device provided by an embodiment of the present invention.

Detailed ways

The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some of the embodiments of the present invention, but not all of them. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without creative efforts fall within the protection scope of the present invention.

It should be noted that the face feature retrieval method provided by the embodiment of the present invention can be applied to the heterogeneous monitoring system shown in FIG. 1 . The system includes N smart terminals, wherein the smart terminals may also be called back-end intelligence, and each smart terminal may be configured with a face model.

Specifically, as shown in FIG. 1 , the system includes: a heterogeneous intelligent cluster 101, a face checkpoint 102 respectively connected to the heterogeneous intelligent cluster 101, a monitoring camera 103, a business scheduling 104, a picture storage library 105, and a large database 106 .

In this embodiment, the heterogeneous intelligent cluster 101 includes four intelligent terminals 1011, and the face mount 102 and the surveillance camera 103 are respectively used to provide image information such as face pictures, surveillance pictures, and surveillance videos to the four intelligent terminals 1011, The smart terminal 1011 can automatically obtain a large face image according to the image information and intercept a small face image from the large face image, and store the intercepted small face image and large face image in the image storage library 105 . And use the face model configured in the smart terminal 1011 to extract the face feature data of the small face image respectively, and store it in the large database 105 in the form of semi-structured data.

After the monitoring personnel input the face image to be retrieved through the business scheduling 104, the face feature data of the face image to be retrieved are respectively extracted through the four intelligent terminals 1011, and the face feature data of the face image to be retrieved are stored in the The facial feature data in the database 105 with the same identification information are subjected to similarity calculation to obtain retrieval results.

Wherein, the retrieval result may include the similarity between the face image to be retrieved and each small face image in the image information provided by the face checkpoint 102 and the monitoring camera 103 .

In addition, the search results can also be output through the service dispatcher 104 or other interfaces, and displayed on a display device.

It should be noted that since the face feature data of the monitored face image can be extracted using different face models, and the accuracy, dimensions, and algorithms used may be different between different face models. If the face feature data obtained by different face models are directly retrieved, the wrong similarity will be obtained.

For example, the dimension of one face model is 128, and the dimension of another face model is 256. In the process of calculating the similarity, the 128-dimensional face model will extract the 128-dimensional vector from a face image. Eigenvalue, the 256-dimensional face model will extract the eigenvalue of the 256-dimensional vector based on another face image. Obviously, the 128-dimensional face feature data cannot be directly compared with the 256-dimensional face feature data. out the similarity. The present invention distinguishes the face feature data obtained by different face models, and only calculates the similarity of the face feature data obtained by using the same face model to obtain the retrieval result.

Of course, the heterogeneous monitoring system shown in FIG. 1 is only used to illustrate an application scenario of the facial feature retrieval method provided by the embodiment of the present invention, and does not limit other application scenarios and specific steps of the facial feature retrieval method.

Please refer to FIG. 2 , which is a flow chart of a face feature retrieval method provided by an embodiment of the present invention. As shown in Figure 2, the method includes the following steps:

Step 201. Obtain a face image to be retrieved.

Wherein, the face image to be retrieved may be the face image input by the user, or the face image received from other devices, for example: the face image input by the user or an external device through the service scheduling 104 shown in FIG. 1 .

Of course, the acquired face image to be retrieved is not limited to the face image input by the user, it can also be a face image obtained by the retrieval system through shooting, analysis, etc. If it is determined that a person's behavior is dangerous by means of analysis and needs to be monitored, the face image of the dangerous person will be automatically obtained.

In addition, the above-mentioned face image to be retrieved may include an image of a face part, and may also include other content other than the face part, such as a background or a body.

As an optional implementation manner, before step 201, the method further includes the following steps:

Obtaining a monitoring face image through an image acquisition unit;

Obtaining N groups of face feature data of the monitoring face image, wherein the N groups of face feature data of the monitoring face image are respectively extracted from the monitoring face images of N kinds of face models;

The N sets of face feature data of the monitored face images are stored in the face feature database.

In specific application scenarios, image acquisition units such as network cameras and monitoring cameras installed in the community can be used to collect face images in the monitoring scene, and match the collected face images with the face images to be retrieved to obtain The similarity between the two, when the similarity is greater than a certain threshold, the two are judged to be the same person, so as to achieve security measures such as monitoring or tracking of the search personnel.

Of course, the image directly acquired by the image acquisition unit may be a large face image, and the monitored face image is obtained by intercepting a part of the face in the large face image as a small face image.

Wherein, the small face image may be a picture of only the face part intercepted from the large face image, and the large face image may also include other content besides the face part.

In addition, the obtained large face image and small face image can also be stored in the image library.

During the application process, the monitored face images are often numerous. The N types of face models can be used to extract the face feature data of one monitoring face image respectively, and repeat in sequence until the face feature data of all the monitoring face images are extracted. The extraction speed of the face feature data of the monitored face image can be improved, so as to improve the overall efficiency of the face feature retrieval method.

