CN115512428A - A method, system, device, and storage medium for live face discrimination - Google Patents

Abstract

The present invention discloses a method, system, device and storage medium for human face identification, wherein the method includes: acquiring a target image to be identified, and acquiring a face detection frame of the target image; Acquire the first human face image and the second human face image in the target image; the first human face image is input into the first human face living body discrimination model to obtain the first feature and the first output result; the second human face image is input The face image is input into the second human face living body discrimination model to obtain the second feature and the second output result; after the first feature and the second feature are spliced, they are input into the fusion weight prediction sub-model to obtain the two human face living body discrimination models. Fusion weight: Acquire the final output result of human face liveness discrimination according to the fusion weight, the first output result and the second output result. The invention can effectively improve the living body discrimination performance under the condition of limited computing resources and storage space, and can be widely used in the field of image processing.

Description

A method, system, device, and storage medium for live face discrimination

technical field

The present invention relates to the field of image processing, in particular to a method, system, device and storage medium for distinguishing living human faces.

Background technique

At present, face recognition technology is widely used in smart security, smart home, financial payment and other fields because of its convenience. However, due to the easy access to face information, lawbreakers can reproduce authorized face images through physical means such as paper printing, electronic screen reproduction, or even 3D printing, thereby easily bypassing the face recognition system. This relatively low-cost deception The method brings serious security risks to the face recognition system. In order to identify the authenticity of facial images and ensure system security, face liveness discrimination technology has become the focus of the industry.

Face recognition system applications are mostly oriented to community access control, smart door locks and other side-end embedded hardware platforms. The computing chip performance, power consumption and storage space are limited, and there are strict restrictions on the number of parameters and computational complexity of the algorithm model. Deploy complex living body discrimination algorithm models. Secondly, on the image acquisition device side, although the introduction of a depth camera based on structured light or laser speckle can effectively help resist planar attacks such as printing and electronic display, but considering the high device cost of the depth camera, currently based on monocular infrared camera The image acquisition scheme is still widely used. How to fully mine the local information of the face, the background area information and the relative relationship information between the face and the context background based on a single frame image with limited computing resources is of great importance for the task of stand-in monocular liveness discrimination. significance.

Contents of the invention

In order to solve one of the technical problems in the prior art at least to a certain extent, the object of the present invention is to provide a method, system, device and storage medium for identifying a living human face.

The technical scheme adopted in the present invention is:

A method for discriminating live human faces, comprising the following steps:

Obtain the target image to be discriminated, and obtain the face detection frame of the target image;

Acquire a first human face image and a second human face image from the target image according to the human face detection frame; wherein, the first human face image is a human face image that does not contain a background, and the second human face image is a human face image that includes a background face image;

Inputting the first human face image into the first human face living body discrimination model to obtain the first feature and the first output result;

Inputting the second human face image into a second human face living body discrimination model to obtain a second feature and a second output result;

After splicing the first feature and the second feature, input the fusion weight prediction sub-model to obtain the fusion weight of two human face living body discrimination models;

The final output result of human face liveness discrimination is obtained according to the fusion weight, the first output result and the second output result.

Further, the acquiring the first face image and the second face image from the target image according to the face detection frame includes:

Cutting out a first human face image from the target image according to the human face detection frame;

Enlarging the face detection frame to obtain a second detection frame;

Cutting out a second face image from the target image according to the second detection frame.

Further, the basic framework of the first human face liveness discrimination model and the second human face liveness discrimination model are the same, and both are lightweight network models;

The live face discrimination model is composed of multiple Invert residual blocks of MoblieNetV2.

Further, the first live face discrimination model and the second live face discrimination model are trained in the following manner:

Insert point convolution layer, batch normalization layer, linear rectification layer at the output end of the intermediate layer of the first human face liveness discriminant model and the second described second human face liveness discriminant model, to build pixel-level classifier;

Obtaining the face image and the face image containing the background according to the training set, and generating the pixel-level classification labels of the face image and the face image containing the background respectively according to the pixel-level classifier;

Inserting a pixel-level classification loss into the first human face living body discrimination model and the second human face living body discrimination model, and forming a target loss function together with Focal Loss;

Optimizing the target loss function according to the binary classification label and the pixel-level classification label, and training the first human face liveness discriminant model and the second human face liveness discriminant model according to the target loss function.

