CN116977819A - Facial expression recognition neural network architecture searching method, device and application - Google Patents

Abstract

The invention discloses a facial expression recognition neural network searching method, a device and application, wherein the searching method comprises the steps of constructing a searching space facing facial expression recognition; generating a plurality of facial expression recognition candidate networks meeting preset parameters in a random search mode; selecting a preset number of zero-shot agents meeting preset conditions from the NAS reference data set; and selecting an optimal candidate network from the plurality of facial expression recognition candidate networks by adopting the selected zero-shot agent as a final facial expression recognition network. The application method of the facial expression recognition network comprises the steps that the facial expression recognition network is obtained by searching by adopting a structure search method, and the facial expression recognition network is trained by selecting a facial expression image data set; acquiring a facial expression image to be identified, and preprocessing the facial expression image; and inputting the preprocessed facial expression image into a trained facial expression recognition network for recognition, and obtaining a recognition result.

Description

Facial expression recognition neural network architecture search method, device and application

Technical field

The invention relates to facial expression recognition technology, and in particular to a facial expression recognition neural network architecture search method, device and application.

Background technique

Facial Expression Recognition (FER) is an important research direction in the field of computer vision. It has broad application prospects and huge commercial value. For example, in the fields of human-computer interaction such as distance education, safe driving and health care, more intelligent and personalized services can be provided by identifying the user's emotional state. Currently, Convolutional Neural Network (CNN) has achieved great success in facial expression recognition tasks due to its powerful feature extraction capabilities and flexible hierarchical expression. However, traditional CNN-based FER networks often require a large amount of computing resources to run, which leads to great limitations in their deployment on resource-constrained smart devices such as mobile phones and laptops.

For example, currently online learning has received many positive reviews and provides students with a flexible learning method, but teachers cannot understand students' learning status in real time to receive classroom feedback and make adjustments. The facial expression recognition model can collect students' learning status in real time, thereby improving the quality of online teaching. However, online learning is usually conducted on mobile devices (mobile phones, tablets, etc.), so the FER model needs to be deployed on mobile devices. However, the existing high-precision FER models occupy a large amount of resources and are limited by mobile devices. The resource limit prevents the application from being deployed directly. Therefore, designing lightweight FER models for these resource-constrained devices to enhance the intelligence of these devices has become an urgent problem to be solved.

In order to implement the deployment of FER models on mobile terminals, lightweight FER model construction methods have emerged on the market. Among them, the sampling-based NAS method samples the network in the search space under certain rules and uses the evaluated model performance As a feedback signal, it guides the algorithm to search for a better FER model. This method requires training multiple FER models, and the training of FER models is a very time-consuming process, which makes it difficult to quickly design FER networks that occupy smaller resources according to different mobile terminals in reality, which further limits the application of FER in different applications. Promotion and application in mobile terminals.

Contents of the invention

In view of the above-mentioned deficiencies in the prior art, the facial expression recognition neural network architecture search method, device and application provided by the present invention solve the problem that it is difficult for the existing architecture search method to quickly design a FER network that occupies smaller resources for different mobile terminals.

In order to achieve the above-mentioned object of the invention, the technical solutions adopted by the present invention are:

In the first aspect, a facial expression recognition neural network architecture search method and device are provided, which includes the steps:

S1. Construct a search space for facial expression recognition;

S2. Use random search to generate several facial expression recognition candidate networks that meet the preset parameters;

S3. Select a preset number of zero-shot agents that meet the preset conditions in the NAS data set;

S4. Use the selected zero-shot agent to select the optimal candidate network from several facial expression recognition candidate networks as the final facial expression recognition network.

Further, the step S3 further includes:

S31. Select several zero-shot proxies, randomly sample several neural networks in multiple NAS benchmark data sets, and use zero-shot proxies to calculate the accuracy-based proxy value of each neural network;

S32. Select a preset number of zero-shot agents based on the correlation between the zero-shot agent value and the verification accuracy of the neural network and the mutual information between the zero-shot agent value and the verification accuracy.