In this embodiment, N kinds of face models are used to extract the face feature data of multiple monitoring face images respectively, avoiding the consumption of a large amount of time by using one face model to sequentially extract the face feature data of multiple monitoring face images , thus improving the efficiency of the face feature retrieval method.

As an optional implementation, the face feature retrieval method also includes:

Preferably, the face model with the highest accuracy among the N types of face models is used to extract the face feature data of the monitoring face images in the most important monitoring scene.

Among them, the most important monitoring scene can be the entrance of the community, around the school, around the government department and other important places or places with a large flow of people, which can be set according to the specific actual situation.

In addition, the face model with the highest accuracy is generally the face model with the latest version.

In this implementation manner, a highly accurate face model can be configured for an image acquisition unit that requires high monitoring accuracy, thereby improving the pertinence in the process of extracting face feature data.

As an optional implementation manner, the identification information of the face model may include a version number.

For example, as shown in Table 1 below, the version number includes 11 digits, the first 4 digits are used to distinguish the face model, the middle 4 digits are used to represent the dimension of the face model, and the last 3 digits are not used to represent the face model version grade:

Table 1

Face model (4 bits) dimension (4 bits) Version (3 digits) Version number (11 digits) 231 128 V1.0 02320128010 231 128 V1.1 02320128011 245 512 V1.0 02450513010 247 2024 V1.1 02472024011 280 128 V1.1 0280128011

In the retrieval process, the version number of each group of face feature data can be obtained first, and two or more sets of face feature data with the same version number can be compared and analyzed, and according to the version number, it can be easily analyzed and extracted. Version information and dimension information of the face model of the face feature data of the version number.

In this embodiment, the version number is used as the identification information, so that information such as the face model, the dimensions of the face model, and the version level of the face model can be easily read.

Step 202. Obtain N sets of face feature data of the face image to be retrieved, wherein the N sets of face feature data are respectively extracted from the face images of the face image to be searched in N types of face models.

Wherein, N may be an integer greater than 1.

It should be noted that, in order to facilitate the retrieval of the face images to be retrieved in the face images in the image database, the face images need to be converted into face feature data first, and then the size, Compare the differences, etc., to determine the similarity between the face image to be retrieved and the face image in the image database.

Wherein, the face feature data may include a plurality of face feature values.

In addition, the N types of face models can be expressed as, each of the N types of face models is a face model with a different algorithm, for example: different types of algorithms, different old and new versions of algorithms, different algorithms The precision or dimensions are not the same.

For example: using the first face model and the second face model to extract the face feature data of the face image to be retrieved respectively, wherein the first face model obtains the first face feature data of the face image to be retrieved, and the first face model obtains the first face feature data of the face image to be retrieved, and the second face model The two-person face model obtains the second face feature data of the face image to be retrieved, and the first face model is different from the second face model, so the obtained first face feature data is different from the second face feature data. The data is also different.

Specifically, the N types of face models can be respectively installed in the N intelligent terminals in the heterogeneous system, or the N types of face models can be installed in one system.

It should be noted that the types of face models may be neural network models, deep machine learning, and the like. According to the progress of the times, the version of the face model also needs to be updated, so there is a situation that the versions of the face model before and after the update are not the same.

Moreover, when the accuracy requirement of face feature data is low, a face model with low accuracy or dimension can be configured; when the accuracy requirement of face feature data is high, the configuration accuracy or dimension is high face model.

In this way, in the case of meeting the accuracy requirements of face feature data, low-precision or low-dimensional face models can be used as much as possible to avoid the excessive complexity of the calculation process of the face model caused by excessive accuracy or dimensions. reduce computational efficiency.

Of course, the face image to be retrieved obtained in step 201 can be a large face image, and a small face image can be intercepted from the large face image through steps such as interception, and step 201 can also include obtaining the large face image and the The face thumbnails are stored in the picture library. Step 202 may refer to obtaining N sets of face feature data of the face thumbnails through N types of face models.

In this step, at least two face models are used to extract the face feature data of the face image to be retrieved respectively, and in the process of upgrading part of the face models, it is convenient to continue to extract the face feature data based on the original face model, In this way, it is avoided that a system using a single face model needs to re-analyze a large amount of data during the upgrade process, resulting in the inability to extract face feature data during the upgrade process.

Step 203: Retrieve the N groups of face feature data in the face feature database to obtain N groups of search results.

Wherein, the retrieving may refer to retrieving face feature data with the same face model.

It should be noted that a large number of face images and face feature data of each face image can be stored in the face retrieval system, and each set of retrieval results can include multiple sets of face feature data and face feature data in the face feature database. The similarity between the data may also include the corresponding face image, and of course, may also include the identification information of the face model corresponding to each set of retrieval results.

In addition, the face feature data of the face images in the face feature database can also be extracted by using the above N types of face models.

It should be noted that since the face feature data of the monitored face image may be extracted using different face models, and the face feature data extracted according to different face models may have different similarity algorithms. Therefore, the face feature data with the same identification information is calculated using a similarity algorithm matched with the face model to calculate the similarity.