Further, the generation process of the pixel-level classification label includes:

For the pixel-level classification label of the face image: construct a label map with the same size as the output of the pixel-level classifier, when the face image is a deceptive face category, the values of the classification label map are all set to 0; when the When the face image is a real face, the values of the classification label map are all set to 1;

For the pixel-level classification label of the face image containing the background: construct a label map with the same size as the pixel-level classifier output, when the input image is a spoofed face, the corresponding value in the spoofed face area in the label map is 2 , the value of the background part is 0; when the input image is a real face, the value of the face area in the label image is 1, and the value of the background part is 0.

Further, the training of the first human face living body discrimination model and the second human face living body discrimination model according to the target loss function includes:

Training the first human face living body discrimination model and the second human face living body discrimination model separately;

，同时计算模型末端输出和二分类标签的Focal Loss损失，记为

；目标损失函数定义为：During training, the cross-entropy classification loss of the output of the pixel-level classifier and the corresponding position label in the label map is calculated point by point, and the loss is denoted as

, and calculate the Focal Loss loss of the end output of the model and the two-category label at the same time, denoted as

; The objective loss function is defined as:

为指定像素级分类损失函数的权重。In the formula,

For the weights of the specified pixel-level classification loss function.

Further, the fusion weight prediction sub-model forms a backbone structure by stacking convolutional layers, batch normalization layers, and linear rectification layers, and introduces a channel attention mechanism to build the basic structure of the fusion weight prediction sub-model;

The fusion weight prediction sub-model is trained in the following manner:

Obtaining the first feature map output by the first human face liveness discrimination model and the second feature map output by the second human face liveness discrimination model;

After splicing the first feature map and the second feature map, input the fusion weight prediction sub-model, and output the weighted fusion weight of the terminal output of the first human face liveness discrimination model and the second human face liveness discrimination model terminal output;

Utilize described weighted fusion weight to the terminal output of the first human face living body discriminant model and the terminal output weighted summation of the second human face live body discriminant model, obtain final output;

Calculate the Focal Loss of the final output and the two classification labels, and optimize the loss using the stochastic gradient descent method to train the fusion weight prediction sub-model;

Wherein, the weight parameters of the first human-face liveness discrimination model and the second human-face liveness discrimination model are frozen during training.

Another technical scheme adopted in the present invention is:

A human face living body discrimination system, comprising:

The target image acquisition and face detection module is used to acquire the target image to be discriminated, and obtain the face detection frame of the target image;

A face image cropping module, configured to obtain a first face image and a second face image from the target image according to the face detection frame; wherein, the first face image is a face image that does not contain a background, The second human face image is a human face image including a background;

The live body discrimination prediction module is used to input the first human face image into the first human face live body discrimination model to obtain the first feature and the first output; the second human face image is input to the second human face live body discrimination model, obtaining a second feature and a second output;

A predictive fusion module, configured to splicing the first feature and the second feature into the fusion weight prediction sub-model to obtain the fusion weight of the two living face discrimination models;

A judging module, configured to obtain a final output of human face liveness discrimination according to the fusion weight, the first output and the second output.

Another technical scheme adopted in the present invention is:

A human face living body discrimination device, comprising:

at least one processor;

at least one memory for storing at least one program;

When the at least one program is executed by the at least one processor, the at least one processor implements the above method.

Another technical scheme adopted in the present invention is:

A computer-readable storage medium stores a processor-executable program therein, and the processor-executable program is used to perform the above method when executed by a processor.

The beneficial effects of the present invention are: the present invention constructs a light-weight living body discrimination model, which can be applied to edge devices; based on the basic model, the first living body discrimination model and the second living body discrimination model are constructed, and input images and pixels are respectively designed Level classification information, so that the two focus on the differentiated living body discrimination clues, and improve the fusion performance of the two; through the fusion weight sub-module, the fusion weight predicted by the first living body discrimination model and the second living body discrimination model can be dynamically determined, which can make full use of The context information of the face area and the image background area is obtained, and the performance of living body discrimination is effectively improved under the condition of limited computing resources and storage space.

