WO2021088640A1

WO2021088640A1 - Facial recognition technology based on heuristic gaussian cloud transformation

Info

Publication number: WO2021088640A1
Application number: PCT/CN2020/122249
Authority: WO
Inventors: 袁正午; 查徐鹏; 李林; 梁星
Original assignee: 重庆邮电大学
Priority date: 2019-11-06
Filing date: 2020-10-20
Publication date: 2021-05-14
Also published as: CN111104852B; KR102593835B1; CN111104852A; KR20210100592A

Abstract

The present invention belongs to the technical field of image recognition, and particularly relates to facial recognition technology based on heuristic Gaussian cloud transformation. The method comprises the following content: acquiring a target facial image by using a camera; inputting the target facial image into an MTCNN model, and outputting a square facial frame image obtained after facial-feature-aligned face cutting; constructing a neural network model based on a random_normal activation function, and defining a new facial recognition loss function; pre-training the constructed neural network model by using a preprocessed facial image data set CASIA-WebFace, and maintaining the structure and parameters of the trained model; and inputting the target facial image and facial images in a facial database into the neural network model, and then using a heuristic Gaussian cloud transformation algorithm to obtain a degree of ambiguity in order to determine a facial recognition result. According to the present invention, the facial recognition technology based on heuristic Gaussian cloud transformation is provided, defines the new facial recognition loss function, replaces a softmax classification method, and does not need to take the problems of the number of samples of a recognition object being small and the number of classification categories being great into consideration, thereby improving the precision.

Description

A face recognition technology based on heuristic Gaussian cloud transformation

Technical field

The invention belongs to the field of image recognition technology, and in particular relates to a face recognition technology based on heuristic Gaussian cloud transformation.

Background technique

In modern society, the application of personal identity authentication is more and more extensive. Among them, the areas of application of biometric authentication technology based on fingerprints, iris, and human faces are increasing, such as: mobile phone fingerprint unlocking, access control iris authentication, and station faces Recognition channels and so on. Although they all have high accuracy and reliability, face recognition is the most natural and convenient, and does not require users to make certain actions or gestures, especially for identity authentication in places with a large amount of people. It is the biggest advantage of face recognition. In the future life, face recognition will have greater application prospects.

Face recognition is a process of analyzing and comparing face images in the database based on digital image processing, computer vision and machine learning technologies, with the help of computer processing technology. At present, face recognition technology mainly uses the convolution training operation of deep convolutional neural networks to extract facial features. For two face images of the same person, the corresponding features belong to the same category; conversely, for two different persons The corresponding features of a face image belong to different categories, so in the face recognition model, a person corresponds to one category. In the early neural network model, the Softmax classification was used directly to obtain the probability of each class, and the highest probability or the first few probabilities were selected as the recognition result. However, this technique has low recognition accuracy due to the small training set and the large number of types.

Summary of the invention

In order to solve the above problems, the present invention provides a face recognition technology based on heuristic Gaussian cloud transformation, defines a new loss function for face recognition, instead of the softmax classification method, and no longer needs to consider the sample of the recognition object Problems such as a small number and a large number of classification categories have improved the accuracy.

To achieve the above objective, the technical solution of the present invention is: a face recognition technology based on heuristic Gaussian cloud transformation, including the following steps:

Step 1) Use the camera to obtain the target face image;

Step 2) Input the target face image into the MTCNN neural network model, and output a square face frame image with only facial features aligned with the face cut;

Step 3) Construct a neural network model based on the random_normal activation function, and define a new face recognition loss function;

Step 4) Pre-train the constructed neural network model with the pre-processed face image data set CASIA-WebFace, and retain the structure and parameters of the trained model;

Step 5) Input the target face image and the face image in the face database into the neural network model, and then use the heuristic Gaussian cloud transform algorithm to obtain the ambiguity to judge the face recognition result.

Preferably, in the step 3), the process of constructing a neural network model based on the random_normal activation function and defining a new face recognition loss function includes:

Step 3-1: Build a neural network model and set the activation function of each layer to random_nomal;

Step 3-2, define the loss function loss of the neural network model;

As a preference, the process of defining the loss function loss of the neural network model is as follows:

Input the face image into the neural network model, extract the feature vector of each face image, and the feature dimension of each face image is the same; each training randomly selects 3 face images from the training set , Are the face sample feature anchor, the positive sample of the anchor (belonging to the same person) positive, and the negative sample of the anchor (not belonging to the same person) negative. After they are input into the neural network model, the corresponding feature vector is obtained.