For example, the dimension of one face model is 128, and the dimension of another face model is 256. In the process of calculating the similarity, it is obvious that the 128-dimensional face feature data and the 256-dimensional face feature data adopt The similarity algorithm is different.

In this step, since the face feature data extracted by the same face model and the face feature data in the face feature database are used for retrieval, the incomparable face feature data extracted by different face models are avoided. The comparison avoids the mixing of face feature data, and improves the accuracy and reliability of face feature retrieval.

As an optional implementation, in order to retrieve the face feature data with the same face model, each set of face feature data in the N sets of face feature data includes the identification information of the face model it uses , each group of face feature data in the face feature database includes the identification information of the face model it uses, and step 203 is specifically:

According to the identification information of the face model used by each set of face feature data, the N sets of face feature data are respectively searched in the face feature database to obtain N sets of search results.

For example: the face model includes the first face model and the second face model, there is one face image to be retrieved, and there are 10 monitoring face images in the face feature database, then the two face models are used to extract Face feature data of 10 monitored face images to obtain 10 sets of face feature data corresponding to the monitored face images one-to-one, of which 4 sets of face feature data are extracted by the first face model, including the first identification information ; The other 6 groups of face feature data are extracted by the second face model, including the second identification information. In addition, two face models are used to retrieve the face feature data of the face image to obtain two sets of face feature data, one set includes the first identification information, and the other set includes the second identification information.

During the retrieval process, the similarity calculation is performed between the face feature data of the four groups of monitoring face images including the first identification information and the face feature data of the face image to be retrieved including the first identification information, and the second The face feature data of the 6 groups of monitoring face images of the identification information and the face feature data of the face image to be retrieved including the second identification information are subjected to similarity calculations to obtain the face image to be retrieved and each monitoring face image similarity between.

In addition, the identification information can be in any form of numbers, codes, characters, etc., and is used to distinguish facial feature data obtained from different facial models, so as to avoid comparing two incomparable facial features extracted by different facial models. The facial feature data is compared, resulting in an incorrect similarity or the inability to perform the comparison process. The method of using identification information to distinguish face feature data extracted from different face models can improve the accuracy of the face feature detection result.

In this embodiment, the identification information of the face model used is added to the face feature data, so that only the face feature data with the same identification information are calculated for similarity, avoiding the use of different face models. The feature data are not comparable or the similarity calculation methods are different, resulting in wrong retrieval results, thereby improving the pertinence and reliability of the retrieval results.

Certainly, in the embodiment of the present invention, it is not limited that the face feature data includes the identification information of the face model. It is assumed that N groups of face feature data are acquired sequentially, so that the face model used by each set of face feature data can be determined in this order, so that in the retrieval process, the face feature data with the same face model can be retrieved.

Step 204 , according to the pre-acquired similarity conversion relationship, the similarities of the N-1 groups of retrieval results of the N-1 types of face models are respectively converted into similarities corresponding to the target face model.

Wherein, the similarity conversion relationship may be a conversion relationship that converts the similarity corresponding to N-1 kinds of face models into the similarity corresponding to the target face model, and the target face model is the N Another face model in the face models other than the N-1 face models.

Among them, the N groups of retrieval results obtained in step 203 are respectively the retrieval results corresponding to the N types of face models, and the algorithms and precisions adopted by the N types of face models are different, so that the similarity in the N groups of retrieval results is Not comparable.

For example: use two face models with different algorithms to extract the face feature data of two face images respectively, and compare the two sets of face feature data extracted by the same face model to obtain a comparison between the two face features. Two similarities corresponding to the face models respectively. At this time, since the algorithms of the two face models are different, the obtained two similarity values are not equal, but these two similarities both represent the similarity between the same two face images. It can be seen that the similarity obtained by using different face models is not comparable. After converting the similarities in the N groups of retrieval results to the same standard through step 204, the similarities in the N groups of retrieval results can be made comparable.

Wherein, the target face model may be a face model with the latest version or the highest accuracy among the N types of face models.

In addition, the pre-acquired similarity conversion relationship may include a numerical conversion relationship between the similarity of N−1 groups of retrieval results and the similarity corresponding to the target face model. After substituting the similarities of N-1 groups of retrieval results, the numerical conversion relationship can output the similarity corresponding to the target face model, so that the similarity of the retrieval results of the target face model is the same as the converted N-1 The similarities are comparable.

For example: using two different face models, the first face model and the second face model, respectively, to calculate the similarities between the first face image and the second face image and the third face image.

Assume that the similarities between the first face image and the second face image calculated by the first face model and the second face model are respectively equal to: 90%, 89%; the first face image and the third person The similarities between face images are equal to: 65%, 67%, respectively. Moreover, the second human face model whose calculated similarity between the first human face image and the second human face image is equal to 89% has higher accuracy, so this human face model is selected as the target human face model.

In addition, the conversion relationship between the similarity between the first human face model and the second human face model is obtained according to the conversion relationship between 90% and 89%, and between 65% and 67%.