Description of drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following describes the accompanying drawings of the embodiments of the present invention or the related technical solutions in the prior art. It should be understood that the accompanying drawings in the following introduction are only In order to clearly describe some embodiments of the technical solutions of the present invention, those skilled in the art can also obtain other drawings based on these drawings without creative work.

Fig. 1 is a flow chart of the steps of a method for discriminating living human faces in an embodiment of the present invention;

Fig. 2 is a schematic diagram of the basic framework of the human face living body discrimination model in the embodiment of the present invention;

Fig. 3 is a schematic diagram of a human face living body discrimination method for multi-model prediction fusion in an embodiment of the present invention;

4 is a schematic diagram of a fusion weight prediction sub-model in an embodiment of the present invention;

Fig. 5 is a schematic structural diagram of a human face living body discrimination system in an embodiment of the present invention;

Fig. 6 is a schematic structural diagram of a human face living body discrimination device in an embodiment of the present invention.

detailed description

Embodiments of the present invention are described in detail below, examples of which are shown in the drawings, wherein the same or similar reference numerals designate the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by referring to the figures are exemplary only for explaining the present invention and should not be construed as limiting the present invention. For the step numbers in the following embodiments, it is only set for the convenience of illustration and description, and the order between the steps is not limited in any way. The execution order of each step in the embodiments can be adapted according to the understanding of those skilled in the art sexual adjustment.

In the description of the present invention, it should be understood that the orientation descriptions, such as up, down, front, back, left, right, etc. indicated orientations or positional relationships are based on the orientations or positional relationships shown in the drawings, and are only In order to facilitate the description of the present invention and simplify the description, it does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus should not be construed as limiting the present invention.

In the description of the present invention, several means one or more, and multiple means more than two. Greater than, less than, exceeding, etc. are understood as not including the original number, and above, below, within, etc. are understood as including the original number. If the description of the first and second is only for the purpose of distinguishing the technical features, it cannot be understood as indicating or implying the relative importance or implicitly indicating the number of the indicated technical features or implicitly indicating the order of the indicated technical features relation.

In the description of the present invention, unless otherwise clearly defined, words such as setting, installation, and connection should be understood in a broad sense, and those skilled in the art can reasonably determine the specific meanings of the above words in the present invention in combination with the specific content of the technical solution.

Explanation of terms:

MoblieNetV2: The second-generation lightweight convolutional neural network infrastructure designed by the Google team.

Invert residual block: Invert residual block refers to a component that constitutes a convolutional neural network, which usually consists of multiple sequentially stacked point convolution layers, depth separable convolution, point convolution layers, and shortcut structures; input features The channel dimension of is expanded and then compressed by a point convolution layer.

Focal Loss: A two-category loss function whose specific expression is as follows:

为模型的输出，该损失函数能动态调节容易样本和难样本对于总体损失的影响。in

As the output of the model, the loss function can dynamically adjust the influence of easy samples and hard samples on the overall loss.

Two classification labels: two classification labels are used to train the first living body discrimination model and the second living body discrimination model, and the value is 1 when the input image is a real face, and the value is 0 when the face is deceived.

As shown in FIG. 1 , this embodiment provides a multi-model predictive fusion-based method for discrimination and prediction of liveness in human faces, so as to realize the performance of discrimination of livingness in monocular under the condition of limited computing power and storage space. The method specifically includes the following steps:

S1. Acquire a target image to be identified, and acquire a face detection frame of the target image.

As an optional implementation manner, the target image can be collected in real time by an image acquisition device, or directly read image file data in a storage device. The face detection frame of the target image can be obtained by inputting the target image to be recognized into any existing face detection algorithm model, such as a deep network model. Wherein, the face detection frame may be a rectangular frame, or a circular frame or a detection frame of other shapes.

S2. Obtain a first human face image and a second human face image from the target image according to the human face detection frame; wherein, the first human face image is a human face image that does not contain a background, and the second human face image is An image of a face that includes a background.