Set the number of concepts of the heuristic Gaussian cloud transformation algorithm to 2, merge the feature vector of the anchor sample and the feature vector of the positive sample, and then use it as the data sample set of the algorithm. After heuristic Gaussian cloud transformation, 2 Gaussian distributions G( μ _k ,σ _k )|k=1, 2; for the k-th Gaussian distribution, calculate the scaling ratio α _{k of} its standard deviation, and then calculate the Gaussian cloud ambiguity CD _k ＝(1-α _k )/(1 +σ _k ), CD ₁ and CD _{2 respectively} ; in the same way, the feature vector of the anchor sample and the feature vector of the negative sample are combined, and then input as the data sample set of the algorithm, and 2 ambiguities are also obtained, respectively CD ₃ , CD ₄ .

For the same person, we hope that their feature vectors belong to the same category, that is, the same concept, then the corresponding ambiguity is large; on the contrary, if they do not belong to the same concept, the corresponding ambiguity is small. due to

We expect

Bigger,

The smaller, therefore, when training the model, as long as you keep optimizing

As a result, it keeps getting close to 0. Therefore, the loss function loss of face recognition is defined as:

Preferably, in step 5), the face image of the database is input into the trained neural network model, and the obtained face feature vector is overlaid on the original face image corresponding to the database, and finally, a face feature vector is obtained. The database of faces is composed.

Preferably, in the step 5), the target face feature vector is merged with the feature vector of the face image in the face database, and then the heuristic Gaussian cloud transform algorithm is used to finally obtain the similarity between the target image and the image in the face database degree.

Beneficial effects. The face recognition technology based on heuristic Gaussian cloud transformation disclosed in the present invention has the following beneficial effects:

1). Propose a face image loss definition based on heuristic Gaussian cloud transformation, which provides a new technology for face recognition technology. It no longer uses the softmax classification method to recognize faces, and solves the problem of fewer training sets and classification categories. Many problems such as low accuracy.

2) The face database no longer uses face images to store personal face information, and converts the feature vector corresponding to the adult face image to store it. This not only protects the privacy of the user, but also reduces the storage space. It is necessary to judge the ambiguity of the feature vectors of the two groups of face images to obtain their similarity. This recognition method shortens the face recognition time.

3) When updating the database, it is no longer necessary to retrain the model, just input the new face image into the trained neural network model, and add the feature vector of the obtained face image to the database.

Description of the drawings

FIG. 1 is a schematic diagram of the evaluation process of the loss function loss of the present invention.

[Corrected according to Rule 26 23.11.2020]

Detailed ways

The technology of the present invention will be further described below in conjunction with the accompanying drawings provided by the present invention:

The present invention discloses a face recognition technology based on heuristic Gaussian cloud transformation. The detailed steps are as follows:

Step 1) Use the camera to obtain the target face image;

Step 3) Construct a neural network model based on the random_normal activation function, and add a new face recognition loss function;

The above steps are described in detail below. Step 1) The target face image can be obtained by using a smart phone or other smart device, and the way to obtain a good face image is to perform a frontal, horizontal, horizontal, and Get close

In the step 2), the target face image is input into the MTCNN neural network model, and the process of outputting a square face frame image with only facial features aligned and cut includes:

Step 2-1. Use the P-Net network to obtain candidate frames and boundary regressors, and at the same time, the candidate frames are calibrated according to the bounding boxes, and then the NMS method is used to remove overlapping frames;

Step 2-2. Train the picture containing the candidate frame determined by the P-Net network in the R-Net network, use the bounding box vector to fine-tune the candidate frame, and then use the NMS method to remove the overlapping frames;

Step 2-3: Use the O-Net network to remove the candidate form and display the location of five key points of the face at the same time.

Through the above steps, the target face image is processed to obtain a square face frame image after facial features aligned face cutting.

The process of constructing a neural network model based on the random_normal activation function in the step 3) and adding a new face recognition loss function includes:

Step 3-1: Build a neural network model and set the activation function of each layer to random_nomal, so that the output features of each layer present a normal distribution state;

Step 3-2, define the loss function loss of a new neural network model, so as to complete the construction of the neural network model;

As shown in Figure 1, the loss function loss process of the neural network model is as follows:

Input the face image to the neural network model, extract the feature vector of each face image, and the feature dimension of each face image is the same; each training randomly selects 3 face images from the training set, They are the face sample feature anchor, the positive sample of the anchor (belonging to the same person) positive, and the negative sample of the anchor (not belonging to the same person) negative. After they are input into the neural network model, the corresponding feature vector is obtained.

