Online car-hailing driver living body face recognition method

Abstract

The invention relates to the field of face recognition, and discloses an online car-hailing driver living body face recognition method, which is used for solving the problem that the effect of a face recognition model is greatly reduced when different scenes need to be migrated, and comprises the following steps: collecting a verification picture when a driver leaves a car under a real condition, and marking the verification picture; the method comprises the following steps: marking as a living body or a non-living body, designing a novel convolutional network model through a splitting convolution operation, carrying out feature extraction through the novel convolutional network model, and after feature extraction, accessing Softmax and a cross entropy loss function after features so as to obtain a feature extraction model; by minimizing the cross entropy loss function, the model is promoted to learn the feature difference between the living body face and the non-living body face, the living body face recognition accuracy is effectively improved, and the reasoning time of the model is shortened.

Description

Technical Field

[0001] The present invention relates to the field of face recognition, and more specifically to a method for live face recognition of online car-hailing drivers. Background Art

[0002] The main goal of early face liveness recognition technology was to artificially design features based on the differences between living and non-living objects. These differences included color texture, non-rigid motion deformation, material, image or video quality, etc., and then send the features to the classifier for decision-making.

[0003] After the era of deep learning, various convolutional neural network models have been used for face detection and liveness recognition tasks. By constructing a binary or multi-classification loss function and minimizing the loss function, the convolutional network is able to learn the difference between live faces and non-live faces, thereby achieving liveness recognition. The training of large-scale neural network models depends on large-scale live and non-live samples. Common data sets include NUAA, Celeb-Spoof, CASIA-SURF, etc. The early data set NUAA only contains 7,500 pictures of 15 people, while CASIA-SURF contains 21,000 videos of 1,000 people. As the size of the data set increases, the performance of the convolutional network model quickly surpasses the manually designed features.

[0004] Existing live face recognition is based on multiple public live face datasets, but the model has poor transferability in different scenarios. A model that performs well in one live face dataset may perform poorly in another dataset. This is because different samples are collected at different angles, lighting, and shooting equipment. The inconsistent collection conditions of training samples and test samples have greatly reduced the model effect. At present, there is no public dataset for the online car-hailing driver login verification scenario, so it is difficult to train and optimize the network model in this scenario.

[0005] In pursuit of improved recognition results, the scale of convolutional neural networks is usually very large, with a large number of parameters and a high amount of multiplication operations, which results in a long model inference time. At the same time, the model can only run on devices such as CPU, GPU, TPU, and cannot be deployed on mobile phones, tablets, and embedded devices.

[0006] In view of the above problems, the present invention proposes a solution. Summary of the invention

[0007] In order to overcome the above-mentioned defects of the prior art, the present invention provides a method for live face recognition of online car-hailing drivers to solve the problems existing in the above-mentioned background technology.

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] A method for live face recognition of online car-hailing drivers, comprising the following steps:

[0010] S001: Collect verification photos of drivers when they are out in real conditions, and mark them as live or non-live, filling the gap in the dataset for live face recognition of online ride-hailing drivers;

[0011] S002: Design a new convolutional network model by splitting the convolution operation, and extract features through the new convolutional network model;

[0012] S003: After feature extraction, Softmax and cross entropy loss functions are connected after the features. By minimizing the cross entropy loss function, the model is prompted to learn the feature differences between living faces and non-living faces.

[0013] Preferably, the step of designing a new convolutional network model by splitting the convolution operation is to split a labeled convolution operation into two convolution operations, the first one has only one convolution kernel with a size of w×h×C1, and each layer of the convolution kernel parameters only acts on the corresponding layer of the input feature map, the second one has C2 convolution kernels, the size of each kernel is 1×1×C2, the parameter quantities of the two convolutions are w×h×C1 and C1×C2 respectively, and the multiplication operations are w×h×C1×W×H and W×H×C2×C1 respectively.

[0014] Preferably, the step of extracting features through a novel convolutional network model is:

[0015] Pre-process the collected verification photos of the driver when he / she leaves the vehicle, including cropping, resizing, graying, etc.

