CN110992522A

CN110992522A - Indoor and outdoor universal human face recognition new algorithm security system

Info

Publication number: CN110992522A
Application number: CN201910804903.8A
Authority: CN
Inventors: 陈工; 宋经纬; 郭二帅
Original assignee: 120 Technology Beijing Co Ltd
Current assignee: 120 Technology Beijing Co Ltd
Priority date: 2019-08-29
Filing date: 2019-08-29
Publication date: 2020-04-10

Abstract

The invention designs an indoor and outdoor universal face recognition new algorithm security system, which comprises: the system comprises a human body data checking/inputting module, a training artificial intelligence model module, an artificial intelligence model verifying and storing module, an artificial intelligence model testing module, a model calculating module, a communication module and a central system authority issuing module; the face photos are deeply learned and trained, the trained models are tested, evaluated and stored, and the central system sets authority and issues the authority to the intelligent door lock terminal.

Description

Indoor and outdoor universal human face recognition new algorithm security system

The technical field is as follows: the invention relates to the field of community security automation, in particular to an indoor and outdoor universal human face recognition new algorithm security system designed by an artificial intelligence technology.

Background art: in recent years, with the advancement of urbanization and the improvement of living standard of residents, more and more residents live into communities, the communities have larger and larger scales, and the community security becomes one of the key problems concerned by public and government security departments. The existing community security is still mainly based on human security, the automation degree is low, the day and night duty is needed, the human consumption is high, and for community residents, the community residents often forget to take keys and the keys are moved by other people to enter the community. The degree of computer automation is continuously improved, and particularly, the coming of an artificial intelligence technology provides a new idea for solving problems.

The invention content is as follows:

the invention aims to provide an indoor and outdoor universal human face recognition new algorithm security system which is more intelligent and has a comprehensive identity recognition database compared with the traditional community security system.

In order to achieve the above purpose, the invention adopts the following technical scheme:

the overall structure design comprises the following steps of inputting a third-party face database, applying the third-party face database to deep learning or transfer learning, and carrying out precision evaluation in the process. And migrating the model trained from the migration learning database, evaluating the model by using the test data, and if the test is passed, storing the model, wherein the model is used for identity verification of face recognition. The management end can issue the model to manage the access authority of each door lock.

Personal identification data check/entry. Before data input (including migration learning data and test data), whether folders of the data are classified according to three folders of a trace identification test or not is checked, whether data names in the three folders are according to a training rule or not is checked, and data proportions of the three folders and a file format are checked.

And training an artificial intelligence model. The model is trained using data in a standard database and can then be saved and migrated (i.e., the ability of the model to learn is applied to another model). And applying the data checked in the previous step to perform new model training (the application is a training data set train). Training artificial intelligence model data was derived from two large-scale face libraries (FERET and FRGC v 2.0).

And verifying and training the artificial intelligence model. And (3) verifying the trained model by using a verification data set (the validation data set in the step 3.1), wherein the verification accuracy is up to the standard, the standard is set by the user, and is generally about 88%. If the requirements are not met, the model training is carried out again; if so, the storage model can be selected. Verification the artificial intelligence model data was derived from randomly sampled portions of two large-scale face libraries (FERET and FRGC v 2.0).

And (6) evaluating the accuracy of the model. Precision assessment exists throughout the process, including the assessment of training data sets, validation data sets, and test data sets, where precision, recall, and a precision-recall curve are calculated to assess model precision

Model test/compute data functions. The testing process comprises the steps that part of target data is used as a testing data set, precision testing and evaluation are conducted on the trained model, if the target data does not reach a set threshold value, the model is judged not to meet requirements, and the model is trained again; and if the test precision meets the requirement and meets the precision requirement, applying the test precision to the target data. When the model testing precision meets the requirement, the target data is calculated and classified, the final calculation result is reserved and stored, and meanwhile, the calculation and classification precision report can be reserved.

And issuing the authority of the central system. The central system authority issuing function comprises the step of transmitting the trained model to access control terminals through a network, wherein the terminals comprise an access control system of a community and access control systems installed by each user of the community. The authority issuing function also includes transmitting personal information to the access control terminal to control whether or not a specific individual is allowed to unlock. The system authority also comprises the step of controlling the terminal to transmit part of the personal identity information stored by the terminal to the overall system. The system may issue a command to upgrade the terminal firmware.