We expect

Bigger,

The smaller, therefore, when training the model, as long as you keep optimizing

After defining the new loss function loss, the neural network model is obtained.

In the step 4), pre-training the constructed neural network model with the pre-processed face image data set CASIA-WebFace, and saving the network model structure and parameters when the model no longer converges;

In the step 5), the feature vector of the target face image is merged with the feature vector of the face image in the face database, the number of concepts of the heuristic Gaussian cloud transform algorithm is set to 2, and the merged vector is used as the algorithm Input the data sample set of, get 2 mixing degrees CD ₁ , CD ₂ , and put

The value of is used as the similarity between the target face image and the face image in the face database. In the actual use process (such as attendance system, user verification system), set a threshold. If the similarity value obtained is higher than the threshold, it will be judged as a similar picture, and the verification is deemed to be passed. If it is lower than the threshold, it will be judged as a different picture. , Indicates that the authentication has failed; the threshold here can be set manually or according to the results of training.

The technical content and technical features of the present invention have been disclosed as above. However, those skilled in the art may still make various substitutions and modifications based on the disclosure of the present invention without departing from the spirit of the present invention. Therefore, the scope of protection of the present invention should not be limited to the embodiments. The disclosed content should include various substitutions and modifications that do not deviate from the present invention, and are covered by the claims of this patent application.

Claims

A face recognition method based on heuristic Gaussian cloud transformation is characterized in that: the method includes the following steps:

Step 1) Use the camera to obtain the target face image;

Step 2) Input the target face image into the MTCNN neural network model, and output a square face frame image with only facial features aligned with the face cut;

Step 3) Construct a neural network model based on the random_normal activation function, and define a new face recognition loss function;

Step 4) Pre-train the constructed neural network model with the pre-processed face image data set CASIA-WebFace, and retain the structure and parameters of the trained model;

Step 5) Input the target face image and the face image in the face database into the neural network model, and then use the heuristic Gaussian cloud transform algorithm to obtain the ambiguity to judge the face recognition result.
A face recognition method based on heuristic Gaussian cloud transformation according to claim 1, characterized in that: in the step 3), a neural network model based on a random_normal activation function is constructed, and a new face recognition loss function is defined The process includes:

Step 3-1: Build a neural network model and set the activation function of each layer to random_nomal;

Step 3-2, define the loss function loss of the neural network model.
The face recognition method based on heuristic Gaussian cloud transformation according to claim 2, characterized in that: the process of defining the loss function loss of the neural network model is as follows:

Input the face image into the neural network model, extract the feature vector of each face image, and the feature dimension of each face image is the same; each training randomly selects 3 face images from the training set , Which are the face sample feature anchor, the positive sample of the anchor (belonging to the same person) positive, and the negative sample of the anchor (not belonging to the same person) negative. After they are input into the neural network model, the corresponding feature vector is obtained;

Set the number of concepts of the heuristic Gaussian cloud transformation algorithm to 2, merge the feature vector of the anchor sample and the feature vector of the positive sample, and then use it as the data sample set of the algorithm. After heuristic Gaussian cloud transformation, 2 Gaussian distributions G( μ k ,σ k )|k=1, 2; for the k-th Gaussian distribution, calculate the scaling ratio α k of its standard deviation, and then calculate the Gaussian cloud ambiguity CD k ＝(1-α k )/(1 +σ k ), CD 1 and CD 2 respectively ; in the same way, the feature vector of the anchor sample and the feature vector of the negative sample are combined, and then input as the data sample set of the algorithm, and 2 ambiguities are also obtained, respectively CD 3 , CD 4 .

For the same person, we hope that their feature vectors belong to the same category, that is, the same concept, then the corresponding ambiguity is large; on the contrary, if they do not belong to the same concept, the corresponding ambiguity is small. due to
We expect
Bigger,
The smaller, therefore, when training the model, as long as you keep optimizing
As a result, it keeps getting close to 0. Therefore, the loss function loss of face recognition is defined as:
A face recognition method based on heuristic Gaussian cloud transformation according to claim 1, characterized in that: in step 5), the face image of the database is input into the trained neural network model, and the obtained The face feature vector covers the corresponding original face image in the database, and finally, a face database composed of face feature vectors is obtained.
The face recognition method based on heuristic Gaussian cloud transformation according to claim 1, characterized in that: in step 5), the target face feature vector is merged with the feature vector of the face image in the face database, Then through the heuristic Gaussian cloud transform algorithm, the similarity between the target image and the image in the face database is finally obtained.