Further, the step S4 further includes:

S41. Use each selected zero-shot agent to calculate the agent value of each candidate FER network, and sort all agent values of the same zero-shot agent in ascending order;

S42. Accumulate the rankings of the same candidate FER network to obtain the cumulative value, and select the candidate facial expression recognition network with the largest cumulative value as the final facial expression recognition network.

further,

Step S32 further includes:

S321. Calculate the Spearman correlation coefficient based on all zero-shot agent values of each zero-shot agent in the same NAS benchmark data set and the verification accuracy of the corresponding neural network;

S322. Determine the relationship between the number of zero-shot agents whose Spearman correlation coefficients are always in the same direction and the preset number. If it is greater, go to S323, if it is less, go to S325, and if it is equal, go to S326;

S323. Perform all possible combinations of all zero-shot agents whose Spearman correlation coefficients are always in the same direction, and make the number of zero-shot agents in each group equal to the preset number;

S324. Calculate the mutual information of each combination in each NAS benchmark data set, calculate the average of all mutual information corresponding to each combination, and select the zero-shot agent combination corresponding to the maximum average;

S325. Delete the zero-shot agents with the smallest number of positive and negative values in the Spearman correlation coefficient, and make the number of selected zero-shot agents equal to the preset number;

S326. Select a zero-shot agent whose Spearman correlation coefficient is always in the same direction.

The beneficial effect of the above technical solution is: using zero-shot agent to evaluate the performance of neural network can greatly reduce the computational complexity and save computational costs. In multiple NAS benchmark data sets, the Spearman correlation coefficient is used to measure the ability of the zero-shot agent to predict the accuracy of the neural network to ensure that the selected zero-shot agent has generalization properties and can be directly transferred to other tasks, thereby ensuring that the zero-shot agent can be used in unknown tasks. The accuracy of neural networks can still be effectively predicted in NAS tasks.

Further, the calculation formula of the mutual information is:

Among them, A is the verification accuracy of the neural network; The agent values calculated for the corresponding 1st... m zero-shot agents respectively;/> is the information entropy corresponding to A;/> for/> The joint information entropy of; for/> The joint information entropy of ;/> Take/> for A The marginal probability distribution function of ; m is the dimension of the random variable;/> for/> Take/> joint probability distribution function; for/> Take/> joint probability distribution function.

The beneficial effect of the above technical solution is that mutual information can intuitively measure the ability of the zero-shot agent to capture accuracy-related features in the neural network, so as to facilitate a theoretical analysis of whether different zero-shot agents are suitable for combination. At the same time, by comparing the ability of multiple and single zero-shot agents to capture features, it can be theoretically demonstrated that there is indeed a complementary relationship between some zero-shot agents, and that combining multiple agents can indeed improve the accuracy of the prediction neural network.

Further, the selection method of the preset number includes:

A1. Perform q rounds of combinations on all selected zero-shot agents to generate all possible zero-shot agent combinations with a number equal to the d value. After one round of selection is completed, let d = d +1, 1≤ d ≤ q , q is the total number of zero-shot agents;

A2. Under each value d , calculate the mutual information of all its corresponding combinations in each NAS benchmark data set, and select the maximum value among all mutual information under the same NAS benchmark data set;

A3. Use the maximum mutual information among different d values corresponding to the same NAS benchmark data set to draw a curve, and select the corresponding d value when all curves tend to be stable as the preset number.

The beneficial effect of the above technical solution is that by calculating the growth trend of mutual information in the process of continuously increasing the number of zero-shot agent mixtures, the optimal number of combinations can be obtained. When the number of combinations increases and the mutual information value no longer changes, it means a waste of computing resources. Select the most appropriate number of combinations to ensure the accuracy and effectiveness of the prediction network while ensuring low computational complexity.

Further, the step S2 further includes:

S201. Each candidate facial expression recognition network includes N sequentially connected stage sequences, and each stage sequence includes l sequentially connected layers;

S202. Select the channel expansion rate of each stage sequence of the current candidate facial expression recognition network;

S203. Randomly select one of standard convolution, moving reverse bottleneck convolution, hybrid depth convolution and skip connection as the layer operation block.