In this way, after the first face model calculates the similarity between the other two face images, the first face can be directly converted according to the conversion relationship between the first face model and the second face model. The similarity calculated by the model is converted into the similarity corresponding to the second face model; according to the conversion relationship between the similarity between the first face model and the third face model, the similarity calculated by the third face model Convert to the similarity corresponding to the first face model.

Therefore, the similarity obtained by different face models is comparable.

Of course, in order to ensure the accuracy of the conversion relationship, two groups of similarities with multiple similarity values can be used for regression calculation to obtain a more accurate conversion relationship.

As an optional implementation, the above conversion relationship includes N-1 regression models, wherein the N-1 regression models respectively represent the similarity between the N-1 types of face models and the target person. The transformation relationship of the similarity corresponding to the face model.

Among them, the number of regression models is the number of face models and minus 1, which means that the regression model corresponds to N-1 kinds of face models one by one, so that in the process of similarity conversion, N-1 kinds of face models can be The obtained similarities are respectively substituted into a corresponding regression model to obtain the similarity corresponding to the target face model.

And repeat this until the similarities obtained by the N-1 face models are converted into the corresponding similarities of the target face model, so as to make the similarities obtained by the N face models comparable.

In this embodiment, N-1 regression models are used to convert the similarities of N-1 types of face models into similarities corresponding to target face models, so that the similarities obtained by N types of face models have a similarity. comparable effect.

Of course, the conversion relationship can also be in other forms such as codes and parameters, for example: the correction code obtained according to the differences in algorithms and dimensions between different face models, the correction code can be based on the algorithm between different face models , dimension and other differences to correct the obtained similarity to the similarity corresponding to the face model with the same algorithm and dimension.

As an optional implementation manner, the N-1 regression models include a first regression model, the independent variable of the first regression model is the similarity corresponding to the first face model, and the first regression model's The dependent variable is the similarity corresponding to the target face model output by the independent variable through the first regression model, and the first face model is one of the N-1 types of face models Model;

The step of converting the similarity of N-1 groups of retrieval results of N-1 face models into the similarity corresponding to the target face model according to the pre-acquired similarity conversion relationship includes:

Substituting the similarity corresponding to the first human face model in the N-1 group of retrieval results into the independent variable of the first regression model, and calculating the dependent variable of the first regression model, so as to obtain The similarity corresponding to the target face model after the corresponding similarity conversion.

Among them, the similarity degree corresponding to the first face model indicates that two sets of face feature data of two face images are respectively extracted by using the first face model, and the two sets of face feature data calculated by the similarity calculation method are Similarity between data.

Of course, the N−1 regression models also include the second regression model, the third regression model up to the N−1 regression model.

And the second regression model, the third regression model up to the N-1th regression model are respectively used to convert the similarity corresponding to the N-1 face models into the similarity corresponding to the target face model .

In this embodiment, only the similarity of the first face model needs to be input as the independent variable of the first regression model, and the value of the dependent variable of the first regression model can be obtained, wherein the value of the dependent variable is the target The similarity corresponding to the face model. Thus, the step of calculating the similarity corresponding to the target face model is simplified.

As an optional implementation manner, the first regression model is determined in the following manner:

The second face feature data of the first face image set and the second face image set are respectively provided through the target face model, wherein the first face image set includes n face images, and the first face image set includes n face images. The two face image collection includes m face images, and the n and the m are both integers greater than 1;

Obtain n pairs of regression sample images, wherein each pair of regression sample images includes a face image belonging to the first human face image set and the second human face image set respectively, and the nth pair of regression sample images A first similarity degree of face image feature data is linearly distributed;

Extracting the second face feature data of the n pairs of regression sample images respectively through the first face model, and calculating the second similarity of the second face feature data of each pair of regression sample images;

According to the first similarity and second similarity of the n pairs of regression sample images, the first regression model is determined, wherein the independent variable of the first regression model is the second similarity, and the dependent variable is the second similarity. the first similarity.

Of course, the second regression model, the third regression model up to the N−1th regression model can also be determined in the above manner.

Among them, the regression model can be represented by the formula y _i = β ₀ + β ₁ x _i + μ _i , where xi is the independent variable, _{y i is} the dependent variable, i represents the i-th regression model, and i ranges from 1 to N-1 Any value between , β ₀ and β ₁ are estimators, and μ _i is a random error. The specific values of the parameters β ₀ , β ₁ , and μ _i in the formula can be determined by a regression algorithm.

It should be noted that β ₀ and μ _i can take any value including 0, so the formula of the regression model is not limited to y _i =β ₀ +β ₁ x _i +μ _i , for example: in β ₀ When equal to 0, the formula of the regression model is y _i =β ₁ x _i +μ _i .

For example: suppose the first face image set includes 10 face images, and the second face image set includes 100 face images, and the new model is used to calculate the similarity between 10 face images and 100 face images , to obtain 10×100 first similarities, after linearly arranging the 1000 first similarities, find n similarity values ( For example, 30, the similarity values are: 61%, 62%, 63%, ..., 90%), and 30 pairs of regression sample images corresponding to the 30 similarity values are obtained. Then, 30 pairs of regression sample images are used to extract facial feature data using the old model, and 30 second similarities are obtained according to the extracted facial feature data.