[0016] According to the requirements of the task and the characteristics of the data set, design the convolutional neural network structure, including convolutional layer, pooling layer and fully connected layer components, and select the pre-trained model for fine-tuning according to the actual situation;

[0017] The pre-processed verification photo of the driver is used as input to the constructed convolutional neural network. In the first few layers of the network, low-level features of the face image are extracted through convolution and pooling operations. These low-level features are used as input to the subsequent network layers to further extract high-level features. High-level features of the face image are extracted through deeper convolution and fully connected layers.

[0018] Add a feature representation learning layer before the last layer or fully connected layer of the convolutional neural network to convert the network's output into a compact feature vector, using global average pooling or other techniques to reduce the feature dimension and extract key features;

[0019] The extracted feature vector is input into the liveness detector, and the face is judged to be a live face or a non-live face according to the representation of the feature vector;

[0020] Use the collected data sets to train the model, including the feature extraction network and the living body detector, adjust the model's hyperparameters, and optimize the loss function to improve the model's performance.

[0021] Preferably, the step of extracting low-level features of the face image through convolution and pooling operations is:

[0022] The convolution operation extracts local features from the image by convolving the filter with the input image. Each filter slides on the image and performs point-by-point multiplication with each pixel of the image. The multiplication results are then added to obtain a pixel value of the output feature map, which is expressed as y(i,j)=∑ m ∑ n x(i+m,j+n)·h(m,n), where y(i,j) represents the pixel value of the output feature map, h(m,n) represents the weight of the filter, m and n are the indexes of the filter, and i and j are the pixel indexes of the output feature map;

[0023] Perform a pooling operation on the input image. The pooling operation is used to reduce the size of the feature map, reduce the amount of calculation, and enhance the translation invariance of the model. The commonly used pooling operation is maximum pooling, and its expression is y(i,j)=max m,n (x(i×s+m,j×s+n)), where y(i,j) represents the pixel value of the pooled feature map, x(i×s+m,j×s+n) represents the local area in the input feature map, s represents the pooling step size, and m, n are the pixel indices of the red area.

[0024] Preferably, the step of extracting high-level features of the face image through deeper convolution and fully connected layers is:

[0025] Using deeper convolutional layers can further extract abstract features from the image, using more filters and more convolutional layers to increase the representation and learning capabilities of the network;

[0026] After the convolutional layer, a fully connected layer is added to learn the complex relationship between the input data and generate the final output. The fully connected layer flattens the output of the convolutional layer and connects it to one or more fully connected layers, and finally outputs a feature vector. The mathematical expression of the fully connected layer is z=f(W·x+b), where x represents the input feature vector, W represents the weight matrix of the fully connected layer, b is the bias vector, and f represents the activation function.

[0027] Preferably, the step of judging whether the face is a live face or a non-live face according to the representation of the feature vector is:

[0028] Collecting a data set containing live face images and non-live face images, and extracting feature vectors of the live face images and non-live face images as training data;

[0029] Through the support vector machine learning model, a liveness detector model is trained using the training data. The model receives the extracted feature vector as input and outputs the probability that the face is a real live body.

[0030] Use the validation set to evaluate the trained liveness detector model and its performance on liveness detection tasks. Adjust and optimize the model based on the evaluation results.

[0031] The extracted feature vector is input into the trained liveness detector model to obtain the probability that the face is truly alive. The obtained probability of true liveness is compared with the preset threshold. If the probability of true liveness is less than the preset threshold, it is judged as a non-live face. If the probability of true liveness is greater than the preset threshold, it is judged as a live face.

