CN117437522B

CN117437522B - Face recognition model training method, face recognition method and device

Info

Publication number: CN117437522B
Application number: CN202311745355.9A
Authority: CN
Inventors: 张飞杨; 李星建; 陈当遥; 林关城; 张怀心
Original assignee: Fujian Toulton Software Co ltd
Current assignee: Fujian Toulton Software Co ltd
Priority date: 2023-12-19
Filing date: 2023-12-19
Publication date: 2024-05-03
Anticipated expiration: 2043-12-19
Also published as: CN117437522A

Abstract

The invention provides a face recognition model training method, a face recognition method and a device, wherein the face recognition model training method comprises the following steps: performing face detection and face alignment operation on the face sample image to obtain a preliminarily processed face sample image; For a pair ofAfter preprocessing operation of image super-resolution and tensor channel sequential adjustment, a preprocessed face sample image set is obtained; based onEstablishing a data setData setTraining a neural network model FaceNet to obtain a FaceNet model; Performing feature extraction on the face image set based on N feature extraction networks to obtain different feature vector sets, and establishing N data sets based on the different feature vector sets; and respectively training N identical KNN cluster models by using the data set to obtain N face recognition results, and weighting the output of the KNN to obtain a final prediction result. The invention solves the problem of lower face recognition accuracy in complex environment, and improves the accuracy of model judgment.

Description

Face recognition model training method, face recognition method and device

Technical Field

The invention belongs to the technical field of data processing of face recognition and machine learning, and particularly relates to a face recognition model training method, a face recognition method and a face recognition device based on ensemble learning.

Background

Currently, many social platforms have a large number of users uploading pictures or videos each day, and the content of the users is related to aspects. According to related regulations, a platform side generally sets rules to avoid illegal contents, but the number of picture contents uploaded by users every day is very large, the manual auditing cost is high and the accuracy is difficult to ensure, so an automatic intelligent face auditing system is urgently needed to ensure the content safety.

The main research of face recognition is in aspects of feature extraction and recognition matching of faces. In the field of face recognition, the existing recognition algorithm mostly uses a single feature extraction mode and a single similarity network, and the robustness of the algorithm is insufficient under the condition that the difference is large due to different light rays or unclear photos. The integrated learning is used for face recognition, so that the robustness and recognition rate of a face recognition algorithm can be improved.

In machine learning, because the real image is influenced by environmental factors, the assumption meeting the training set does not necessarily have the same good performance in practical application, a certain risk is faced when a single mode is selected to judge the similarity of the face features, and the risk of misjudgment of the model can be reduced by integrating a plurality of face feature extraction modes and clustering algorithms. The method can effectively detect the specific characters appearing in the images and the videos, and has higher detection efficiency and accuracy.

Meanwhile, the problems of specific figure picture leakage and model safety are considered, a database containing the specific figure picture, a web server providing an auditing platform and a server cluster running an integrated learning model are built in an intranet environment, and a user can upload the audited picture to the auditing server through a webpage by an http protocol, so that data leakage and model running results are prevented from being tampered. Meanwhile, the task of the user is distributed to the idle virtual machines in the server cluster, so that the computing power of the server is fully utilized, and the running efficiency of the whole environment can be improved.

For this reason, an invention patent with publication number CN111898413a discloses a face recognition method, device, electronic equipment and medium. The method comprises the following steps: acquiring a face image to be identified; extracting features of the face image to be identified based on a first network structure to obtain face feature data of the face image to be identified; adjusting the face feature data of the face image to be identified through target dictionary parameters to obtain adjustment feature data; processing the adjustment feature data based on a second network structure to obtain a target face feature vector; by comparing the target face feature vector with the template face feature vector, the recognition result of the face image to be recognized is determined, so that face recognition under different situations can be realized, the accuracy of shielding face recognition can be improved, and the method has strong universality.

Another patent of the invention with publication number CN111291740a discloses a training method of face recognition model, a face recognition method, and the training method includes: and acquiring face characteristic data of at least two modal images corresponding to the sample object. And carrying out feature fusion on the face feature data of the sample object corresponding to at least two modal images to obtain the face fusion feature data corresponding to the sample object. And taking the face fusion characteristic data corresponding to the sample object as the input of the face recognition model, and taking the recognition classification label corresponding to the sample object as the output of the face recognition model to train the face recognition model. The face recognition method comprises the following steps: and acquiring face characteristic data of the object to be identified corresponding to at least two modal images. And carrying out feature fusion on the face feature data of the object to be identified corresponding to at least two modal images to obtain the face fusion feature data of the object to be identified. And inputting the face fusion characteristic data of the object to be identified into the face identification model to obtain an identification result corresponding to the object to be identified.

Another patent of the invention, CN103903004a, discloses a method and apparatus for fusion of multiple feature weights for face recognition, which includes: step 1, acquiring face images under different shielding conditions or different light source conditions, and constructing a training sample set; step 2, training a shielding model or a light source model by using the training sample set; step 3, dividing the samples in the training sample set into a test image and a template image, and using the test image and the template image which accord with specific conditions for constructing a weight training set; step 4, constructing weight functions for the N identification features based on the weight training set to carry out weighted fusion, and determining the optimal value of each weighted parameter; and 5, calculating N template images which are closest to the image to be identified under N identification features according to the image to be identified, and carrying out weighted fusion on the N identification features by using the weight function and the optimal weight when the N template images are the same person.

Another patent of the invention with publication number CN109255340a discloses a face recognition method integrating multiple improved VGG networks, which is based on the existing VGG19 network, by deleting other convolution layers or full connection layers and different combinations thereof, or changing the convolution kernel number and the number of full connection layer nodes of different level convolutions, two or two improved VGG networks with different structures are generated; preprocessing a training sample face image and expanding a data set; putting the expanded data set into an improved VGG network for training, wherein each network correspondingly adopts different training methods to generate a plurality of stable VGG models; and simultaneously putting the face image to be identified into a plurality of models for identification, and selecting a final identification result from a plurality of identification results.

Another patent of the invention with publication number CN103136504a discloses a face recognition method and device, the face recognition method comprises: a clustering feature extraction step, which is used for extracting clustering features of the preprocessed face images; determining, namely determining a clustering feature class which is matched with the face image and is obtained by training in advance according to the clustering feature extracted from the face image; an identification feature extraction step, which is used for extracting P identification features of the preprocessed face image, wherein P is a natural number larger than 1; and a calculation step, which is used for respectively calculating the similarity between the P recognition features and the corresponding features of the P recognition features in the face templates registered in advance, and determining the optimal weight combination of the P recognition features when the P recognition features are subjected to weighted fusion according to the classification of the clustering features determined in the determination step so as to obtain the comprehensive similarity between the face image and the face templates.