Taking the 30 first similarities as the independent variable xi and the corresponding 30 second similarities as the dependent variable y _i , and substituting them into _the formula y _i =β ₀ +β ₁ x _i +μ _i respectively, we can find The specific values of the parameters β ₀ , β ₁ , μ _i in the formula. resulting in a regression model.

Wherein, the least square method or the maximum likelihood method may be used to solve the formula.

The solution procedure can be, assuming:

can get:

Wherein, the random error μ _i is assumed to be 0, and the corresponding unary thread equation can be obtained.

Of course, multiple regression equations can also be established to establish regression models.

Afterwards, through this regression equation, the similarity obtained by two or more face models can be converted into the corresponding similarity of the latest model, and the feedback can be sorted and sorted according to the converted similarity value to achieve the purpose of fusion retrieval .

The sample table shown in Table 2 below, in which the similarity of the same pair of regression sample images under different models is selected, wherein, assuming that the above solution method is used, the regression linear equation corresponding to the regression model is obtained: y _i =0.00484+1.0936 x _i +μ _i :

Table 2

serial number old model new model Regression equation calculation Deviation μ _i 1 0.2210 0.2730 0.246512 0.026475 2 0.2673 0.3043 0.297182 0.007137 3 0.3137 0.3847 0.347852 0.036800 4 0.3560 0.4180 0.394153 0.023837 5 0.4042 0.4752 0.446897 0.028323 6 0.4525 0.5045 0.499642 0.004808 7 0.5020 0.5860 0.553830 0.032170 8 0.5231 0.6041 0.576905 0.027195 9 0.5587 0.5797 0.615837 (0.036137) 10 0.5643 0.6313 0.621962 0.009338 11 0.5794 0.6584 0.638475 0.019925 12 0.6084 0.6724 0.670190 0.002210 13 0.6343 0.6653 0.698514 (0.033214)

In addition, the same threshold can also be set for the converted similarity, and the face images corresponding to the similarity greater than the threshold are judged as the same person.

It should be noted that the greater the number of regression sample images, the higher the accuracy of the regression model obtained. In addition, after the regression equation is obtained, the value of the corresponding estimated deviation μ _i can be calculated by comparing the values shown in Table 2 The sample table is obtained by looking up the table.

In this step, the similarity of N-1 groups of retrieval results of N-1 face models is converted into the similarity corresponding to the target face model, so that the similarities obtained by using different face models are comparable . It is convenient for monitoring personnel to accurately judge whether the face images are of the same person only according to the numerical relationship between the transformed similarities.

Step 205, outputting the retrieval result corresponding to the converted target face model.

Wherein, step 205 can be understood as: outputting converted N−1 groups of retrieval results among the N groups of retrieval results, and another non-transformed retrieval result among the N groups of retrieval results.

Wherein, according to the converted N-1 group of retrieval results among the N groups of retrieval results and another non-converted retrieval result, the converted N-1 group of retrieval results can be output in descending order of similarity , and the retrieval result corresponding to the target face model.

In this way, it is convenient for monitoring personnel to obtain detection results with high similarity more clearly.

In addition, the output retrieval result can also be a retrieval result corresponding to a similarity greater than a certain value. For example, if the similarity is less than 50%, the possibility that the two are the same person can be excluded, and only the similarity greater than or equal to 50 can be output. % of search results.

In this step, the converted N−1 groups of retrieval results and the retrieval results corresponding to the target face model are output, so as to facilitate identification by monitoring personnel.

In the embodiment of the present invention, the face image to be retrieved is obtained; N groups of face feature data of the face image to be retrieved are obtained, wherein the N groups of face feature data are obtained from N types of face models respectively. Extract from the face images to be retrieved; search the N groups of face feature data in the face feature database to obtain N groups of retrieval results; The similarities of the retrieval results of the N−1 groups of models are respectively converted into similarities corresponding to the target face model; and the converted retrieval results corresponding to the target face model are output. In this way, multiple sets of retrieval results corresponding to various face models can be obtained, and the similarity in multiple sets of retrieval results can be converted into similarity corresponding to the same human face model, and the converted similarity retrieval results can be output, so that These retrieval results are comparable, thereby improving the performance of the face retrieval system.

In addition, in the embodiment of the present invention, since multiple face models are included, and the similarity in multiple sets of retrieval results can be converted into the similarity corresponding to the same face model, it can also solve the problem of the old data after the new model is upgraded. The problem of incompatibility between the similarity of the old model in the old data and the similarity of the new model caused by the fact that it does not exist in the new model can improve the data compatibility of the face retrieval system, and the use of multiple face models can also improve the algorithm the accuracy.