[0032] Preferably, the steps of connecting Softmax and cross entropy loss function after the features and minimizing the cross entropy loss function to enable the model to learn the feature differences between living faces and non-living faces are:

[0033] The feature vector is input into the Softmax classifier for classification, and its function expression is Where P(y=j|x) is the probability of category j given input x, and K represents the total number of categories;

[0034] The cross entropy loss function is used to measure the difference between the probability distribution of the model output and the true label. Its function expression is: where y i is represented as the true label, is the predicted output of the model, and N is the number of samples;

[0035] Backpropagation and parameter update using the cross entropy loss function on the training data, using gradient descent or its variants to update the model's parameters;

[0036] The trained model is evaluated using the validation set, and adjustments and optimizations are made based on the evaluation results. The technical effects and advantages of the present invention are:

[0037] Verification photos of drivers when they drive under real conditions are collected and annotated as living or non-living. A new convolutional network model is designed through split convolution operations, and features are extracted through the new convolutional network model. After feature extraction, Softmax and cross-entropy loss functions are connected after the features. The cross-entropy loss function is minimized to enable the model to learn the feature differences between living faces and non-living faces.

[0038] It fills the gap of large-scale data sets for liveness verification scenarios of online ride-hailing drivers, and proposes a standard convolution decomposition method. This method greatly reduces the number of parameters and computational complexity of the model while ensuring the feature extraction effect, allowing the model to be deployed on embedded devices such as mobile phones. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is the overall flow chart of the present invention.

[0040] Figure 2 This is a standard convolution operation flowchart.

[0041] Figure 3 Split the flowchart for the convolution operation.

[0042] Figure 4 The following is a flow chart of the specific calculation of the two losses. DETAILED DESCRIPTION

[0043] The technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the present invention. In addition, the forms of the various structures recorded in the following embodiments are only examples. The method for live face recognition of an online car-hailing driver involved in the present invention is not limited to the various structures recorded in the following embodiments. All other implementations obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0044] The present invention provides a method for live face recognition of online car-hailing drivers, comprising the following steps:

[0045] S001: Collect verification photos of drivers when they are driving under real conditions, and mark them as living or non-living, filling the gap in the data set in this scenario;

[0046] S002: Design a new convolutional network model by splitting the convolution operation, and extract features through the new convolutional network model;

[0047] S003: After feature extraction, Softmax and cross entropy loss functions are connected after the features. By minimizing the cross entropy loss function, the model is prompted to learn the feature differences between living faces and non-living faces.

[0048] Real photos represent real scenes and situations. Using real photos to train models can better simulate actual application scenarios, allowing the model to better adapt to changes and noise in real situations. Real photos contain a variety of faces, covering different races, ages, genders, expressions, postures and other factors, which can help the model learn richer and more comprehensive facial features and improve the generalization ability of the model. Using real photos for training can reduce the risk of model overfitting, because real photos represent the real data distribution. The model is more likely to learn universal features rather than overfitting specific samples in the training set. Real photos may contain a variety of environmental factors, such as lighting conditions, background noise, etc. Using real photos to train models can make the model more robust and better adapt to face recognition tasks under various complex environmental conditions.

[0049] The goal of convolutional neural networks is to extract features and make the output features able to distinguish between living and non-living objects. At the same time, the network is as lightweight as possible, that is, the network parameters and computational complexity are compressed as much as possible. Figure 2 As shown, the standard convolution operation steps are:

[0050] The feature map f in As the input signal, the input channel number C1 is used to determine the filter C2 used for the convolution operation. The number of filters C2 is K. And the number of convolution kernel parameters is w×h×C1×C2. The number of convolution kernel parameters refers to the number of parameters that need to be learned in the convolution kernel. The number of multiplication operations is w×h×C1×W×H and W×H×C2×C1 respectively. The overall number of multiplication operations refers to the total number of multiplication operations performed in the entire network. This value is one of the important indicators for measuring the computational complexity of the network and can be used to evaluate the computational resource requirements and operating efficiency of the network. The feature map refers to the output image generated in the convolution layer. Since the convolution layers are stacked in the network, the input and output of the convolution layer are both feature maps. Feature maps at different levels can capture image features of different scales and semantic levels. The feature map has three dimensions, namely, length Width, height Height, and number of channels Channel, usually expressed as W×H×C. The convolution kernel is also called a filter, which is the smallest unit of feature extraction and detection. A convolution operation has the following dimensions: padding, stride, length Width, height Height;