As can be seen from the above, in the training method for the face recognition model in the prior art, a deep convolutional neural network structure is generally adopted, and further training and learning are performed by using a learning target of a related loss function based on a large amount of training data, so that network weights capable of characterizing and distinguishing face information of different people are finally learned for face recognition. However, the inventor realizes that at least the following problems exist in the prior art in the practical training process of implementing the face recognition model: 1. the existing training method for the face recognition model uses a single feature extraction method, the feature processing method is improved continuously, and the single feature extraction method is selected to possibly cause insufficient robustness of the trained network. 2. Existing face recognition algorithm methods do not enhance the training and testing samples, resulting in low resolution images that may be difficult to use for recognition.

Disclosure of Invention

The following presents a simplified summary of embodiments of the invention in order to provide a basic understanding of some aspects of the invention. It should be understood that the following summary is not an exhaustive overview of the invention. It is not intended to identify key or critical elements of the invention or to delineate the scope of the invention. Its purpose is to present some concepts in a simplified form as a prelude to the more detailed description that is discussed later.

In order to solve the technical problems, the invention provides a face recognition model training method, a face recognition method and a face recognition device, which can be used for effectively detecting specific characters appearing in images and videos by designing a specific face recognition model, and have higher detection efficiency and accuracy. Meanwhile, a server running the face recognition method and the face recognition device is configured in an intranet environment, so that a user issues tasks through a web server, and data leakage and falsification of model running results are prevented.

According to an aspect of the present application, there is provided a face recognition model training method, including

Step A1: collecting a specific character image required to be identified as a face sample image(I=1, 2,., i) is a natural number;

step A2: for human face sample image Performing face detection and face alignment operation to obtain a primarily processed face sample image set/>(i=1，2，...，i)；

Step A3: for face sample image collection of preliminary processingPerforming an image preprocessing operation, including: image of face sample/>Is converted into a vector/>, comprising the number of RGB channels, height and width of the pictureJudging the vector/> according to a preset valueWhether the channel sequence is RGB, and whether the size and dimension meet the preset resolution; />, not RGB to channel orderProcessing is performed so that the order of channels is changed to RGB, and the resolution is judged to be lowPerforming super-resolution processing by using SRGAN to obtain a preprocessed face sample image set(i=1，2，...，i)；

Step A4: for preprocessed human face sample imageLabeling and establishing a dataset/>; The labels are used for classifying subsequent face images, and the data set is used for training a face feature vector extraction network FaceNet;

step A5: use of data sets Training a neural network model FaceNet to obtain a FaceNet model/>; The neural network model FaceNet is a network for face feature extraction according to the present application.

Step A6: the preprocessed face sample image is subjected to extraction based on N types of feature extraction networks (N is more than or equal to 3)Extracting features to obtain different feature vector sets; for example, the N-based feature extraction networks can be 3-based feature extraction networks to perform feature extraction so as to obtain three face feature vector sets，/>，The 3 feature extraction methods used may employ FaceNet networks/>, respectively68 Face feature point extraction features and PCA algorithm extraction features of dlib.

Step A7: and respectively storing the obtained different feature vector sets into different databases, facilitating comparison between subsequent pictures uploaded by users, and establishing N data sets based on the feature vector sets obtained by different feature extraction methods. For example, the face feature vector is assembled，/>，Respectively stored in 3 independent databases/>,/>,/>For subsequent feature comparison, and simultaneously establishing a dataset/>, by using the face feature vector sets，/>，/>For training an ensemble learning model/>。

Step A8: establishing and training an integrated learning model: respectively training N identical KNN cluster models by using N data sets to obtain N face recognition results, obtaining a prediction result by using a weighted calculation mode, and using(Accuracy) is used as an evaluation index of the integrated model, and an integrated learning discriminant/> isobtained after training is completed. The output of the KNN is weighted to obtain a final prediction result, the weight is obtained by calculating the accuracy of the output of the KNN through a softmax layer, and N identical KNN cluster models and corresponding output weights after training are used as discriminators/>, for integrated learning face recognition. For example, use dataset/>，/>，/>Respectively training three identical KNN cluster models，/>，/>And obtain the face recognition result/>,/>,/>Output weight corresponding to result，/>，/>And a weighted calculation mode is used to obtain the final prediction result/>UsingAs the evaluation index of the integrated model, three identical KNN cluster models/>, after training, are obtained，，/>Corresponding output weights are used as discriminators/>, of integrated learning face recognition。

The application provides an integrated learning face recognition model training method through the scheme, which comprises the following scheme ideas: 1. collecting training samples, wherein the training samples comprise face sample images of the same person in specific characters, with different expressions and different illumination as much as possible; 2. preprocessing a training sample; 3. preprocessing a training sample to obtain a feature vector set of various feature extraction modes; 4. and respectively training the feature vector sets of the multiple feature extraction modes into different machine learning models, and carrying out weighted calculation on the final output through the evaluation index to obtain a final prediction result. The application combines a plurality of feature extraction methods and an integrated learning method, and usually, the integrated learning is to train a plurality of base models by using a feature vector obtained by a feature extraction method, or train the same model by using a plurality of feature extraction methods, thereby integrating the advantages of the above 2 methods and further improving the robustness of the model. Most of the existing face recognition algorithm methods do not enhance the training and testing samples, so that the low-resolution images can be difficult to use for recognition, and in the application, whether the person appears in the corresponding media needs to be checked as much as possible in consideration of the specificity of a specific person in a recognition scene, so that the sample is enhanced by adopting a network. The existing face recognition scheme does not emphasize the configuration and matched operation method, the application comprises the operation of a face recognition system and the configuration process of a server, and the system is enabled to operate efficiently through the configuration of the server as much as possible.

Further, the step A1 specifically includes: collecting specific characters to be identified(J=1, 2,., j) to obtain a face sample image/>(I=1, 2,) i. The collected images are the front face of the same person in the specific person to be detected, and the face sample images with different expressions and different illumination as many as possible are used as training samples, so that the robustness of the model is improved, namely j < < i >.

Further, the step A2 is performed on a face sample imagePerforming face detection and face alignment operations, specifically performing face detection by dlib and face alignment by using a Haar feature classifier, thereby obtaining a preliminarily processed face sample image/>(I=1, 2,., i), in particular, this step A2 comprises:

step A21: sample image of human face The face sample image is segmented by a dlib detector, so as to obtain a face image only containing a face, and the formula of the face image is as follows:

；

Wherein, For the face image only containing faces obtained after segmentation,/>For dividing functions,/>For face sample image/>The i-th face sample image of (a). As one example, the partitioning method is completed by dlib.