Please refer to FIG. 3 , which is a structural diagram of a face feature retrieval device provided by an embodiment of the present invention. The face feature retrieval device 300 includes:

The first obtaining module 301 is used to obtain the face image to be retrieved;

The second acquiring module 302 is configured to acquire N sets of face feature data of the face images to be retrieved, wherein the N sets of face feature data are the face images to be retrieved from N types of face models respectively extract from;

The retrieval module 303 is used to retrieve the N groups of facial feature data respectively in the facial feature database to obtain N groups of retrieval results;

The conversion module 304 is used to convert the similarities of N-1 groups of retrieval results of N-1 types of face models into similarities corresponding to the target face model according to the similarity conversion relationship obtained in advance;

The output module 305 outputs the converted retrieval result corresponding to the target face model.

Wherein, N may be an integer greater than 1.

In addition, the retrieving may refer to retrieving face feature data with the same face model.

In addition, the similarity conversion relationship may be a conversion relationship that converts the similarity corresponding to N-1 types of face models into the similarity corresponding to the target face model, and the target face model is the N Another face model in the face models other than the N-1 face models.

The content output by the output module 305 can also be referred to as: the N-1 group of search results after conversion among the N groups of search results, and another search result without conversion among the N groups of search results.

Optionally, each set of face feature data in the N sets of face feature data includes the identification information of the face model used by it, and each set of face feature data in the face feature database includes the face model used by it. The identification information of the face model, the retrieval module 304 is specifically used for:

According to the identification information of the face model used by each set of face feature data, the N sets of face feature data are respectively searched in the face feature database to obtain N sets of search results.

Optionally, the pre-acquired similarity conversion relationship includes:

N-1 regression models, wherein the N-1 regression models respectively represent the conversion relationship between the similarity corresponding to the N-1 types of human face models and the similarity corresponding to the target human face model.

Optionally, the N-1 regression models include a first regression model, the independent variable of the first regression model is the similarity corresponding to the first face model, and the dependent variable of the first regression model is the The similarity corresponding to the target face model output by the independent variable through the first regression model, the first face model is one of the N-1 face models;

The conversion module 304 is specifically used for:

Substituting the similarity corresponding to the first human face model in the N-1 group of retrieval results into the independent variable of the first regression model, and calculating the dependent variable of the first regression model, so as to obtain The similarity corresponding to the target face model after the corresponding similarity conversion.

Optionally, the first regression model is determined in the following manner:

The second face feature data of the first face image set and the second face image set are respectively provided through the target face model, wherein the first face image set includes n face images, and the first face image set includes n face images. The two face image collection includes m face images, and the n and the m are both integers greater than 1;

Obtain n pairs of regression sample images, wherein each pair of regression sample images includes a face image belonging to the first human face image set and the second human face image set respectively, and the nth pair of regression sample images A first similarity degree of face image feature data is linearly distributed;

Extracting the second face feature data of the n pairs of regression sample images respectively through the first face model, and calculating the second similarity of the second face feature data of each pair of regression sample images;

According to the first similarity and second similarity of the n pairs of regression sample images, the first regression model is determined, wherein the independent variable of the first regression model is the second similarity, and the dependent variable is the second similarity. the first similarity.

Optionally, as shown in FIG. 4, the device 300 further includes:

The third acquiring module 306 is used to acquire the monitor face image through the image acquisition unit;

The fourth acquisition module 307 is used to acquire N sets of face feature data of the monitored face image, wherein, the N sets of face feature data of the monitored face image are obtained from the monitor through N types of face models respectively. extracted from face images.

A storage module 308, configured to store the N sets of fourth face feature data in the face feature database.

Optionally, the identification information of the face model includes a version number.

It should be noted that the face feature retrieval device provided by the embodiment of the present invention can realize each step in the method embodiment shown in FIG. 1 or FIG. 2 and achieve the same beneficial effect, so it is not repeated here to avoid repetition.

Please refer to FIG. 5 , which is a composition diagram of a human face feature retrieval device provided by an embodiment of the present invention. As shown in Figure 5, this human face feature retrieval device comprises: transceiver 501, processor 502, memory 503 and are stored on the memory 503 and can run on the computer program 5031 of processor 502, computer program 5031 is processed by processor 502 The following process is realized during execution:

Wherein, the transceiver 501 is used for: obtaining the face image to be retrieved;

Processor 502, configured to acquire N sets of face feature data of the face images to be retrieved, wherein the N sets of face feature data are respectively extracted from the face images to be retrieved of N types of face models .

The processor 502 is further configured to search the N sets of face feature data respectively in the face feature database to obtain N sets of search results.

The processor 502 is further configured to convert the similarities of the N-1 groups of retrieval results of the N-1 types of face models into similarities corresponding to the target face model according to the pre-acquired similarity conversion relationship.

The transceiver 501 outputs the converted retrieval result corresponding to the target face model.

Optionally, each set of face feature data in the N sets of face feature data includes the identification information of the face model used by it, and each set of face feature data in the face feature database includes the face model used by it. For the identification information of the face model, the processor 502 executes the step of retrieving the N groups of face feature data respectively in the face feature database to obtain N groups of retrieval results, specifically including:

According to the identification information of the face model used by each set of face feature data, the N sets of face feature data are respectively searched in the face feature database to obtain N sets of search results.