As an optional implementation manner, the face image refers to an image cut out from the target image based on the original face detection frame, and the face image including the background refers to appropriately amplifying the size of the face detection frame, and then An image cropped from the target image.

，坐标中

表示人脸所在区域的左上角坐标，坐标中

表示人脸所在区域的右下角坐标，由左上角和右下角即可确定一个人脸框，抠取目标图像

矩形区域内的图像作为人脸图像；以

为中心，抠取目标图像

为左上角、

矩形区域内的图像内容包含作为包含背景的人脸图像。Specifically, given face coordinates

, in coordinates

Indicates the coordinates of the upper left corner of the area where the face is located, in the coordinates

Indicates the coordinates of the lower right corner of the area where the face is located. A face frame can be determined from the upper left corner and the lower right corner, and the target image is selected.

The image in the rectangular area is used as the face image;

Take the target image as the center

for the upper left corner,

The image content within the rectangular area contains the face image as the containing background.

S3. Input the first human face image into the first human face living body discrimination model to obtain the first feature and the first output result; input the second human face image into the second human face living body discrimination model to obtain the second feature and the second output result.

S4. Splicing the first feature and the second feature into the fusion weight prediction sub-model to obtain the fusion weight of the two living face discrimination models.

S5. According to the fusion weight, the first output result and the second output result, obtain the final output result of human face liveness discrimination.

The multi-scale model prediction fusion living body discrimination method proposed in this embodiment fully utilizes the context information of the face area and the image background area, and can effectively improve the living body discrimination performance under the condition of limited computing resources and storage space.

Referring to Fig. 2, as an optional implementation manner, the basic framework of the first human face living body discrimination model and the second human face living body discrimination model are the same, and both are lightweight network models; The discriminant model is formed by stacking multiple Invert residual blocks of MoblieNetV2.

In this embodiment, a human face liveness discrimination model includes 4 Invert residual blocks, and the parameters of the two overall models are less than 760KB, which can be effectively deployed on edge devices with low computing power, low power consumption and low memory. application.

The construction method and training method of the human face living body discrimination model will be explained in detail below in conjunction with FIG. 2 .

Step 1. Build the basic framework of the first live face discrimination model and the second live face discrimination model by stacking several Invert residual blocks of MoblieNetV2

Step 2. Insert a point convolution layer, a batch normalization layer, and a linear rectification layer at the output ends of the intermediate layers of the first human face liveness discrimination model and the second human face liveness discrimination model to construct a pixel-level classifier.

For the convenience of description, the above-mentioned step 1 and step 2 are combined for description. As shown in FIG. The basic framework of the human face discrimination model is the same, and will be described in a unified manner below), further, a pixel-level classification branch structure is additionally inserted at the output end of the third Invert residual block of the living body discrimination framework, where Conv1 is a 3x3 convolutional layer, and Bn is Batch normalization layer, Relu is a linear rectification layer, and Conv2 is a 1x1 convolutional layer.

Step 3: Process the original training image into a face image and a face image including a background.

Specifically, the face image and the face image including the background can be acquired in the same manner as in the above step S2, which will not be repeated here.

Step 4: Generating pixel-level classification labels (also referred to as classification label maps) for the face image and the face image including the background, respectively.

As an optional implementation, the specific construction method of the pixel-level label map is as follows:

First, generate a label image with the same size as the input image. The value of the label image is determined according to the category of the input image. For a face image: when the image is an attacking face, the value of the label image is all set to 0; when the input image is For a real face, the values in the label map are all set to 1; for a face image that includes a background: the value of the background area in the label map is set to 0, the value of the real face area is set to 1, and the value of the deceptive face area corresponds to The value of is set to 2; finally, the label map is scaled to the size of the output of the pixel-level classifier.

Step 5. Insert pixel-level classification losses into the first human face liveness discrimination model and the second human face liveness discrimination model respectively, and form a target loss function with Focal Loss.

Step 6, using the face image and the binary classification label to train the first human face liveness discriminant model, and using the face image including the background and the binary classification label to train the second human face liveness discriminant model.