Step A22: face alignment is achieved by using OpenCV to perform image rotation transformation, and a primarily processed face sample image which is displayed by taking eyes and mouth of the face as centers is obtainedFace sample image/>The formula of (c) is as follows:

；

Wherein, To achieve face-aligned face sample images,/>For face alignment function,/>Is a face image containing only faces. As an example, the face alignment function/>Is implemented by the Haar feature classifier of OpenCV.

Further, the step A3 specifically includes:

Step A31: face sample image to be preliminarily processed Conversion to vector/>The subsequent processing of the image is convenient;

；

Wherein, The face picture is converted into a tensor vector function by the method; /(I)For vector set/>Contains (C, H, W) information of the image; /(I)And the face sample images are aligned. As an example, the picture turns to function tensor/>Implemented by transform.

Step A32: when the face sample imageThere are pictures with suffix ". Png", which when converted to tensor have 4 channels, there may also be pictures with channels in order "BGR" or "GBR", if/>For tensor of the above channel number not being 3 or tensor of the channel sequence not being "RGB", the channel sequence is converted into "RGB", and the formula is as follows:

；

Wherein, For the processed channel to be a vector of RGB,/>To/>Is converted into a function of 'RGB'For vector set/>The channel order is not tensor of "RGB". As an example,/>Implemented by the image.

Step A33: judgingAnd performing super-resolution operation on the image with low resolution, if the size of the image is larger than a preset value (for example, 500 kb), or one dimension of the dimensions H, W (Height, number of pixels representing the vertical dimension of the image, W: width, number of pixels representing the horizontal dimension of the image) of tensor is larger than a preset value (for example, 1000), the resolution of the image is considered to be sufficient, and no operation is performed. If the above condition is not satisfied, the process proceeds to step A34.

Step A34: the resolution determined in step A33 is lowFace sample image enhancement is performed as input. Using prior art super-resolution model based on generation of countermeasure network GAN, the model can learn mapping relation between low-resolution image and high-resolution image to generate high-quality super-resolution preprocessed human face sample image。

The formula for carrying out face sample image enhancement is as follows:

；

Wherein, A loss function representing an countermeasure target, G and G (z) representing a generator and a generation model concerning noise data z, respectively, and D (x) representing a discriminator and a discrimination model concerning true data x; x is real data, z is noise data,/>Representing the distribution/>, of x-coincidence with real data，/>Representing the z-coincidence noise distribution/>；/>The result representing the noise input generation model G (z) is input into the discrimination model D (x); e () > stands for desired calculation,/>For/>Desired calculation of/>For/>Is used for the calculation of the expected calculation of (a). /(I)The goal is to minimize the loss function of generator G and maximize the loss function of arbiter D; as one example, the GAN model used for migration is a pre-training model SRGAN.

Step A4: for the preprocessed face sample imageLabeling for classifying subsequent face images to establish a dataset/>Dividing the data set for training of the face feature vector extraction network FaceNet;

Further, the step A4 specifically includes the following steps:

Step A41: for preprocessed human face sample image Labeling, namely taking names or classifications of different faces as labels, and establishing a data set/>. As an example, the number of samples is relatively fixed or less, it can be noted by One Hot Encoding;

Step A42: data set Divided into training sets/>Verification set/>Test set/>And (2) and:/>:/>=/>。

Step A5: use of data setsTraining a neural network model FaceNet to obtain a FaceNet model/>，/>Is a network for extracting the face characteristics; the neural network model FaceNet is a network for face feature extraction according to the present application.

Further, the step A5 specifically includes the following steps:

Step A51: the deep CNN layer of the neural network model FaceNet carries out convolution operation on the face sample image, and as the image has three RGB channels, the convolution operation is required to be carried out on the three channels; the formula for face feature extraction by convolution operation on a single channel is as follows:

；

Wherein the spatial coordinates of the single channel input image are (x, y), the convolution kernel size is p×q, and the kernel weight is The luminance value of the image is/>For all/>×/>Summing the results of (2);

Step A52: the extracted face features are subjected to L2 regularization and embedding (embedding method) operation, wherein the L2 regularization formula is as follows:

；

Wherein, （/>) Representing the operation of L2 regularization, X is a face feature with n-dimensional feature values, x= (X1, X2, X3,..once., xn),/>Any face characteristic value is used;

Step A53: calculating a value of Loss output by the neural network model, and carrying out back propagation through a loss_function, namely back () function to adjust parameters of the model, wherein the Loss function of the neural network model is a Triplet Loss, and the formula is as follows:

；

Wherein the method comprises the steps of Representing the result of Triplet Loss,/>（/>/>) Calculation representing euclidean distance, f () (/ >)，/>，/>) Functions fitted to FaceNet models,/>Face features as reference examples,/>For positive example,/>As a negative example, these 3 samples can be considered as a triplet T (/ >),/>,/>) Parameter a represents/>And/>Distance between and/>And/>A threshold value between the distances between them.

As an example, 3-tuple T #,/>,/>) The sample is selected by using an Online TRIPLET MINING method, and feature vectors of each face image are traversed for each batch (comprising b face images) divided when the model is trained, and Euclidean distances between the feature vectors and other images are calculated, wherein the Euclidean distance is expressed as follows:

Wherein the method comprises the steps of ，/>For samples of two face feature vectors in n-dimensional space,/>Representative/>，/>Euclidean distance between them.

Then find the maximum 3-tuple T,/>,/>) Wherein the triplet should conform to the formula:

Wherein, Calculation representing Euclidean distance, f () > is a function of FaceNet model fits,/>Face features as reference examples,/>For positive example,/>As a negative example, these 3 samples can be considered as a triplet T (/ >),/>,/>)

Further, the step A6 is to collect the preprocessed face sample imagesFeature extraction is carried out by using 3 different feature extraction modes, so that a face feature vector set is obtained，/>，The method specifically comprises the following steps:

step A61: based on FaceNet network extraction characteristics, a trained FaceNet model is obtained Input of preprocessed face sample image/>The method is used for extracting the face feature vectors and finally obtaining the corresponding 128-bit face feature vector set；

Step A62: feature extraction is carried out on the 68 face feature points based on dlib, and the face sample image is obtained by loading the detector of the 68 feature points in dlib of pythonConverting the face feature vector into a face feature vector corresponding to 128 bits; the formula for feature extraction is as follows:

；

Wherein, For face feature vectors obtained after feature extraction using dlib,/>Method for extracting face feature vector,/>For/>, in a preprocessed set of face sample imagesAny one image. As an example, the method of extracting the face feature vector is completed by dlib.