S204. When the standard convolution is selected, randomly select a convolution kernel size from the three convolution kernels {3×3, 5×5, 7×7}, and then enter step S208;

S205. When the moving reverse bottleneck convolution is selected, randomly select a convolution kernel size from the three convolution kernels {3×3, 5×5, 7×7}, and then enter step S207;

When the hybrid depth convolution is selected, the number of convolution kernels of different scales to be mixed is randomly selected from three groups of {2, 3, 4}, and then step S207 is entered;

S206. When the skip connection is selected, enter step S208;

S207. Generate a random number in [0, 1]. When the random number is greater than or equal to 0.5, the attention module is added to the operation block. When the random number is less than 0.5, the attention module is not added, and then step S208 is entered;

S208. Determine whether all layers of each stage in the current candidate facial expression recognition network have been configured with operation blocks. If so, the generation of the current candidate facial expression recognition network is completed and step S209 is entered. Otherwise, return to step S203;

S209. Determine whether the parameter amount of the current candidate facial expression recognition network meets the preset parameter amount. If so, proceed to step S210. Otherwise, delete the current candidate facial expression recognition network and return to step S202;

S210. Determine whether the generated candidate facial expression recognition network is the same as the current candidate facial expression recognition network. If so, delete the current candidate facial expression recognition network and return to step S202. Otherwise, proceed to step S211;

S211. Determine whether the number of generated candidate facial expression recognition networks meets the preset number of networks. If so, complete the generation of candidate networks. Otherwise, return to step S202.

The beneficial effect of the above technical solution is: a hierarchical search space is designed for facial expression recognition tasks, allowing more flexible generation of FER networks with excellent performance. The hybrid depth convolution introduced in the search space can learn the features of facial expression images more comprehensively; the attention mechanism introduced can pay more attention to the key features in expression images, thereby effectively improving the recognition accuracy of the FER network. At the same time, in the process of generating candidate FER networks, a preset constraint is added to ensure that a FER network that meets the resource constraints of the target device is generated.

In a second aspect, a facial expression recognition neural network architecture search device is provided, which includes:

Search space building module, used to build a search space for facial expression recognition;

The candidate network generation module is used to use random search to generate several facial expression recognition candidate networks that meet the preset parameters;

The zero-shot agent selection module is used to select a preset number of zero-shot agents that meet the preset conditions in the NAS data set;

Facial expression recognition network performance evaluation module, used to use the selected zero-shot agent to select the best candidate network from several facial expression recognition candidate networks as the final facial expression recognition network;

The facial expression recognition module is used to select a facial expression image data set to train the facial expression recognition neural network. After training, it is used to recognize the facial expression image to be recognized.

In the third aspect, an application method of a facial expression recognition network is provided. The facial expression recognition network is searched using a facial expression recognition neural network architecture search method, and is characterized by including:

Select a facial expression image data set to train the facial expression recognition network;

Obtain a facial expression image to be recognized and preprocess the facial expression image;

Input the preprocessed facial expression image into the trained facial expression recognition network for recognition, and obtain the recognition result.

Further, the expression images in the facial expression image data set include seven emotions: anger, disgust, fear, happiness, sadness, surprise and nature.

Furthermore, a stochastic gradient descent optimizer was used to train the facial expression recognition neural network. During training, the initial learning rate was set to 0.01 and trained for 300 epochs.

The beneficial effects of the present invention are: 1. The hybrid zero-shot agent used in this solution can predict the accuracy of the network in the initialization stage, which greatly reduces the calculation time of the design process. In this way, according to the resource conditions of different mobile terminals, we can quickly design a mobile terminal that occupies a smaller area. The FER network of resources makes it possible to promote and apply the FER network in different mobile terminals.

2. This program designs a new search space for the field of facial expression recognition. By increasing the degree of freedom in model construction, it enhances the diversity of the FER model, thereby designing a FER model with better performance. A search space with higher degrees of freedom can achieve a better balance between model complexity and accuracy.

3. This solution uses an attention module in the search space, so that the FER model can increase the weight of important features based on global information and pay more attention to expression features that are important for recognition accuracy, thereby further enhancing the recognition accuracy of the FER network while only adding very little The amount of parameters makes it possible for the FER network with higher recognition accuracy and smaller memory footprint to be deployed on mobile terminals with different computing resources.

Description of the drawings

Figure 1 is a flow chart of the neural network architecture search method for facial expression recognition.