，同时计算模型末端输出和二分类标签的Focal Loss损失，记为

，同时为了调节像素级分类损失对模型的影响，给与像素级分类损失权重，于是最终的目标损失函数可定义为

，

为指定像素级分类损失函数的权重。通过梯度下降法优化最终目标损失函数训练活体判别模型。Specifically, the first human face living body discrimination model and the second human face living body discrimination model are trained separately. During training, the cross-entropy classification loss of the output of the pixel-level classifier and the corresponding position label in the label map is calculated point by point, and the loss is denoted as

, and calculate the Focal Loss loss of the end output of the model and the two-category label at the same time, denoted as

, at the same time, in order to adjust the influence of pixel-level classification loss on the model, the weight of pixel-level classification loss is given, so the final target loss function can be defined as

,

For the weights of the specified pixel-level classification loss function. The final target loss function is optimized by the gradient descent method to train the living body discriminant model.

In order to facilitate those skilled in the art to understand the technical solution provided by the embodiment of the present invention, referring to FIG. 3 , the technical solution provided by the embodiment of the present invention will be described in detail below to obtain the fusion weight prediction sub-model as an example.

Step 1, as shown in Figure 3, construct the basic structure of the fusion weight prediction sub-model. Specifically, the backbone structure is formed by stacking convolutional layers, batch normalization layers, and linear rectification layers continuously. At the same time, a channel attention mechanism is introduced to assign importance weights to feature channels. Finally, a global pooling layer and a fully connected structure are inserted. The end of the fully connected layer outputs a 2-dimensional vector, which represents the weighted fusion weight output from the end of the first live face discrimination model and the second live face discrimination model.

Step 2, first input the human face image data and the human face image containing the background into the first human face living body discrimination model and the second human face living body discrimination model, extract the first human face living body discrimination model and the The feature map output by the fourth Invert residual block of the second human face living body discrimination model; then, according to the channel dimension, the above feature map is spliced, input to the fusion weight prediction sub-model to obtain the terminal output of the first human face living body discrimination model and The weighted fusion weight output by the end of the second human face liveness discrimination model.

Step 3: Based on the weighted fusion weights, the terminal outputs of the first liveness discrimination model and the second liveness discrimination model are weighted and summed to obtain a final liveness discrimination output.

Step 4: Optimizing the FocalLoss between the final in vivo discrimination output and the binary classification labels by using the gradient descent method to train the weight parameters of the fusion weight prediction sub-model. It should be noted that this step must be performed after the training of the first human face liveness discrimination model and the second human face liveness discrimination model is completed. Furthermore, the first human face liveness discrimination model and The weight of the second human face living body discrimination model only updates the parameters of the fusion weight prediction sub-model.

Apply the above algorithm to the near-infrared face intelligent face recognition module, and the performance comparison is shown in Table 1 and Table 2:

Table 1 Performance comparison between fusion and fusion models on near-infrared data

Model tpr@fpr=0.1%（↑） tpr@fpr=0.01%（↑） tpr@fpr=0.001%（↑） Model size (M) The first in vivo discriminant model 84.82 65.57 46.71 0.34 The second in vivo discriminant model 88.68 69.37 47.76 0.34 Multi-living discriminant model fusion 97.88 94.89 84.34 0.7

Table 2 Comparison of comprehensive performance with the same level of parameter model on near-infrared data

Model tpr@fpr=0.1% tpr@fpr=0.01% tpr@fpr=0.001% Params(M) ResNet18 88.10 73.54 \ 11.6 CDCNpp 83.43 36.55 12.82 2.26 CDCN 95.74 90.42 80.75 2.24 The first in vivo discriminant model 95.96 86.75 53.70 0.34 The second in vivo discriminant model 94.71 86.77 73.41 0.34 Multi-living discriminant model fusion (weighted average fusion) 97.73 94.43 80.27 0.68 Multi-living discriminant model fusion (fusion prediction module fusion) 97.88 94.89 84.34 0.7

Combining the above Table 1 and Table 2, the performance obtained through calculation can be obtained, which proves that the technical solutions provided by the embodiments of the present invention are practical (feasible).