Step A63: feature extraction is performed based on a PCA algorithm, and the global face image is reduced from high-dimensional data to low-dimensional data. The model of the PCA algorithm is:

Wherein the method comprises the steps of For original data vector with dimension m×n,/>Is a vector of dimension k x n,/>For/>Is a vector of dimension n x k; the goal of the PCA algorithm is to find one/>Matrix, let pass/>After the matrix projects the original data to a low-dimensional space, the reconstruction error is minimum; multiplying X/>Transpose/>Finally, a face feature vector/>, of m multiplied by k, is obtainedI.e./>。

Further, the step A7 is to collect the obtained face feature vectors，/>，Respectively stored in 3 independent databases/>,/>,/>For subsequent feature comparison, and simultaneously establishing a dataset/>, by using the face feature vector sets，/>，/>For training an ensemble learning model/>. The method specifically comprises the following steps:

Step A71: for face feature vector set ，，/>Carrying out data normalization; the z-score normalization is used for each dimension of the face feature vector, and the formula is as follows:

；

Wherein, And/>Respectively normalizing a value of a certain dimension of the face vector and a value before normalization; /(I)Is the mean value of x,/>For/>Standard deviation of (2);

Step A72: the normalized face feature vector set ，，/>Labeling samples in thereby creating a dataset/>，/>，/>And labeling the names of different faces as labels of the sample face feature vectors. As an example, if the number of samples is fixed or small, it may be noted by One Hot Encoding.

Step A73: the three data sets are combinedDivided into training sets/>Test set/>And is also provided with=/>。

Further, in the step A8, an ensemble learning model is built and trained by using the data set，/>，Training three identical KNN cluster models/>, respectively，/>，/>And obtain the face recognition result,/>,/>Output weight corresponding to result/>，/>，/>And a weighted calculation mode is used to obtain the final prediction result/>Use/>As the evaluation index of the integrated model, three identical KNN cluster models/>, after training, are obtained，/>，/>Corresponding output weights are used as discriminators/>, of integrated learning face recognitionThe method specifically comprises the following steps:

consists of 3 identical KNN cluster models, which are marked as/> ，/>，/>By/>，/>，/>Training three KNN cluster models so as to determine the K value of each model, inputting different feature vectors of the same face, obtaining a corresponding result by each KNN model, and judging that two feature vectors are the result of the same person by a majority voting method. The calculation formula of the euclidean distance used by KNN is as follows:

Wherein y1, y2 represent two face features, and the distance between y1 and y2 is calculated by selecting a method for calculating the Euclidean distance. And traversing and calculating the distance between y1 and each face feature in the training set, then circumscribing the nearest K adjacent training objects, and determining the K value through multiple tests.

Wherein, confirm，/>，/>The method for outputting weights of the 3 sub-models is as follows, the result of each training is counted, and the evaluation index of the model is calculated, wherein the evaluation index of the model is/>Obtaining the accuracy rate/>, of 3 models，/>，/>Inputting 3 accuracy rates of the model into a Softmax layer to obtain the weight output by the model, wherein the formula is as follows:

Wherein the above formula is a Softmax (i.e., a) formula, and the weight of each model for output is calculated through Softmax (i.e., a) ，/>，/>The model output with high accuracy is weighted higher, where exp (i) represents an exponential function that maps the input value to a positive number, where/>The evaluation index is adopted for the model.

When modelAfter the K value and the model weight are determined, a new face image is input, the model traverses the image in the face database to obtain a final result, and the final model is output as follows:

Wherein/> ，/>，/>For the weight of the model to the output,/>For the feature of the face image to be detected,/>For traversing the features of any face image, the 3 feature extraction modes/>, of the step A6, are usedIs a sub-model and the result obtained after inputting the sample into the sub-model, wherein/>, is consideredThe image considered to be of the same person may be modified according to the actual accuracy requirements.

The evaluation index adopted by the integrated learning face recognition model is Accuracy, and the formula is as follows:

For evaluation index,/> (True Positives) is a true case, showing that in fact the sample is a positive case, and the model predicts the sample as a positive case; /(I)(True Negatives) true counterexamples, representing the fact that this sample is counterexample, the model also predicts this sample as counterexample; /(I)(False Positives) is a false positive, indicating that in reality this sample is a negative, but the model predicts a positive case; /(I)(FALSE NEGATIVES) is a false negative, indicating that in reality this sample is positive, but the model predicts the case as negative.

According to another aspect of the present application, there is provided a face recognition method, including:

collecting an image to be identified;

Inputting the image to be recognized into a face recognition model Image recognition is carried out to obtain a recognition result;

The face recognition model is obtained by training the face recognition model by the training method.

The face recognition method specifically comprises the following steps:

Step B1: the user uploads the file to be detected, inquires whether the information of the file exists in the redis cache, if so, directly returns the face recognition result stored in the redis, and refreshes the recorded existence time.

Step B2: judging whether the type of the uploaded file is a picture or not, and respectively processing the picture.

The type of the uploaded file is judged through the suffix name of the file, and the suffix name of the picture type file is as follows: '. xbm ', ' pjp ', ' jpg ', ' jpeg ', ' ico ', ' webp ', ' png ', ' bmp ', ' pjpeg;

step B3: optionally, step a32 is executed, and if the uploaded file is a picture, the number of channels and the channel sequence are judged, and the file is uniformly converted into an RGB 3-channel image.

Step B4: repeating the steps A2-A3 and A6, extracting the facial features of the image uploaded by the user to obtain a facial feature vector set, and returning a result that the image uploaded by the user does not contain the face if the image containing the face does not exist.

Step B5: traversing and comparing the face feature vector set with face feature vectors stored in different databases respectively according to the feature extraction mode, and inputting the face feature vector set into an integrated learning discriminatorAnd obtaining a comparison result, outputting a corresponding result and finishing traversal if the result obtained by the discriminator is a face with a specific person, and returning the person to be detected as the undetermined person if the traversal result does not have the face.

Step B6: the result is stored in the redis, the result can be quickly judged when the same image is uploaded, the content in the redis is emptied after a period of time, and when the redis result is requested each time, the record is refreshed for the existing time, so that the record can be accessed for a longer time.

According to still another aspect of the present application, there is further provided an intranet environment of a face recognition system and a server building method, including the steps of:

step C1: server cluster configured with operation integrated learning model in intranet environment Running multiple virtual machines/>, in parallel on one server(K=1, 2,., k), while the ensemble learning model/>Run in VM Environment, each/>Face recognition/>, distributed by different users(l=1，2，...，l)。

Wherein whenLong-term unassigned/>When the server should set policy to make this/>And entering a shutdown state to reduce the power consumption of the server.