Optionally, the pre-acquired similarity conversion relationship includes:

N-1 regression models, wherein the N-1 regression models respectively represent the conversion relationship between the similarity corresponding to the N-1 types of human face models and the similarity corresponding to the target human face model.

Optionally, the N-1 regression models include a first regression model, the independent variable of the first regression model is the similarity corresponding to the first face model, and the dependent variable of the first regression model is the The similarity corresponding to the target face model output by the independent variable through the first regression model, the first face model is one of the N-1 face models;

Processor 502 executes: according to the pre-acquired similarity conversion relationship, the step of converting the similarity of N-1 groups of retrieval results of N-1 face models into the similarity corresponding to the target face model, specifically including:

Substituting the similarity corresponding to the first human face model in the N-1 group of retrieval results into the independent variable of the first regression model, and calculating the dependent variable of the first regression model, so as to obtain The similarity corresponding to the target face model after the corresponding similarity conversion.

Optionally, the first regression model is determined in the following manner:

The first human face feature data of the first human face image set and the second human face image set are respectively provided through the target human face model, wherein the first human face image set includes n pieces of human face images, and the first human face image set includes n pieces of human face images, and the first human face image set includes n face images, and The two face image collection includes m face images, and the n and the m are both integers greater than 1;

Obtain n pairs of regression sample images, wherein each pair of regression sample images includes a face image belonging to the first human face image set and the second human face image set respectively, and the nth pair of regression sample images A first similarity degree of face image feature data is linearly distributed;

Extracting the second face feature data of the n pairs of regression sample images respectively through the first face model, and calculating the second similarity of the second face feature data of each pair of regression sample images;

According to the first similarity and second similarity of the n pairs of regression sample images, the first regression model is determined, wherein the independent variable of the first regression model is the second similarity, and the dependent variable is the second similarity. the first similarity.

Optionally, before the transceiver 501 acquires the face image to be retrieved, it is also used to:

Obtaining a monitoring face image through an image acquisition unit;

The processor 502 is further configured to obtain N sets of face feature data of the monitored face image, wherein the N sets of face feature data of the monitored face image are the monitored person from N types of face models respectively. Extraction from face images;

The processor 502 is further configured to store N sets of face feature data of the monitored face image in the face feature database.

Optionally, the identification information of the face model includes a version number.

The embodiment of the present invention can implement any step in the method embodiment corresponding to FIG. 1 or FIG. 2 , and can achieve the same beneficial effect. To avoid repetition, details are not repeated here.

Those of ordinary skill in the art can understand that all or part of the steps for implementing the methods of the above embodiments can be completed by program instructions related hardware, and the program can be stored in a computer-readable medium. An embodiment of the present invention also provides a computer storage medium, on which a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the face feature retrieval method as shown in FIG. 1 or FIG. 2 is realized. steps, and can achieve the same beneficial effect, in order to avoid repetition, it is not repeated here.

In the several embodiments provided in this application, it should be understood that the disclosed methods and devices may be implemented in other ways. For example, the device embodiments described above are only illustrative. For example, the division of the modules is only a logical function division. In actual implementation, there may be other division methods. For example, multiple modules or components can be combined or May be integrated into another system, or some features may be ignored, or not implemented. In another point, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces, and the indirect coupling or communication connection of devices or modules may be in electrical, mechanical or other forms.

The above-mentioned integrated modules implemented in the form of software function modules may be stored in a computer-readable storage medium. The above-mentioned software function modules are stored in a storage medium, and include several instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute some steps of the face feature retrieval method described in various embodiments of the present invention . The aforementioned storage media include: U disk, mobile hard disk, read-only memory (Read-Only Memory, referred to as ROM), random access memory (Random Access Memory, referred to as RAM), magnetic disk or optical disk, etc., which can store program codes. medium.

The above is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Anyone familiar with the technical field can easily think of changes or replacements within the technical scope disclosed in the present invention, and should cover all Within the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims (10)