Figure 2 is a schematic diagram of the framework of the search space.

Figure 3 is a schematic diagram of the curves plotted at different values of d for each NAS benchmark data set.

Detailed ways

The specific embodiments of the present invention are described below to facilitate those skilled in the art to understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the technical field, as long as various changes These changes are obvious within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions and creations utilizing the concept of the invention are protected.

Referring to Figure 1, Figure 1 shows a flow chart of a neural network architecture search method for facial expression recognition; as shown in Figure 1, the method S includes steps S1 to S4.

In step S1, a search space for facial expression recognition is constructed;

In step S2, a random search method is used to generate several facial expression recognition candidate networks that meet the preset parameter amounts;

In one embodiment of the present invention, step S2 further includes:

S201. Each candidate facial expression recognition network includes N sequentially connected stage sequences, and each stage sequence includes l sequentially connected layers;

S202. Select the channel expansion rate of each stage sequence of the current candidate facial expression recognition network;

The channel expansion rate of this scheme can take values {1, 4, 6}, which are randomly selected each time. After the channel expansion rate is introduced in this scheme, the nonlinear expression ability can be increased by increasing the number of output channels of the separated convolution.

S203. Randomly select one of standard convolution, moving reverse bottleneck convolution, hybrid depth convolution and skip connection as the layer operation block.

S204. When the standard convolution is selected, randomly select a convolution kernel size from the three convolution kernels {3×3, 5×5, 7×7}, and then enter step S208;

S205. When the moving reverse bottleneck convolution is selected, randomly select a convolution kernel size from the three convolution kernels {3×3, 5×5, 7×7}, and then enter step S207;

When the hybrid depth convolution is selected, the number of convolution kernels of different scales to be mixed is randomly selected from three groups of {2, 3, 4}, and then step S207 is entered;

S206. When the skip connection is selected, enter step S208;

S207. Generate a random number in [0, 1]. When the random number is greater than or equal to 0.5, the attention module is added to the operation block. When the random number is less than 0.5, the attention module is not added, and then step S208 is entered;

S208. Determine whether all layers of each stage in the current candidate facial expression recognition network have been configured with operation blocks. If so, the generation of the current candidate facial expression recognition network is completed and step S209 is entered. Otherwise, return to step S203;

S209. Determine whether the parameter amount of the current candidate facial expression recognition network meets the preset parameter amount. If so, proceed to step S210. Otherwise, delete the current candidate facial expression recognition network and return to step S202;

S210. Determine whether the generated candidate facial expression recognition network is the same as the current candidate facial expression recognition network. If so, delete the current candidate facial expression recognition network and return to step S202. Otherwise, proceed to step S211;

S211. Determine whether the number of generated candidate facial expression recognition networks meets the preset number of networks. If so, complete the generation of candidate networks. Otherwise, return to step S202.

Since facial expression images are usually obtained under different lighting, occlusions and different postures, it greatly affects the ability of the FER model to extract expression features. The hybrid depth convolution introduced in this solution can extract features of different scales to learn Subtle differences in facial expressions in order to improve the model's expression recognition accuracy.

The skip connection (skip operation) introduced in this solution can directly output the input feature map without any calculation, allowing the depth of the FER network architecture to be automatically adjusted, thereby designing more diverse FER models. The introduction of the attention module SEblock can use global information to selectively increase the weight of important facial expression features, allowing the model to pay more attention to learning features useful for expression recognition, thereby increasing recognition accuracy and reducing the effects of light, posture and occlusion. coming influence.

In step S3, a preset number of zero-shot agents that meet the preset conditions are selected from the NAS data set:

S31. Select several zero-shot proxies, randomly sample several neural networks in multiple NAS benchmark data sets, and use zero-shot proxies to calculate the accuracy-based proxy value of each neural network;

This solution randomly samples 1000 neural networks from each NAS benchmark data set for calculation. The benchmark data sets used are NAS-Bench-101, NAS-Bench-201 and Network Design Space (NDS), except for NAS-Bench- 201 is obtained by training on three data sets: CIFAR10, CIFAR100 and ImageNet16-120. The others are only trained on CIFAR10. That is, this solution uses 9 NAS benchmark data sets and randomly samples 1,000 in each NAS benchmark data set. Neural networks are used for zero-shot agent selection.