Server clusterThe ping-forbidden policy should be set to prevent icmp message attacks.

Step C2: configuring web servers in an intranet environmentThe method is used for providing an http service to establish a network auditing platform, and an inspector (user) uploads images through a website,/>, and the network auditing platform is used for providing http service to establish a network auditing platformGenerating task releases to/>Differences/>。

Step C3: configuring a firewall on a gateway router while configuring a firewall on a gateway routerAllowing a user to access a web server of an intranet through the internet.

Wherein the gateway routerShould map/>Http service, waiting until SSL certificates are applied to CA, mapping 443 ports for providing https service, wherein the router configures the port mapping command as follows:

ip nat inside source static tcp80/>80 extendable

ip nat inside source static tcp443/>443 extendable

The command is a command of a Cisco router, wherein For/>Intranet ip,/>Is thatThe public network ip.

Step C4: registering and issuing domain names, DNS servers provided by domain name registrarsThe mapping of ip address and domain name is completed. /(I)

Step C5: the user uploads the picture to be audited through the web browser and the web serverWill/>Assigned to/>Idle/>。

If there is no idleAttempting to initiate shutdown/>If there is still no shutdown/>Then C6 is entered.

Step C6: if issued by the userToo much, the user's/>Store in the redis queue until there is free/>Reassigning the/>, when present。

Step C7: when the integrated learning model returns a result, the server returns a corresponding result to the web service, and the web server issues the corresponding result to the user.

Among other things, the servers involved in the identification system may include a central processing unit that may perform various appropriate actions and processes according to programs stored in a read-only memory or programs loaded from a storage portion into a random access memory (RAM, random Access Memory). In the RAM, various programs and data required for the system operation are also stored. The CPU, ROM and RAM are connected to each other by a bus. An input/output (I/O) interface is also connected to the bus.

The server may also relate to a communication part including a keyboard, a mouse, a liquid crystal display, a speaker, and a network interface card such as a LAN (Local Area Network ) card, a modem, and the like.

The application provides a specific person recognition system based on ensemble learning, which comprises a face recognition model training method, a face recognition device, a specific person recognition system execution method, a face recognition system intranet environment and a server building. Compared with the prior art, the application has the following beneficial effects: the method and the system overcome the problem of lower face recognition accuracy in a complex environment by combining an integrated learning method, and simultaneously reduce the influence of noise in a sample on a model result as much as possible by a plurality of sample processing methods, thereby improving the accuracy of model judgment. Compared with the existing face recognition model, the face recognition method based on the face recognition technology uses a technology combining multiple different face feature extraction to train the same kind of different clustering models so as to obtain a face recognition result. Meanwhile, the security problem of specific figure picture leakage is considered, a database containing the specific figure picture, a web server providing an auditing platform and a server cluster running an integrated learning model are built in an intranet environment, and a user can upload the audited picture to the auditing server only through a built website of the web server and cannot directly access the server cluster and the database of the intranet, so that data leakage and falsification of a model running result are prevented. Meanwhile, the task of the user is distributed to the idle virtual machines in the server cluster, so that the computing power of the server is fully utilized, and the running efficiency of the whole environment can be improved.

Drawings

The invention may be better understood by referring to the following description in conjunction with the accompanying drawings in which like or similar reference numerals are used to indicate like or similar elements throughout the several views. The accompanying drawings, which are included to provide a further illustration of the preferred embodiments of the invention and together with a further understanding of the principles and advantages of the invention, are incorporated in and constitute a part of this specification. In the drawings:

Fig. 1 is a schematic diagram of a training step of an integrated learning model for face recognition according to an embodiment of the present invention;

FIG. 2 is a schematic diagram illustrating steps performed by a specific person identification module according to an embodiment of the present invention;

fig. 3 is a schematic diagram of an intranet environment and a server construction of the face recognition system according to the embodiment of the invention;

FIG. 4 is a schematic diagram of a network architecture of FaceNet model used in the present invention;

FIG. 5 is a schematic diagram of an ensemble learning face recognition model according to an embodiment of the present invention;

Fig. 6 is a schematic diagram of an example intranet environment of an embodiment of the invention.

Detailed Description

Embodiments of the present invention will be described below with reference to the accompanying drawings. Elements and features described in one drawing or embodiment of the invention may be combined with elements and features shown in one or more other drawings or embodiments. It should be noted that the illustration and description of components and processes known to those skilled in the art, which are not relevant to the present invention, have been omitted in the drawings and description for the sake of clarity.

Example 1

As shown in fig. 1, an embodiment of the present invention provides a face recognition model training method, including the following steps:

step A1: for specific persons needing to be identified (J=1, 2,., j) collecting its image, resulting in a face sample image/>(i=1，2，...，i)。

The collected images are the front face of the same person in the specific person to be detected, different expressions and as many face sample images as possible under different illumination are used as training samples, so that the robustness of the model is improved, namely j < < i.

Step A2: for human face imageFace detection and face alignment operations are performed. Performing face detection by dlib, and performing face alignment by using a Haar feature classifier to obtain a preliminarily processed face image/>(i=1，2，...，i)。

Step A21: will beAnd (3) dividing the face image by using a dlib detector to obtain an image only containing the face. The formula for obtaining the segmented face image is as follows:

Wherein, For the face image only containing faces obtained after segmentation,/>For segmentation method,/>For face sample image/>The i-th face sample image of (a). As one example, the partitioning method is completed by dlib.

Step A22: face alignment is achieved by using OpenCV to perform rotation transformation of images, and a face image which is displayed by taking eyes and mouth of the face as centers is obtained. The formula for obtaining the face image after transformation is as follows:

Wherein, To realize the preliminary processing of face sample images,/>For the collection/>I-th face image of (i)/>For face alignment method,/>Is a face image containing only faces. As one example, a face alignment method/>Is implemented by the Haar feature classifier of OpenCV.

Step A3: face sample image to be preliminarily processedPerforming image preprocessing operation, wherein the image preprocessing operation is divided into converting the image into a vector/>, which contains the image (C, H, W)/>, Not RGB to channelProcessing while judging the graph/>For the size and dimension of (3) determined to be low resolution >Super-resolution processing is performed by using SRGAN, so that a preprocessed face image/>, is obtained(I=1, 2,) i. The image preprocessing is performed as follows.

Step A31: picture is madeTurning to vector tensor to get/>And the subsequent images are convenient to process.