Description of the drawings:

FIG. 1 flow chart of personal identification data checking/inputting function

FIG. 2 functional flow diagram for training artificial intelligence models

FIG. 3 is a functional flow diagram of a verification training artificial intelligence model

FIG. 4 flow chart of model accuracy assessment

FIG. 5 model test/calculate data function flow diagram

FIG. 6 flow chart of the central system rights issuing function

The specific implementation mode is as follows:

personal identification data check/entry. Before data input (including migration learning data and test data), whether folders of the data are classified according to three folders of a train evaluation test or not is checked, whether data names in the three folders are named according to a training rule or not is checked, and the data proportions of the three folders are as follows, for example, 3: 1: 1, and file format.

And training an artificial intelligence model. The model is trained using data in a standard database and can then be saved and migrated (i.e., the ability of the model to learn is applied to another model). And applying the data checked in the previous step to perform new model training (the application is a training data set train).

Training artificial intelligence model data was derived from two large-scale face libraries (FERET and FRGC v 2.0).

And verifying and training the artificial intelligence model. And (3) verifying the trained model by using a verification data set (the validation data set in the step 3.1), wherein the verification accuracy is up to the standard, the standard is set by the user, and is generally about 88%. If the requirements are not met, the model training is carried out again; if so, the storage model can be selected.

In the model test/calculation data function, global and local facial features are integrated in a serial and parallel combination mode, firstly, global features are used for rough matching, and then the global and local features are integrated for fine confirmation.

In the aspect of spectrum analysis, the global features correspond to low frequencies, and the local features correspond to high frequencies.

And extracting global Fourier features. The low-frequency part of a two-dimensional Discrete Fourier Transform (DFT) coefficient is used as a global feature, and the two-dimensional discrete Fourier transform of an image can be represented by the following formula:

where f (x, y) represents a two-dimensional image of size M N, u and v are frequency domain variables image f (x, y) is a real-valued function, so that the output of the Fourier transform is complex, i.e., the Fourier transform is complex

The method comprises the following steps of obtaining a face image, wherein R (u, v) and I (u, v) respectively represent a real part and an imaginary part of F (u, v), after Fourier transformation, the image can be represented as a real part and an imaginary part transformation coefficient of all frequency bands, and although the transformation coefficients of all the frequency bands contain information in the image, the global information of the image is mostly contained in low-frequency coefficients.

And extracting local Gabor characteristics. Two-dimensional Gabor wavelet transform (2D Gabor wavelet transform, GWT for short) was originally proposed by Daugman [11] for modeling the spatial receptive field of primary visual cortical simple cells, and has recently been considered one of the most successful face description methods.

Wherein k is_uvKv denotes the frequency (scale) of the kernel function, ϕ_uBy setting different scales and directions, a group of Gabor wavelet kernel functions can be obtained.

And constructing global and local classifiers and performing serial and parallel integration of the global and local classifiers. The global characteristics mainly describe the overall attributes of the human face and are used for rough matching; while local features mainly describe changes in details of the face for fine confirmation, therefore, to improve the accuracy and speed of recognition, we propose to construct a two-layered classifier: the global classifier is used to perform coarse matching at level 1, and the global and local classifiers are integrated (i.e., global classifier) at level 2 to perform fine validation, as can be seen from the classifier construction process in the previous section, the global classifier uses fewer features and is therefore faster, but the precision is lower; the whole classifier is composed of N +1 component classifiers, utilizes more features, has lower speed, therefore, the speed of recognition can be improved by adopting the global classifier in the layer 1, and the accuracy of recognition can be improved by adding the local classifier in the layer 2.

After global and local feature extraction, we can get N +1 feature vectors, 1 global feature vector and N local feature vectors, because of the higher dimension of these feature vectors, we propose to use Linear Discriminant Analysis (LDA) to further reduce the dimension of these N +1 feature vectors, for two face images, by comparing the corresponding feature vectors, we can get N +1 similarities, in this document, we use the commonly used normalized cross-correlation (NCC) method to calculate the similarity of the corresponding feature vectors, in the task of pattern recognition, once the similarity between samples (or feature vectors) is calculated, the task of classifier becomes very simple (for example, classification can be done with nearest neighbor classifier), therefore, for convenience of description, global and local feature vectors after LDA dimension reduction are respectively called a Global Classifier (GC) and a Local Component Classifier (LCC), because the facial features used by the classifiers E are different, the classifiers have large difference (diversity), and the classification error rate can be effectively reduced by integrating the classifiers in a certain form according to the ensemble learning (ensemble) theory.

In this document, we perform classifier integration at the level of similarity (score level), that is, perform weighted summation on the similarities output by multiple classifiers to obtain the final similarity.