S32. Select a preset number of zero-shot agents based on the correlation between the zero-shot agent value and the verification accuracy of the neural network and the mutual information between the zero-shot agent value and the verification accuracy.

The zero-shot proxy method is a performance evaluation method that does not require training. It predicts the relative accuracy of a neural network from forward or backward propagation of a single batch of data based on a network characteristic related to network accuracy. Accuracy assessment of individual network architectures is completed in seconds, accelerating neural network searches for facial expression recognition.

During implementation, the preferred selection methods for the preset quantity of this plan include:

A1. Perform q rounds of combinations on all selected zero-shot agents to generate all possible zero-shot agent combinations with a number equal to the d value. After one round of selection is completed, let d = d +1, 1≤ d ≤ q , q is the total number of zero-shot agents;

In step A1, assuming a total of 12 different zero-shot agents are selected, then 12 rounds of combinations are needed. The number of zero-shot agent combinations obtained in each round is ,/> ,/> ,…/> …/> ,/> in each round The number of zero-shot agents in each combination is d .

A2. Under each value d , calculate the mutual information of all its corresponding combinations in each NAS benchmark data set, and select the maximum value among all mutual information under the same NAS benchmark data set;

A3. Use the maximum mutual information among different d values corresponding to the same NAS benchmark data set to draw a curve, and select the corresponding d value when all curves tend to be stable as the preset number. The curve drawn can be referred to Figure 3. It can be seen from Figure 3 that when the value d is greater than 4, the curve becomes stable, and the value of the preset quantity is 4 at this time.

In one embodiment of the present invention, step S32 further includes:

S321. Calculate the Spearman correlation coefficient based on all zero-shot agent values of each zero-shot agent in the same NAS benchmark data set and the verification accuracy of the corresponding neural network;

Assume that this solution selects 12 zero-shot agents (refer to the 12 zero-shot agents shown in the first horizontal row of Table 1), combined with the 9 NAS benchmark data sets mentioned above and 1000 randomly sampled neural Network, the final calculated Spearman correlation coefficient can be referred to Table 1.

Table 1 Calculated correlation coefficients on different NAS benchmark data sets

In Table 1, the higher the absolute value of the Spearman correlation coefficient, the more reliable the prediction accuracy is.

S322. Determine the relationship between the number of zero-shot agents whose Spearman correlation coefficients are always in the same direction and the preset number. If it is greater, go to S323, if it is less, go to S325, and if it is equal, go to S326;

S323. Perform all possible combinations of all zero-shot agents whose Spearman correlation coefficients are always in the same direction, and make the number of zero-shot agents in each group equal to the preset number;

S324. Calculate the mutual information of each combination in each NAS benchmark data set, calculate the average of all mutual information corresponding to each combination, and select the zero-shot agent combination corresponding to the maximum average;

S325. Delete the zero-shot agents with the smallest number of positive and negative values in the Spearman correlation coefficient, and make the number of selected zero-shot agents equal to the preset number;

S326. Select a zero-shot agent whose Spearman correlation coefficient is always in the same direction.

In step S4, the selected zero-shot agent is used to select the optimal candidate network from several facial expression recognition candidate networks as the final facial expression recognition network:

S41. Use each selected zero-shot agent to calculate the agent value of each candidate FER network, and sort all agent values of the same zero-shot agent in ascending order;

S42. Accumulate the rankings of the same candidate FER network to obtain the cumulative value, and select the candidate facial expression recognition network with the largest cumulative value as the final facial expression recognition network.

When implemented, the calculation formula of the mutual information is:

Among them, A is the verification accuracy of the neural network;/> The agent values calculated for the corresponding 1st... m zero-shot agents respectively; is the information entropy corresponding to A;/> for/> The joint information entropy of ;/> for The joint information entropy of ;/> Take/> for A The marginal probability distribution function of ; m is the dimension of the random variable; for/> Take/> joint probability distribution function;/> for/> Take/> joint probability distribution function.