Wherein,To put the face picture/>, by this methodA method of converting into tensor vectors; /(I)(I=1, 2,., i is the vector set/>Comprises (C, H, W) information of the image, wherein C is the number of RGB channels, H is the height of the image, and W is the width of the image; /(I)And the face sample images are aligned. As an example, the method of picture transition tensor/>Implemented by transform.

Step A32: optionally, the face sample imageThere are pictures with suffix ". Png" and 4 channels, there may be pictures with channels in order "BGR" or "GBR", if/>Tensor, wherein the number of channels is not 3, or the channel sequence is not 'RGB', the channel sequence is converted into 'RGB', and the formula is as follows:

Wherein, For the processed channel to be a vector of RGB,/>To/>Method for converting channel of (E) into 'RGB'For vector set/>The channel order is not tensor of "RGB". As an example,/>Implemented by the image.

Step A33: judgingIf the picture is larger than a preset value (e.g., 500 kb), or if one of the dimensions of the H, W (Height, number of pixels representing the vertical dimension of the image, W: width, number of pixels representing the horizontal dimension of the image) of tensor is larger than a preset value (e.g., 1000), the resolution of the picture is considered to be sufficient, and no operation is performed. If the above condition is not satisfied, the process proceeds to step A34.

Step A34: the resolution determined in step A33 is lowFace sample image enhancement is performed as input: the migration is used based on generating a super-resolution model against the network GAN, the model can learn the mapping relation between the low-resolution image and the high-resolution image, and a high-quality super-resolution face sample image/>, is generated。

The formula of GAN for face enhancement is as follows:

Wherein, Loss functions representing countermeasure targets, G and G () representing generation models (GENERATIVE MODEL), and D () representing discrimination models (DISCRIMINATIVE MODEL). x represents real data, z represents noise data,Representing the distribution/>, of x-coincidence with real data，/>Representing the z-coincidence noise distribution/>。The result of inputting noise into the generation model G () is input into the discrimination model D (). E () stands for performing the desired calculation. /(I)The goal of (a) is to minimize the loss function of the generator G and maximize the loss function of the arbiter D so that real data can be generated by noise data. The GAN model used is a pre-training model SRGAN.

Step A4: for preprocessed human face sample imageLabeling for classifying subsequent face images to establish a dataset/>Dividing the data set for training of the face feature vector extraction network FaceNet;

step A41 for preprocessed face sample image Labeling different face names or classifications as labels, and establishing a data set/>. If the number of samples is relatively fixed or small, the number of samples can be marked by One Hot Encoding.

Step A42: data setDivided into training sets/>Verification set/>Test set/>And (2) and:/>:/>=/>。

Step A5: data setTraining a neural network model FaceNet to obtain a FaceNet model/>; The neural network model FaceNet is a network for face feature extraction according to the present application.

Step A51: the deep CNN layer of FaceNet model needs to convolve the face image, and because the image has three channels of RGB, the convolution operation needs to be performed on all three channels. The formula for face feature extraction by convolution operation on a single channel is as follows:

Step A52: the extracted features are subjected to L2 regularization and embedding operations, where the formula for L2 regularization is as follows.

Wherein,Representing the operation of L2 regularization, X is a face feature with n-dimensional feature values, x= (X1, X2, X3,..once., xn),/>Any face characteristic value is used;

Step A53: calculating a value of Loss output by the model, and carrying out back propagation through a loss_function, namely back () function to adjust parameters of the model, wherein the Loss function of the model is a triple Loss, and the formula is as follows:

Wherein the method comprises the steps of Representing the result of Triplet Loss,/>Calculation representing Euclidean distance, f () > is a function of FaceNet model fits,/>Face features as reference examples,/>For positive example,/>As a negative example, these 3 samples can be considered as a triplet T (/ >),/>,/>) Parameter a represents/>And/>Distance between and/>And/>A threshold value between the distances between them.

Wherein, the 3-element group T is%,/>,/>) For each batch (containing b face images) divided during training of the model, traversing the feature vector of each face image, and calculating the Euclidean distance between the feature vector and other images, wherein the Euclidean distance is expressed as follows:

Wherein, Calculation representing Euclidean distance, f () > is a function of FaceNet model fits,/>Face features as reference examples,/>For positive example,/>As a negative example, these 3 samples can be considered as a triplet T (/ >),/>,/>)/>

Step A6: face image to be preprocessedFeature extraction is performed in 3 different ways, so that a face feature vector set/> isobtained，/>，. The 3 feature extraction modes used are FaceNet network extraction features, dlib 68 face feature point extraction features and PCA algorithm extraction features respectively.

Step A61: based on FaceNet network extraction features, a trained FaceNet model is to be trainedInput pre-processed face image/>And the method is used for extracting the face feature vector and finally obtaining the corresponding 128-bit face feature vector.

Step A62: feature extraction is carried out on the 68 face feature points based on dlib, and the face image is obtained by loading the detector of the 68 feature points in dlib of pythonAnd converting the face feature vector into a face feature vector corresponding to 128 bits. The formula for extracting the face features is as follows:

Wherein, For face feature vectors obtained after feature extraction using dlib,/>Method for extracting face feature vector,/>For/>, in a preprocessed set of face sample imagesAny one image. Specifically, the method for extracting the face feature vector is completed by dlib.

Step A7: the obtained face feature vector，，/>Respectively stored in 3 independent databases/>,/>,/>For subsequent feature comparison, and simultaneously establishing a data set from the face feature vector sets，/>，/>For training an ensemble learning model/>。

Step A71: for face feature vector set，，/>Data normalization was performed. The z-score normalization is used for each dimension of the face feature vector, and its formula is as follows.

Wherein,And/>The face vector is normalized and pre-normalized value of a certain dimension of the face vector. /(I)For the average of all samples x,/>For all samples/>Standard deviation of (2).

Step A72: step A72: the normalized face feature vector set，/>，/>Labeling samples in thereby creating a dataset/>，/>，/>And labeling the names of different faces as labels of the sample face feature vectors. As an example, if the number of samples is fixed or small, it may be noted by One Hot Encoding.

Step A8: using datasets by means of ensemble learning，/>，/>Training three identical KNN cluster models/>, respectively，/>，/>And obtain the face recognition result/>,/>,/>Output weight corresponding to result/>，/>，/>And a weighted calculation mode is used to obtain the final prediction result/>Use/>As the evaluation index of the integrated model, three identical KNN cluster models/>, after training, are obtained，，/>Corresponding output weights are used as discriminators/>, of integrated learning face recognition。

The structure of the ensemble learning model is shown in fig. 5, in which,，/>，/>Are all KNN, through，/>，/>Training three KNN models so as to determine the K value of each model, inputting different feature vectors of the same face, obtaining a corresponding result by each KNN model, and judging that two feature vectors are the result of the same person by a majority voting method. The calculation formula of the euclidean distance used by KNN is as follows:

For evaluation index,/> For a true example, this sample is actually shown as a positive example, and the model predicts this sample as a positive example; /(I)For true counterexamples, this sample is actually counterexample, which the model also predicts as counterexample; /(I)Representing the case where in fact this sample is a counterexample for a false positive example, but the model predicts as a positive example; /(I)The case where this sample is actually a positive example is represented as a false negative example, but the model predicts a negative example.