[0051] Pad the input signal as needed to ensure that the output is the same size as the input or has the required size;

[0052] Before the convolution operation, it is usually necessary to flip the filter to align it with the input signal;

[0053] Multiply the flipped filter by the input signal point by point, and add the results of the product to get each pixel or point of the output;

[0054] When performing convolution calculations, the filter usually moves along the input signal, gradually covering the entire input signal;

[0055] Special processing may be required at the boundaries, such as using zero padding or reflection padding, to ensure that the convolution operation is performed correctly at the boundaries;

[0056] The output result is the result of the convolution operation f out , usually a feature map or a signal processed by a filter, and

[0057] In this embodiment, it should be specifically explained that the step of designing a new convolutional network model by splitting the convolution operation is to split a labeled convolution operation into two convolution operations. The first time, there is only one convolution kernel, whose size is w×h×C1, and each layer parameter of the convolution kernel only acts on the corresponding layer of the input feature map. The second time, there are C2 convolution kernels, and the size of each kernel is 1×1×C2, such as Figure 3 As shown in the figure, the parameters of the two convolutions are w×h×C1 and C1×C2, and the multiplication operations are w×h×C1×W×H and W×H×C2×C1, respectively.

[0058] The steps of splitting a labeled convolution operation into two convolution operations are to decompose a standard convolution into single-channel features and then learn cross-channel deep features. This reduces the number of convolution parameters and the amount of computation. Usually, we use 3×3 convolution, and we can conclude that the number of parameters and the amount of multiplication operations are approximately the same as before. This network can also realize feature extraction, but the network parameter volume is only 4.2M, which greatly improves the model inference speed. The specific network structure is described in Table 1, where Conv, Conv1, and Conv2 represent the standard convolution, the first convolution, and the second convolution mentioned above, respectively.

[0059]

[0060]

[0061] Table 1

[0062] In this embodiment, it should be specifically explained that the step of extracting features through the novel convolutional network model is:

[0063] Pre-process the collected verification photos of drivers when they leave the vehicle, including cropping, resizing, graying, etc., to ensure the consistency and availability of input data;

[0064] According to the requirements of the task and the characteristics of the dataset, design a suitable convolutional neural network structure, including components such as convolutional layers, pooling layers, and fully connected layers. Depending on the actual situation, you can choose to use a pre-trained model (such as VGG, ResNet, MobileNet, etc.) and perform fine-tuning;

[0065] The pre-processed verification photo of the driver when driving is used as input to the constructed convolutional neural network. In the first few layers of the network, low-level features of the face image, such as edges and textures, are extracted through convolution and pooling operations. In the latter layers of the network, these low-level features can be used as inputs for subsequent network layers to further extract high-level features. High-level features of the face image, such as facial structure and expression, are extracted through deeper convolution and fully connected layers.

[0066] Add a feature representation learning layer before the last layer or fully connected layer of the convolutional neural network to convert the network's output into a compact feature vector. Global average pooling or other techniques can be used to reduce the feature dimension and extract key features.

[0067] The extracted feature vector is input into a liveness detector, which can be a classifier (such as a support vector machine, logistic regression, etc.) or a deep learning model, and determines whether the face is a live face or a non-live face based on the representation of the feature vector;

[0068] Use the collected data sets to train the model, including the feature extraction network and the living body detector, adjust the model's hyperparameters, optimize the loss function, and improve the model's performance.