In summary, compared with the prior art, this embodiment has the following advantages and beneficial effects:

(1) The embodiment of the present invention is implemented based on two ultra-lightweight near-infrared living body discrimination models, and the overall model parameter volume is <760KB, which can be deployed and applied on edge devices with low computing power, low power consumption and low memory.

(2) In the model training stage of the embodiment of the present invention, different scales of face input and supervision labels are used for the first and second live face discrimination models, so that the first live body discrimination model pays attention to the details of the face, and the second live body discrimination model focuses on face details. The discriminant model focuses on the contextual connection between the face and the background, increases the difference between the first living body discrimination model and the second living body discrimination model, and effectively improves the performance of the fusion of the two.

(3) In the model prediction fusion stage of the example of the present invention, the fusion weight sub-module is designed to dynamically determine the fusion weights predicted by the first living body discrimination model and the second living body discrimination model, and further, channel attention is introduced into the fusion weight sub-module The force mechanism enables the model to explicitly consider the contribution of face image information of different scales to the liveness discrimination task, further improving the performance of multi-model prediction fusion.

As shown in Figure 5, the present embodiment also provides a human face living body discrimination system, including:

Target image acquisition and face detection module 101, used to acquire the target image to be discriminated, and obtain the face detection frame of the target image;

The face image cropping module 102 is used to obtain a first face image and a second face image from the target image according to the face detection frame; wherein, the first face image is a face image that does not contain a background , the second face image is a face image containing a background;

The living body discrimination prediction module 103 is used to input the first human face image into the first human face living body discrimination model to obtain the first feature and the first output; input the second human face image into the second human face living body discrimination model , get the second feature and the second output;

The prediction fusion module 104 is used to input the fusion weight prediction sub-model after the first feature and the second feature are spliced to obtain the fusion weight of the two human face living body discrimination models;

A judging module 105, configured to obtain a final output of human face liveness discrimination according to the fusion weight, the first output and the second output.

A human face living body identification system in this embodiment can execute a human face living body identification method provided by the method embodiment of the present invention, can execute any combination of implementation steps of the method embodiment, and has the corresponding functions and beneficial effects of the method .

Referring to Fig. 6, the present embodiment also provides a human face living body discrimination device, including:

A near-infrared image collector 201, configured to collect near-infrared images;

at least one memory 202 for storing computer programs and near-infrared image data;

At least one processor 203 is configured to implement the steps of the living body discrimination method based on multi-model prediction fusion as shown in FIG. 1 when executing the computer program.

A human face living body identification device in this embodiment can execute a human face living body identification method provided by the method embodiment of the present invention, can execute any combination of implementation steps of the method embodiment, and has the corresponding functions and beneficial effects of the method .

The embodiment of the present application also discloses a computer program product or computer program, where the computer program product or computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. The processor of the computer device can read the computer instruction from the computer-readable storage medium, and the processor executes the computer instruction, so that the computer device executes the method shown in FIG. 1 .

This embodiment also provides a storage medium, which stores an instruction or a program that can execute a human face living body discrimination method provided by the method embodiment of the present invention. When the instruction or program is executed, any of the method embodiments can be executed. Combining the implementation steps has the corresponding functions and beneficial effects of the method.

In some alternative implementations, the functions/operations noted in the block diagrams may occur out of the order noted in the operational diagrams. For example, two blocks shown in succession may, in fact, be executed substantially concurrently or the blocks may sometimes be executed in the reverse order, depending upon the functionality/operations involved. Furthermore, the embodiments presented and described in the flowcharts of the present invention are provided by way of example in order to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logical flow presented herein. Alternative embodiments are contemplated in which the order of various operations is changed and in which sub-operations described as part of larger operations are performed independently.