First, we perform weighted summation on N local component classifiers to obtain a Local Classifier (LC):

then, the global classifier and the local classifier are also integrated in parallel in a weighted summation mode to obtain an overall classifier (UC):

where wG represents the weights of the global classifier.

In view of the different characteristics of the global and local features, the resolution of the face image is different when extracting the two features (as shown in fig. 4). the global features mainly reflect the overall properties of the face, so a lower resolution face image can meet the requirements.

In the face recognition problem, a system inputs identity information declared by a user while inputting a face image, and the system judges whether the identity of the input face image is consistent with the declared identity, so that the process is one-to-one.

For the face recognition problem, we first use the global classifier to find the similarity between the input face image and all the candidate face images in the database in the 1 st layer, and sort the similarity to eliminate the candidate with larger difference with the input face image, then use the global classifier to further recognize the rest of the candidates similar to the input image in the 2 nd layer, thus, the slower global classifier only needs to process a smaller subset of the original face database, so the recognition speed can be improved significantly, in this process, there is a problem to be noted that we need to ensure that the candidate face image with the same identity as the input face image is not basically eliminated by the global classifier in order to maintain the accuracy of the whole recognition system, this needs to adjust the number of the candidates (assumed as M) kept in the 1 st layer classification, obviously, the larger the value of M, the more candidates that are retained, the less likely candidates that are identical to the input face image will be excluded, however, as the number of retained candidates increases, the speed of the level 2 classification will decrease accordingly.

For the human face confirmation problem, firstly, a global classifier is utilized in a layer 1 to obtain the similarity between an input human face image and a human face image in a database which is consistent with the declared identity of a user, the similarity is rough, but certain judgment information is also included, if the similarity is lower than a certain small threshold (T1), the two human face images are very likely to belong to the same person, so that the system can directly judge that the identities of the two human face images are different, otherwise, if the similarity is higher than a certain large threshold (T2), the two human face images are very likely to belong to the same person, so that the system can directly judge that the identities of the two human face images are the same, except for the two situations, namely, when the similarity given by the global classifier is between T1 and T2, a correct decision is difficult to be made according to the similarity, based on the above strategy we can see that the system speed is significantly increased if in most cases the system can make decisions based on the similarity given by the global classifier, rather than having to use a slower global classifier, but the system accuracy is reduced if it always relies on the global classifier, so we also need to adjust the thresholds T1 and T2 to achieve a balance between speed and accuracy.

In Fourier feature extraction, the size of the face image is 64 × 80 and the distance between the two eyes is 28. to use Fast Fourier Transform (FFT), the image is expanded to 128 × 128. therefore, considering symmetry, the width of the Fourier spectrum is 64. global information is mostly contained in the Fourier transform coefficients of the low frequency band.

Claims

1. Indoor outer general face identification new algorithm security protection system, its characterized in that includes: the system comprises a personal identification data checking/inputting module, a training artificial intelligence model module, an artificial intelligence model verifying and storing module, an artificial intelligence model testing module, a model calculating module, a communication module and a central system authority issuing module.

2. The system of claim 1, wherein the pin id data check/input module checks whether folders of data are classified by three folders of trail validationtest, whether data naming in three folders is according to training rules, three folder data proportions, and file format before data input (including migratory learning data and test data).

3. The system of claim 1, wherein the training artificial intelligence model module checks whether the folders of the data are classified according to three folders of a trace validation test, checks whether the data names in the three folders are according to a training rule, three folder data proportions, and a file format before inputting the data (including migrating learning data and testing data);

the verification and storage artificial intelligence model module verifies the trained model by using a verification data set, if the verification precision meets the standard, the standard is set by the user, and if the verification precision does not meet the requirement, the model training is carried out again; if so, the storage model can be selected.

4. The system of claim 1, wherein the artificial intelligence model test module is configured to perform an accuracy assessment that includes a training dataset, a validation dataset, and a test dataset, wherein the accuracy assessment is configured to compute accuracy, recall, and plot accuracy-recall curves to assess model accuracy.

5. The system of claim 1, wherein the model calculation module calculates and classifies the target data when the model test accuracy meets the requirement, and the final calculation result is retained and stored while the calculation and classification accuracy report can be retained.

6. The system of claim 1, wherein the central system authority issuing module is configured to transmit the trained model to the access terminals via a network, and the terminals include a community access system and access systems installed by users in the community.

7. The authority issuing function also includes transmitting personal information to the access control terminal to control whether or not a specific individual is allowed to unlock.

8. The system authority also comprises the step of controlling the terminal to transmit part of the personal identity information stored by the terminal to the overall system.

9. The system may issue a command to upgrade the terminal firmware.