[0069] In this embodiment, it should be specifically explained that the steps of extracting low-level features of the face image through convolution and pooling operations are:

[0070] The convolution operation performs convolution operation with the input image through a filter (convolution kernel) to extract local features in the image. Each filter slides on the image and performs point-by-point product with each pixel of the image. The results of the product are then added to obtain a pixel value of the output feature map. The expression of the convolution operation is y(i,j)=∑ m ∑ n x(i+m,j+n)·h(m,n), where y(i,j) represents the pixel value of the output feature map, h(m,n) represents the weight of the filter, m and n are the indexes of the filter, and i and j are the pixel indexes of the output feature map;

[0071] Perform a pooling operation on the input image. The pooling operation is used to reduce the size of the feature map, reduce the amount of calculation, and enhance the translation invariance of the model. The commonly used pooling operation is maximum pooling, and its expression is y(i,j)=maxm,n (x(i×s+m,j×s+n)), where y(i,j) represents the pixel value of the pooled feature map, x(i×s+m,j×s+n) represents the local area in the input feature map, s represents the pooling step size, and m, n are the pixel indices of the red area.

[0072] The weight of the filter is a key component in a convolutional neural network (CNN) and is used to extract features from input data. In a convolution operation, the weight is a parameter in a convolution kernel (also called a filter) and is used to perform a convolution operation on the input data. Specifically, the weight of the filter is a multidimensional matrix whose size is determined by the size of the convolution kernel and the number of channels of the input data. Each element in the weight matrix of each filter is a parameter that needs to be learned and is used as a weight coefficient when performing a convolution operation on the input data.

[0073] The filter index generally refers to the identification number or serial number of the filter in the convolution layer. In a convolution layer, multiple filters may be included, each filter is used to extract different features in the input data, and the filter index is used to specify which filter is used for convolution operation at which position in the convolution layer.

[0074] The pixel index of the feature map refers to the position identifier of each pixel in the feature map. The feature map is a two-dimensional matrix, and each element in it is called a pixel. The pixel index of the feature map is usually represented by a two-dimensional coordinate, such as (i, j), where i represents the row index of the feature map, j represents the column index of the feature map, and the index starts from 0, for example (0, 0) represents the upper left corner pixel of the feature map.

[0075] In this embodiment, it should be specifically explained that the steps of extracting high-level features of the face image through deeper convolution and fully connected layers are:

[0076] Using deeper convolutional layers can further extract abstract features in the image, such as facial structure, expression, etc., and using more filters and more convolutional layers to increase the representation and learning capabilities of the network;

[0077] After the convolution layer, a fully connected layer is usually added to learn the complex relationship between the input data and generate the final output. The fully connected layer flattens the output of the convolution layer and connects it to one or more fully connected layers, and finally outputs a feature vector. The mathematical expression of the fully connected layer is z=f(W·x+b), where x represents the input feature vector, W represents the weight matrix of the fully connected layer, b is the bias vector, and f represents the activation function.

[0078] The weight matrix of the fully connected layer is a matrix used to convert the input feature vector into the output feature vector. Specifically, if the dimension of the input feature vector of the fully connected layer is n and the dimension of the output feature vector is m, then the size of the weight matrix of the fully connected layer is m×n. Each element in this weight matrix represents the connection weight between each feature in the input feature vector and each node in the output feature vector.

[0079] The bias vector is a parameter in the neural network, which is used to adjust the output of the neuron. In the fully connected layer and the partial convolutional layer, each neuron has a bias parameter, which is associated with the output of the neuron. The role of the bias vector is to add a fixed offset to the output of the neuron, thereby changing the activation state of the neuron, which enables the neural network to fit the data more flexibly and learn more complex functional relationships.

[0080] The activation function is a function in a neural network, which introduces nonlinear transformation on the output of neurons to increase the network's expressiveness and learning capabilities. In deep learning, common activation functions include Sigmoid function, ReLU function, Leaky ReLU function, and Softmax function. ReLU and its variants are widely used in practice because of their fast calculation speed and stable training. The calculation formula is ReLU(x)=max(0,x).

[0081] In this embodiment, it should be specifically explained that the step of judging whether a face is a live face or a non-live face according to the representation of the feature vector is as follows:

[0082] Collecting a data set containing live face images and non-live face images, and extracting feature vectors of the live face images and non-live face images as training data;

[0083] Through the support vector machine learning model, a liveness detector model is trained using the training data. The model receives the extracted feature vector as input and outputs the probability that the face is a real live body.