In a second aspect, a facial expression recognition neural network architecture search device is provided, which includes:

Search space building module, used to build a search space for facial expression recognition;

The candidate network generation module is used to use random search to generate several facial expression recognition candidate networks that meet the preset parameters;

The zero-shot agent selection module is used to select a preset number of zero-shot agents that meet the preset conditions in the NAS data set;

Facial expression recognition network performance evaluation module, used to use the selected zero-shot agent to select the best candidate network from several facial expression recognition candidate networks as the final facial expression recognition network;

The facial expression recognition module is used to select a facial expression image data set to train the facial expression recognition neural network. After training, it is used to recognize the facial expression image to be recognized.

In the third aspect, this solution also provides an application method of facial expression recognition network. The facial expression recognition network is searched using the facial expression recognition neural network architecture search method. It is characterized by:

Select a facial expression image data set to train the facial expression recognition network;

Obtain a facial expression image to be recognized and preprocess the facial expression image;

Input the preprocessed facial expression image into the trained facial expression recognition network for recognition, and obtain the recognition result.

Among them, the expression images in the facial expression image data set cover at least seven emotions: anger, disgust, fear, happiness, sadness, surprise and natural; a stochastic gradient descent optimizer is used to train the facial expression recognition neural network. During training, the initial learning is set The rate is 0.01 and the training is for 300 epochs.

The following is a verification of the lightweight facial expression recognition network capabilities designed in this solution:

In order to verify the performance of the neural architecture search algorithm that integrates the zero-shot agent as a performance evaluation strategy to automatically design a lightweight FER model, this program conducts evaluation work on the FER2013 dataset, which is widely used in the field of facial expression recognition, and will predict based on the zero-shot agent. The FER model with the best performance is compared with the existing latest hand-designed lightweight FER model and the automatically designed lightweight FER model.

A. Dataset construction

This solution selects FER2013, a challenging data set in the field of facial expression recognition. This data set is an expression image obtained in a real environment, including seven emotions including anger, disgust, fear, happiness, sadness, surprise and nature.

B． Task evaluation index

For designing lightweight facial expression recognition networks, model complexity and recognition accuracy are used as evaluation indicators. In order to facilitate calculation and comparison here, the parameter amount is used as the evaluation index of model complexity, in megabytes (M) as the unit.

C． Algorithm parameter settings

For example, the Kaiming initialization method commonly used in the zero-shot proxy method initializes the model parameters before calculating the proxy value of the FER model. This plan sets a large enough sampling number of 10,000 to make the random search FER model close to the optimal solution. In order to further illustrate that the effectiveness of the method of this scheme is not accidental, this scheme ran the algorithm 10 times and used different random seeds. In addition, considering that the parameter amount of the FER model designed by the existing NAS method is around 3M, this solution sets the constraint that the parameter amount is less than 3M.

To verify the accuracy of the optimal model predicted by training, use the stochastic gradient descent optimizer, set the initial learning rate to 0.01, and train for 300 epochs.

D. Results comparison and analysis

Comparison of results: This solution (NAS based on hybrid zero-shot proxy, HZS-NAS), in the data set FER2013, compared with the manually designed lightweight model BKVGG12, SHCNN, Light-CNN, eXnet, MBCC-CNN and automatically designed The lightweight models MNAS-based and Auto-FERNet are compared. The comparison results are shown in Table 2.

Table 2 Experimental results on the data set FER2013

Result analysis: On the data set FER2013, in terms of the two evaluation indicators, the recognition accuracy has been significantly improved, and the number of parameters has been reduced. Compared with Light-CNN and Auto-FERNet, although the number of parameters increased by 1.32M and 0.33M respectively, the recognition accuracy was also improved by 6.02% and 0.91%.

In addition, this solution only consumes 10 GPU hours. Compared with the two automatic design methods, it greatly reduces the computational overhead. The Auto-FERNet method requires 46 GPU hours to implement. Although the MNAS-based method does not report the calculation time , but on the ImageNet data set, the MNAS algorithm consumed a total of 40,000 GPU hours, which still requires a lot of computing overhead even in FER2013.

From the above comparative analysis, it can be seen that the search method of this solution greatly shortens the search time, and the number of parameters is relatively small. In this way, according to the resource conditions of different mobile terminals, a FER network that occupies smaller resources corresponding to the mobile terminal can be quickly designed. So that the FER network is expected to be popularized on different smart terminals.