1. A face feature retrieval method, characterized in that, comprising: Obtain the face image to be retrieved; Obtaining N groups of face feature data of the face images to be retrieved, wherein the N groups of face feature data are respectively extracted from the face images to be retrieved of N types of face models; Retrieve the N groups of face feature data respectively in the face feature database to obtain N groups of retrieval results; According to the pre-acquired similarity conversion relationship, the similarities of N-1 groups of retrieval results of N-1 face models are respectively converted into similarities corresponding to the target face model, wherein the pre-acquired similarity conversion The relationship includes: N-1 regression models, the N-1 regression models respectively represent the conversion relationship between the similarity corresponding to the N-1 face models and the similarity corresponding to the target face model, the The target face model is another face model other than the N-1 face models among the N types of face models; Outputting the retrieval result corresponding to the converted target face model. 2. The method according to claim 1, wherein the N-1 regression models include a first regression model, and the independent variable of the first regression model is the corresponding similarity of the first human face model, so The dependent variable of the first regression model is the similarity corresponding to the target face model output by the independent variable through the first regression model, and the first face model is the N-1 types of face models A face model in ; The step of converting the similarity of N-1 groups of retrieval results of N-1 face models into the similarity corresponding to the target face model according to the pre-acquired similarity conversion relationship includes: Substituting the similarity corresponding to the first human face model in the N-1 group of retrieval results into the independent variable of the first regression model, and calculating the dependent variable of the first regression model, so as to obtain The similarity corresponding to the target face model after the corresponding similarity conversion. 3. The method according to claim 2, wherein the first regression model is determined in the following manner: The first human face feature data of the first human face image set and the second human face image set are respectively provided through the target human face model, wherein the first human face image set includes n pieces of human face images, and the first human face image set includes n pieces of human face images, and the first human face image set includes n face images, and The two face image collection includes m face images, and the n and the m are both integers greater than 1; Obtain n pairs of regression sample images, wherein each pair of regression sample images includes a face image belonging to the first human face image set and the second human face image set respectively, and the nth pair of regression sample images A first similarity degree of face image feature data is linearly distributed; Extracting the second face feature data of the n pairs of regression sample images respectively through the first face model, and calculating the second similarity of the second face feature data of each pair of regression sample images; According to the first similarity and second similarity of the n pairs of regression sample images, the first regression model is determined, wherein the independent variable of the first regression model is the second similarity, and the dependent variable is the second similarity. the first similarity. 4. according to the method described in any one in claim 1 to 3, it is characterized in that, before described obtaining to be retrieved face image, described method also comprises: Obtaining a monitoring face image through an image acquisition unit; Obtaining N groups of face feature data of the monitoring face image, wherein the N groups of face feature data of the monitoring face image are respectively extracted from the monitoring face images of N kinds of face models; The N sets of face feature data of the monitored face images are stored in the face feature database. 5. A face feature retrieval device, characterized in that, comprising: The first obtaining module is used to obtain the face image to be retrieved; The second acquisition module is used to acquire N sets of face feature data of the face images to be retrieved, wherein the N sets of face feature data are respectively obtained from the face images to be retrieved of N types of face models extract; A retrieval module, configured to retrieve the N groups of face feature data in the face feature database respectively to obtain N groups of retrieval results; The conversion module is used to convert the similarities of N-1 groups of retrieval results of N-1 face models into similarities corresponding to the target face model according to the similarity conversion relationship obtained in advance, wherein the pre-acquired The obtained similarity conversion relationship includes: N-1 regression models, the N-1 regression models respectively represent the difference between the similarity corresponding to the N-1 face models and the similarity corresponding to the target face model Conversion relationship, the target face model is another face model of the N types of face models except the N-1 types of face models; The output module is used to output the retrieval result corresponding to the converted target face model. 6. The device according to claim 5, wherein the N-1 regression models include a first regression model, and the independent variable of the first regression model is the similarity corresponding to the first human face model, so The dependent variable of the first regression model is the similarity corresponding to the target face model output by the independent variable through the first regression model, and the first face model is the N-1 types of face models A face model in ; The conversion module is used for substituting the similarity corresponding to the first face model in the N-1 groups of retrieval results into the independent variable of the first regression model, and calculating the dependent variable of the first regression model, so as to obtain the The similarity corresponding to the target human face model after the conversion of the similarity corresponding to the first human face model. 7. The device according to claim 6, wherein the first regression model is determined as follows: The first human face feature data of the first human face image set and the second human face image set are respectively provided through the target human face model, wherein the first human face image set includes n pieces of human face images, and the first human face image set includes n pieces of human face images, and the first human face image set includes n face images, and The two face image collection includes m face images, and the n and the m are both integers greater than 1; Obtain n pairs of regression sample images, wherein each pair of regression sample images includes a face image belonging to the first human face image set and the second human face image set respectively, and the nth pair of regression sample images The first similarity of the face image feature data is linearly distributed; Extracting the second face feature data of the n pairs of regression sample images respectively through the first face model, and calculating the second similarity of the second face feature data of each pair of regression sample images; According to the first similarity and second similarity of the n pairs of regression sample images, the first regression model is determined, wherein the independent variable of the first regression model is the second similarity, and the dependent variable is the second similarity. the first similarity. 8. The device according to any one of claims 5 to 7, wherein the device further comprises: The third acquisition module is used to acquire the monitoring face image through the image acquisition unit; The fourth acquisition module is used to acquire N sets of face feature data of the monitored face image, wherein, the N sets of face feature data of the monitored face image are the monitored person from N types of face models respectively. Extraction from face images; A storage module, configured to store N sets of face feature data of the monitored face images in the face feature database. 9. A face feature retrieval device, characterized in that it comprises a processor, a memory and a computer program stored on the memory and operable on the processor, when the computer program is executed by the processor Realize the steps of the human face feature retrieval method as described in any one of claims 1 to 4. 10. A computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the method according to any one of claims 1 to 4 is realized. The steps of the face feature retrieval method.