Furthermore, although the invention has been described in the context of functional modules, it should be understood that one or more of the described functions and/or features may be integrated into a single physical device and/or unless stated to the contrary. or software modules, or one or more functions and/or features may be implemented in separate physical devices or software modules. It will also be appreciated that a detailed discussion of the actual implementation of each module is not necessary to understand the present invention. Rather, given the attributes, functions and internal relationships of the various functional blocks in the devices disclosed herein, the actual implementation of the blocks will be within the ordinary skill of the engineer. Accordingly, those skilled in the art can implement the present invention set forth in the claims without undue experimentation using ordinary techniques. It is also to be understood that the particular concepts disclosed are illustrative only and are not intended to limit the scope of the invention which is to be determined by the appended claims and their full scope of equivalents.

If the functions described above are realized in the form of software function units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the essence of the technical solution of the present invention or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including Several instructions are used to make a computer device (which may be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage media include: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk, and other media that can store program codes. .

The logic and/or steps represented in the flowcharts or otherwise described herein, for example, can be considered as a sequenced listing of executable instructions for implementing logical functions, can be embodied in any computer-readable medium, For use with an instruction execution system, device, or device (such as a computer-based system, a system including a processor, or other systems that can fetch instructions from an instruction execution system, device, or device and execute instructions), or in conjunction with such an instruction execution system, device or equipment used. For the purposes of this specification, a "computer-readable medium" may be any device that can contain, store, communicate, propagate or transmit a program for use in or in conjunction with an instruction execution system, device or device.

More specific examples (non-exhaustive list) of computer-readable media include the following: electrical connection with one or more wires (electronic device), portable computer disk case (magnetic device), random access memory (RAM), Read Only Memory (ROM), Erasable and Editable Read Only Memory (EPROM or Flash Memory), Fiber Optic Devices, and Portable Compact Disc Read Only Memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program can be printed, since the program can be read, for example, by optically scanning the paper or other medium, followed by editing, interpretation or other suitable processing if necessary. The program is processed electronically and stored in computer memory.

It should be understood that various parts of the present invention can be realized by hardware, software, firmware or their combination. In the embodiments described above, various steps or methods may be implemented by software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented by any one or combination of the following techniques known in the art: Discrete logic circuits, ASICs with suitable combinational logic gates, Programmable Gate Arrays (PGAs), Field Programmable Gate Arrays (FPGAs), etc.

In the above description of this specification, the description with reference to the terms "one embodiment/example", "another embodiment/example" or "some embodiments/example" means that the description is described in conjunction with the embodiment or example. A particular feature, structure, material, or characteristic is included in at least one embodiment or example of the invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the described specific features, structures, materials or characteristics may be combined in any suitable manner in any one or more embodiments or examples.

Although the embodiments of the present invention have been shown and described, those skilled in the art can understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principle and spirit of the present invention. The scope of the invention is defined by the claims and their equivalents.

The above is a specific description of the preferred implementation of the present invention, but the present invention is not limited to the above-mentioned embodiments, and those skilled in the art can also make various equivalent deformations or replacements without violating the spirit of the present invention. Equivalent modifications or replacements are all within the scope defined by the claims of the present application.

Claims (10)

1. A face living body distinguishing method is characterized by comprising the following steps:

acquiring a target image to be distinguished, and acquiring a face detection frame of the target image;

acquiring a first face image and a second face image from the target image according to the face detection frame; the first face image is a face image without a background, and the second face image is a face image with a background;

inputting the first face image into a first face living body distinguishing model to obtain a first characteristic and a first output result;

inputting the second face image into a second face living body distinguishing model to obtain a second characteristic and a second output result;

splicing the first feature and the second feature and then inputting the spliced first feature and the spliced second feature into a fusion weight predictor model to obtain fusion weights of two human face living body discrimination models;

and acquiring a final output result of the living human face judgment according to the fusion weight, the first output result and the second output result.

2. The method for determining the living human face according to claim 1, wherein the acquiring a first human face image and a second human face image from the target image according to the human face detection frame includes:

cutting out a first face image from the target image according to the face detection frame;

amplifying the face detection frame to obtain a second detection frame;

and cutting out a second face image from the target image according to a second detection frame.

3. The living human face identification method according to claim 1, wherein the first living human face identification model and the second living human face identification model have the same basic framework and are both lightweight network models;

the living face discrimination model is formed by stacking a plurality of invirt residual blocks of mobrieetv 2.