Due to the specificity of a specific person identification scene, the safety of the face identification system needs to be ensured, the sample of the face identification system needs to be incapable of being leaked, and the result is prevented from being tampered due to the fact that the system is attacked. The invention emphasizes the environment in which the face recognition model is operated, and the face recognition model is configured in an intranet environment, so that a public network user can only upload pictures to be checked through a web server, and attacks from an untrusted public network user to the face recognition system and a database are broken from network segments.

In addition, there are prior art techniques that use multiple networks to process a face feature vector (e.g., CN111898413a uses 2 networks to process a face feature vector), and the end result is to output a face feature vector. The invention adopts a plurality of (network) methods to extract characteristics of the same picture, finally obtains a plurality of different face characteristic vectors, further processes the plurality of different face characteristic vectors, weights the output of KNN to obtain a final prediction result, the weight is obtained by calculating the accuracy of the output of KNN through a softmax layer, and N identical KNN cluster models and corresponding output weights are used as discriminators for integrated learning face recognition after training is completedFinally obtained integrated learning face recognition discriminator/>Has the advantage of good robustness.

Example 2

As shown in fig. 2, the embodiment of the present invention provides a specific person identification system, which includes the following steps:

Example 3

As shown in fig. 6, the embodiment of the invention provides an intranet environment of a face recognition system and server construction, which comprises the following steps:

Step C1: as shown in fig. 6, a server cluster configured to run an ensemble learning model in the intranet environment in fig. 6 Running multiple virtual machines/>, in parallel on one server(K=1, 2,., k), while the learning model is integratedRun in VM Environment, each/>Face recognition/>, distributed by different users(l=1，2，...，l)。

ip nat inside source static tcp80/>80 extendable

ip nat inside source static tcp443/>443 extendable

Step C4: registering and issuing domain names, DNS servers provided by domain name registrarsThe mapping of ip address and domain name is completed.

Those of ordinary skill in the art will appreciate that the various illustrative elements and algorithm steps described in connection with the embodiments disclosed herein may be implemented as electronic hardware, computer software, or combinations of both, where the various illustrative elements and steps are described generally in terms of functionality in order to clearly illustrate the interchangeability of hardware and software. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the solution. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present invention.

The block diagrams shown in the figures of the specific person identification method, system and server based on ensemble learning are only functional entities and do not necessarily correspond to physically independent entities. That is, the functional entities may be implemented in software, or in one or more hardware modules or integrated circuits, or in different networks and/or processor devices and/or microcontroller devices.

The server involved in the method and system for ensemble learning-based specific person identification provided in the present invention is units and algorithm steps of each example described in connection with the embodiments disclosed herein, and can be implemented in electronic hardware, computer software, or a combination of both, and to clearly illustrate the interchangeability of hardware and software, each example's composition and steps have been generally described in terms of functions in the above description. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the solution. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present invention.

It will be appreciated by those skilled in the art that embodiments of the present application may be provided as a method, system, or computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.

The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the application. It will be understood that each flow and/or block of the flowchart illustrations and/or block diagrams, and combinations of flows and/or blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

Furthermore, the methods of the present invention are not limited to being performed in the time sequence described in the specification, but may be performed in other time sequences, in parallel or independently. Therefore, the order of execution of the methods described in the present specification does not limit the technical scope of the present invention.

While the invention has been disclosed in the context of specific embodiments, it should be understood that all embodiments and examples described above are illustrative rather than limiting. Various modifications, improvements, or equivalents of the invention may occur to persons skilled in the art and are within the spirit and scope of the following claims. Such modifications, improvements, or equivalents are intended to be included within the scope of this invention.

Claims

1. A face recognition model training method is characterized in that: comprising the following steps:

Step A1: collecting a specific character image required to be identified as a face sample image I is a natural number;

step A2: for human face sample image Performing face detection and face alignment operation to obtain a preliminarily processed face sample image/>；

Step A3: for the face sample image of preliminary processingImage preprocessing operation is carried out to obtain preprocessed face sample image/>; For the preliminarily processed face sample image/>The image preprocessing operation comprises the following steps: face sample image to be primarily processed/>Is converted into a vector/>, comprising the number of RGB channels, height and width of the pictureJudging the vector/> according to a preset valueWhether the channel order is RGB, and whether the size and dimension meet the preset resolution; />, not RGB to channel orderProcessing is performed so that the order of channels is changed to RGB, and the channel is changed to RGB for the channel judged to be low in resolutionPerforming super-resolution processing by using SRGAN;

Step A4: for preprocessed human face sample image Labeling and establishing a dataset/>；

Step A5: use of data setsTraining a neural network model FaceNet to obtain a FaceNet model/>；

Step A6: the preprocessed face sample image is subjected to network extraction based on N featuresExtracting features to obtain different feature vector sets, wherein N is more than or equal to 3;

Step A7: respectively storing the obtained different feature vector sets into different databases, and establishing N data sets based on the different feature vector sets;

step A8: establishing and training an integrated learning model: respectively training N identical KNN cluster models by using N data sets in the step A7 to obtain N face recognition results, weighting to obtain a prediction result, and obtaining an integrated learning discriminant after training is completed ；

The step A3 specifically includes:

Step A31: face sample image to be preliminarily processed Conversion to vector/>：

；

Wherein,A function for converting the face picture into tensor vectors; /(I)Is a vector setComprises RGB channel number, height and width information of the image; /(I)The face sample images are aligned;

Step A32: if the vector is If the channel number is not 3 or the channel sequence is not 'RGB', converting the channel sequence into 'RGB', and the formula is as follows:

；

Wherein, For the processed channel to be a vector of RGB,/>To/>Is converted into a function of 'RGB'For vector set/>Vectors with middle channel order not "RGB";

Step A33: judging Super-resolution manipulation of low resolution images, i.e., to be determined as low resolution/>, is performed on the size and dimensions of the imagePerforming face sample image enhancement as input;

the step A5 specifically includes:

；

step A52: and carrying out L2 regularization and embedding operation on the extracted face features, wherein the L2 regularization formula is as follows:

；

Wherein, Representing the operation of L2 regularization, X is a face feature with n-dimensional feature values, x= (X1, X2, X3,..once., xn),/>Any face characteristic value is used;

step A53: calculating a value of Loss output by the neural network model, and performing back propagation through a loss_function of backspace () function so as to adjust parameters of the model, wherein a Loss function of the neural network model is a Triplet Loss;

the step A6 specifically comprises the following steps:

step A61: based on FaceNet network extraction characteristics, a trained FaceNet model is obtained Input of preprocessed face sample image/>The method is used for extracting the face feature vectors and finally obtaining the corresponding 128-bit face feature vector set/>；

；

Wherein, For face feature vectors obtained after feature extraction using dlib,/>To extract the function of the face feature vector,/>For/>, in a preprocessed set of face sample imagesAny one image;

Step A63: feature extraction is carried out based on a PCA algorithm, and the global face image is reduced from high-dimensional data to low-dimensional data; the model of the PCA algorithm is:

Wherein the method comprises the steps of For m rows and n columns of original data vectors with dimension m x n,/>For k row n column vectors in the k n dimension,/>For/>Is a vector of dimension n x k; the goal of the PCA algorithm is to find one/>Matrix, let pass/>After the matrix projects the original data to a low-dimensional space, the reconstruction error is minimum; multiplying X/>Transpose/>Finally, a face feature vector/>, of m multiplied by k, is obtainedI.e./>。

2. The face recognition model training method of claim 1, wherein: the step A2 specifically comprises the following steps:

；

Wherein, For the face image only containing faces obtained after segmentation,/>For dividing functions,/>For face sample image/>The ith face sample image in (a);

step A22: face alignment is achieved by using OpenCV to perform image rotation transformation, and a primarily processed face sample image which is displayed by taking eyes and mouth of the face as centers is obtained Face sample image/>The formula of (c) is as follows:

；

Wherein, To achieve face-aligned face sample images,/>For face alignment function,/>Is a face image containing only faces.

3. The face recognition model training method of claim 1, wherein: the step A4 specifically comprises the following steps:

Step A41: for preprocessed human face sample image Labeling and establishing a dataset/>；

Step A42: data setDivided into training sets/>Verification set/>Test set/>And (2) and:/>:/>=7:1:2。

4. The face recognition model training method of claim 1, wherein: the step A7 is to collect the obtained face feature vector set，/>，Respectively stored in 3 independent databases/>,/>,/>For subsequent feature comparison; simultaneously using these sets of face feature vectors to build a dataset/>，/>，/>For training an ensemble learning model/>; The method specifically comprises the following steps:

Step A71: for face feature vector set ，，/>Carrying out data normalization; performing normalization on each dimension of the face feature vector by using z-score;

Step A72: the normalized face feature vector set ，，/>Labeling samples in thereby creating a dataset/>，/>，/>Labeling different face names as labels to the sample face feature vectors;

Step A73: the three data sets are combined (I=1, 2, 3) divided into training sets/>And test set/>And is also provided with:/>=7：3。

5. The face recognition model training method of claim 4, wherein: in the step A8, an integrated learning model is built and trained by using the data set，/>，/>Respectively training three identical KNN cluster models，/>，/>And obtain the face recognition result/>,/>,/>Output weight corresponding to result，/>，/>And a weighted calculation mode is used to obtain the final prediction result/>UsingAs the evaluation index of the ensemble learning model, three identical KNN cluster models/>, after training is completed，/>，/>Corresponding output weights are used as discriminators/>, of integrated learning face recognitionThe method specifically comprises the following steps:

consists of 3 identical KNN cluster models, which are marked as/> ，/>，/>By/>，，/>Training three KNN cluster models so as to determine the K value of each model, inputting different feature vectors of the same face, obtaining a corresponding result by each KNN model, and judging that two feature vectors are the result of the same person by a majority voting method; KNN uses the Euclidean distance calculation method, and the K value is determined through multiple tests;

Wherein, confirm ，/>，/>The method for outputting the weights of the 3 sub-models is as follows, the result of each training is counted, the evaluation index of the integrated learning model is calculated, and the accuracy rate/>' of the 3 models is obtained，/>，Inputting 3 accuracy rates of the model into a Softmax layer to obtain the weight output by the model, wherein the formula is as follows:

wherein exp ()'s represent an exponential function;

When model After the K value and the model weight are determined, a new face image is input, the model traverses the image in the face database to obtain a final result, and the final model is output as follows:

Wherein/> ，/>，/>Weights output for model correspondence,/>For the feature of the face image to be detected,/>、/>、/>Pair/>, respectively, of ways of using 3 feature extractions of step A6Face feature vector obtained by feature extraction,/>For traversing the features of any face image,/>，/>，/>Pair/>, respectively, of modes using 3 feature extraction of step A6Carrying out feature extraction to obtain a face feature vector; /(I)The result obtained after the sample is input into the submodel, namely the result obtained after the sample is input into 3 KNN models; when/>When it is regarded as/>And/>Is an image of the same person, and modifies the threshold according to the actual precision requirement;

For evaluation index,/> For the real example,/>For true and inverse example,/>For false positive,/>Is a false counter example.

6. A face recognition method, comprising:

collecting an image to be identified;

Inputting the image to be identified into a face recognition model for image identification to obtain an identification result;

Wherein the face recognition model is trained by the face recognition model training method according to any one of claims 1 to 5.

7. A face recognition device for implementing the face recognition model training method according to any one of claims 1 to 5, characterized in that: comprising the following steps:

and the acquisition module is used for: for collecting a specific person image to be identified as a face sample image (i=1，2，...，i)；

And a preliminary processing module: for human face sample imagePerforming face detection and face alignment operation to obtain a preliminarily processed face sample image/>；

And a pretreatment module: for human face sample imageImage preprocessing operation is carried out to obtain preprocessed face sample images；

A data set establishing module: for the preprocessed face sample imageLabeling, and establishing a data set/>, through division；

The neural network model training module: data setTraining a neural network model FaceNet to obtain a FaceNet model；

And the feature extraction module is used for: face sample image based on N feature extraction networksExtracting features to obtain different feature vector sets, wherein N is more than or equal to 3;

A data set establishing module: respectively storing the obtained different feature vector sets into different databases, and establishing N data sets based on the different feature vector sets;

The integrated learning discriminant building module: establishing and training an integrated learning model: respectively training N identical KNN cluster models by using N data sets in the step A7 to obtain N face recognition results, obtaining a prediction result by using a majority voting mode, and obtaining an integrated learning discriminator after training is completed 。