[0084] Use the validation set to evaluate the trained liveness detector model and its performance on liveness detection tasks. Based on the evaluation results, adjust and optimize the model to improve its performance in practical applications.

[0085] The extracted feature vector is input into the trained liveness detector model to obtain the probability that the face is truly alive. The obtained probability of true liveness is compared with the preset threshold. If the probability of true liveness is less than the preset threshold, it is judged as a non-live face. If the probability of true liveness is greater than the preset threshold, it is judged as a live face.

[0086] In this embodiment, it should be specifically explained that Figure 3 As shown, the steps of connecting Softmax and cross entropy loss function after the features and minimizing the cross entropy loss function to enable the model to learn the feature differences between living faces and non-living faces are as follows:

[0087] The feature vector is input into the Softmax classifier for classification. The Softmax function converts each element of the feature vector into a value representing the probability of the corresponding category. Its function expression is: Where P(y=j|x) is the probability of category j given input x, and K represents the total number of categories;

[0088] The cross entropy loss function is used to measure the difference between the probability distribution of the model output and the true label. Its function expression is: where y i is represented as the true label, is the predicted output of the model, and N is the number of samples;

[0089] Use the cross entropy loss function on the training data for back propagation and parameter update to minimize the loss function, and use gradient descent or its variants to update the parameters of the model so that the model can better distinguish the feature differences between living and non-living faces;

[0090] Use the validation set to evaluate the trained model and adjust and optimize it based on the evaluation results to improve the performance of the model on the test set.

[0091] The cross entropy loss function consists of two parts. The specific calculation process of the two parts of the loss function is as follows: Figure 4 As shown in the figure, the feature map output by the convolutional network enters the classification network, outputs the category probability of live recognition and then calculates the SoftmaxLoss. On the other hand, the feature map passes through a Fourier transform learning network, outputs the prediction of the Fourier transform map of the original image, and then calculates the loss of the predicted Fourier transform map and the real Fourier transform map, adding the Fourier transform of the input image as the learning target.

[0092] During the training process, the final optimized loss function is the sum of FT Loss and Softmax Loss. Through the stochastic gradient descent algorithm, the loss value is continuously reduced to obtain the optimal network model parameters.

[0093] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

[0094] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for live face recognition of online car-hailing drivers, characterized in that: The following steps are involved: S001: Collect verification photos of drivers when they are out in real conditions, and mark them as live or non-live, filling the gap in the dataset for live face recognition of online ride-hailing drivers; S002: Design a new convolutional network model by splitting the convolution operation, and extract features through the new convolutional network model; S003: After feature extraction, Softmax and cross entropy loss functions are connected after the features. By minimizing the cross entropy loss function, the model is prompted to learn the feature differences between living faces and non-living faces.

2. A method for live face recognition of online car-hailing drivers according to claim 1, characterized in that: The steps of designing a new convolutional network model by splitting the convolution operation are as follows: splitting a labeled convolution operation into two convolution operations, the first one has only one convolution kernel with a size of w×h×C1, and each layer of the convolution kernel parameters only acts on the corresponding layer of the input feature map, the second one has C2 convolution kernels, the size of each kernel is 1×1×C2, the parameter quantities of the two convolutions are w×h×C1 and C1×C2 respectively, and the multiplication operations are w×h×C1×W×H and W×H×C2×C1 respectively.

3. A method for live face recognition of online car-hailing drivers according to claim 1, characterized in that: The feature extraction steps using the novel convolutional network model are as follows: Pre-process the collected verification photos of the driver when he / she leaves the vehicle, including cropping, resizing, graying, etc. According to the requirements of the task and the characteristics of the data set, design the convolutional neural network structure, including convolutional layer, pooling layer and fully connected layer components, and select the pre-trained model for fine-tuning according to the actual situation; The pre-processed verification photo of the driver is used as input to the constructed convolutional neural network. In the first few layers of the network, low-level features of the face image are extracted through convolution and pooling operations. These low-level features are used as input to the subsequent network layers to further extract high-level features. High-level features of the face image are extracted through deeper convolution and fully connected layers. Add a feature representation learning layer before the last layer or fully connected layer of the convolutional neural network to convert the network's output into a compact feature vector, using global average pooling or other techniques to reduce the feature dimension and extract key features; The extracted feature vector is input into the liveness detector, and the face is judged to be a live face or a non-live face according to the representation of the feature vector; Use the collected data sets to train the model, including the feature extraction network and the living body detector, adjust the model's hyperparameters, and optimize the loss function to improve the model's performance.