4. The living human face identification method according to claim 3, wherein the first living human face identification model and the second living human face identification model are trained by:

inserting a point convolution layer, a batch normalization layer and a linear rectification layer into output ends of intermediate layers of a first human face living body discrimination model and a second human face living body discrimination model to construct a pixel-level classifier;

acquiring a face image and a face image containing a background according to a training set, and respectively generating pixel level classification labels of the face image and the face image containing the background according to a pixel level classifier;

inserting pixel level classification Loss into the first face living body discrimination model and the second face living body discrimination model, and forming a target Loss function together with the Focal local;

and optimizing the target loss function according to the two classification labels and the pixel classification label, and training the first face living body discrimination model and the second face living body discrimination model according to the target loss function.

5. The living human face identification method according to claim 4, wherein the generation process of the pixel-level classification label comprises:

for pixel level classification labels of face images: constructing a label graph with the same size as the output of the pixel-level classifier, and setting all values of the classified label graph to be 0 when the face image is in a deception face category; when the face image is a real face, setting the values of the classification label images to be 1;

for pixel level classification labels of face images containing backgrounds: constructing a label graph with the same size as the output of the pixel-level classifier, wherein when the input image is a deception face, the corresponding value in a deception face area in the label graph is 2, and the value of a background part is 0; when the input image is a real face, the value of the face area in the label image is 1, and the value of the background part is 0.

6. The method according to claim 5, wherein the training the first living human face identification model and the second living human face identification model according to the target loss function comprises:

training the first face living body distinguishing model and the second face living body distinguishing model separately;

during training, calculating the cross entropy classification loss of the output of the pixel-level classifier and the corresponding position label in the label graph point by point, and recording the loss as

Calculate the Loss of the Focal local of the model's end output and the binary label simultaneously, and record it as

(ii) a The objective loss function is defined as:

in the formula (I), the compound is shown in the specification,

the weights of the loss function are classified for a given pixel level.

7. The method for discriminating living human faces according to claim 1, wherein the fusion weight predictor model constructs a basic structure of the fusion weight predictor model by stacking a convolution layer, a batch normalization layer and a linear rectifying layer to form a main structure and introducing a channel attention mechanism;

the fusion weight predictor model is trained in the following way:

acquiring a first feature map output by the first human face living body distinguishing model and a second feature map output by the second human face living body distinguishing model;

after the first feature map and the second feature map are spliced, inputting a fusion weight predictor model, and outputting a weighted fusion weight output by the tail end of the first face living body discrimination model and the tail end of the second face living body discrimination model;

weighting and summing the tail end output of the first face living body distinguishing model and the tail end output of the second face living body distinguishing model by using the weighted fusion weight to obtain a final output;

calculating the Focal local of the final output and the two classification labels, and optimizing the Loss by using a random gradient descent method to train a fusion weight predictor model;

and during training, freezing the weight parameters of the first human face living body discrimination model and the second human face living body discrimination model.

8. A living human face discrimination system, comprising:

the target image acquisition and face detection module is used for acquiring a target image to be distinguished and acquiring a face detection frame of the target image;

the face image cutting module is used for acquiring a first face image and a second face image from the target image according to the face detection frame; the first face image is a face image without a background, and the second face image is a face image with a background;

the living body distinguishing and predicting module is used for inputting the first face image into a first face living body distinguishing model to obtain a first characteristic and a first output; inputting the second face image into a second face living body distinguishing model to obtain a second characteristic and a second output;

the prediction fusion module is used for splicing the first feature and the second feature and then inputting the spliced first feature and second feature into a fusion weight prediction submodel to obtain fusion weights of the two human face living body distinguishing models;

and the judging module is used for acquiring the final output of the human face living body judgment according to the fusion weight, the first output and the second output.

9. A living human face discrimination apparatus, comprising:

at least one processor;

at least one memory for storing at least one program;

when executed by the at least one processor, cause the at least one processor to implement the method of any one of claims 1-7.

10. A computer-readable storage medium, in which a program executable by a processor is stored, wherein the program executable by the processor is adapted to perform the method according to any one of claims 1 to 7 when executed by the processor.

Images