4. A method for live face recognition of online car-hailing drivers according to claim 3, characterized in that: The steps of extracting low-level features of face images through convolution and pooling operations are as follows: The convolution operation extracts local features from the image by convolving the filter with the input image. Each filter slides on the image and performs point-by-point multiplication with each pixel of the image. The multiplication results are then added to obtain a pixel value of the output feature map, which is expressed as y(i,j)=∑ m ∑ n x(i+m,j+n)·h(m,n), where y(i,j) represents the pixel value of the output feature map, h(m,n) represents the weight of the filter, m and n are the indexes of the filter, and i and j are the pixel indexes of the output feature map; Perform a pooling operation on the input image. The pooling operation is used to reduce the size of the feature map, reduce the amount of calculation, and enhance the translation invariance of the model. The commonly used pooling operation is maximum pooling, and its expression is y(i,j)=max m,n (x(i×s+m,j×s+n)), where y(i,j) represents the pixel value of the pooled feature map, x(i×s+m,j×s+n) represents the local area in the input feature map, s represents the pooling step size, and m, n are the pixel indices of the red area.

5. A method for live face recognition of online car-hailing drivers according to claim 3, characterized in that: The steps of extracting high-level features of face images through deeper convolution and fully connected layers are: Using deeper convolutional layers can further extract abstract features from the image, and using more filters and more convolutional layers can increase the representation and learning capabilities of the network; After the convolutional layer, a fully connected layer is added to learn the complex relationship between the input data and generate the final output. The fully connected layer flattens the output of the convolutional layer and connects it to one or more fully connected layers, and finally outputs a feature vector. The mathematical expression of the fully connected layer is z=f(W·x+b), where x represents the input feature vector, W represents the weight matrix of the fully connected layer, b is the bias vector, and f represents the activation function.

6. A method for live face recognition of online car-hailing drivers according to claim 3, characterized in that: The steps of judging whether a face is a live face or a non-live face according to the representation of the feature vector are as follows: Collecting a data set containing live face images and non-live face images, and extracting feature vectors of the live face images and non-live face images as training data; Through the support vector machine learning model, a liveness detector model is trained using the training data. The model receives the extracted feature vector as input and outputs the probability that the face is a real live body. Use the validation set to evaluate the trained liveness detector model and its performance on liveness detection tasks. Adjust and optimize the model based on the evaluation results. The extracted feature vector is input into the trained liveness detector model to obtain the probability that the face is truly alive. The obtained probability of true liveness is compared with the preset threshold. If the probability of true liveness is less than the preset threshold, it is judged as a non-live face. If the probability of true liveness is greater than the preset threshold, it is judged as a live face.

7. A method for live face recognition of online car-hailing drivers according to claim 1, characterized in that: The steps of connecting Softmax and cross entropy loss function after the features and minimizing the cross entropy loss function to enable the model to learn the feature differences between living faces and non-living faces are as follows: The feature vector is input into the Softmax classifier for classification, and its function expression is Where P(y=j|x) is the probability of category j given input x, and K represents the total number of categories; The cross entropy loss function is used to measure the difference between the probability distribution of the model output and the true label. Its function expression is: where y i is represented as the true label, is the predicted output of the model, and N is the number of samples; Backpropagation and parameter update using the cross entropy loss function on the training data, using gradient descent or its variants to update the model's parameters; Use the validation set to evaluate the trained model and adjust and optimize it based on the evaluation results.

Images