Claims (10)

1. Facial expression recognition neural network search method, which is characterized by including the following steps: S1. Construct a search space for facial expression recognition; S2. Use random search to generate several facial expression recognition candidate networks that meet the preset parameters; S3. Select a preset number of zero-shot agents that meet the preset conditions in the NAS benchmark data set; S4. Use the selected zero-shot agent to select the optimal candidate network from several facial expression recognition candidate networks as the final facial expression recognition network. 2. The facial expression recognition neural network search method according to claim 1, characterized in that the step S3 further includes: S31. Select several zero-shot proxies, randomly sample several neural networks in multiple NAS benchmark data sets, and use zero-shot proxies to calculate the accuracy-based proxy value of each neural network; S32. Select a preset number of zero-shot agents based on the correlation between the zero-shot agent value and the verification accuracy of the neural network and the mutual information between the zero-shot agent value and the verification accuracy. 3. The facial expression recognition neural network search method according to claim 1, characterized in that the step S4 further includes: S41. Use each selected zero-shot agent to calculate the agent value of each candidate FER network, and sort all agent values of the same zero-shot agent in ascending order; S42. Accumulate the rankings of the same candidate FER network to obtain the cumulative value, and select the candidate facial expression recognition network with the largest cumulative value as the final facial expression recognition network. 4. The facial expression recognition neural network architecture search method according to claim 2, characterized in that step S32 further includes: S321. Calculate the Spearman correlation coefficient based on all zero-shot agent values of each zero-shot agent in the same NAS benchmark data set and the verification accuracy of the corresponding neural network; S322. Determine the relationship between the number of zero-shot agents whose Spearman correlation coefficients are always in the same direction and the preset number. If it is greater, go to S323, if it is less, go to S325, and if it is equal, go to S326; S323. Perform all possible combinations of all zero-shot agents whose Spearman correlation coefficients are always in the same direction, and make the number of zero-shot agents in each group equal to the preset number; S324. Calculate the mutual information of each combination in each NAS benchmark data set, calculate the average of all mutual information corresponding to each combination, and select the zero-shot agent combination corresponding to the maximum average; S325. Delete the zero-shot agents with the smallest number of positive and negative values in the Spearman correlation coefficient, and make the number of selected zero-shot agents equal to the preset number; S326. Select a zero-shot agent whose Spearman correlation coefficient is always in the same direction. 5. The facial expression recognition neural network architecture search method according to claim 2 or 4, characterized in that the calculation formula of the mutual information is: I(Z ₁ ,…, _m ;)=(A)+(Z ₁ ,…, _m )-(A,Z ₁ ,…, _m ) Among them, A is the verification accuracy of the neural network; Z ₁ ,..., _m are respectively the agent values calculated by the corresponding 1...mth zero-shot agent; H(A) is the information entropy corresponding to A; H(Z ₁ ,…, _m ) is the joint information entropy of Z ₁ ,…, _m ; H(A,Z ₁ ,…, _m ) is the joint information entropy of A, Z ₁ ,…, _m ; (a _i ) is the joint information entropy of A The marginal probability distribution function of a _i ; m is the dimension of the random variable; P(z ₁ ,...z _m ) is the joint probability distribution function of Z ₁ ,..., _m takes z ₁ ,...z _m ; P(a,z ₁ ,...z _m ) is the joint probability distribution function of a,z ₁ ,... _{z m} for A,Z ₁ ,...,m. 6. The facial expression recognition neural network architecture search method according to claim 1, characterized in that the selection method of the preset number includes: A1. Perform q rounds of combinations on all selected zero-shot agents to generate all possible zero-shot agent combinations with a number equal to the value of d. After one round of selection is completed, let d=d+1, 1≤d≤q, q is the total number of zero-shot agents; A2. Under each value d, calculate the mutual information of all its corresponding combinations in each NAS benchmark data set, and select the maximum value among all mutual information under the same NAS benchmark data set; A3. Use the maximum mutual information among different d values corresponding to the same NAS benchmark data set to draw a curve, and select the corresponding d value when all curves tend to be stable as the preset number. 7. The facial expression recognition neural network architecture search method according to any one of claims 1-4 and 6, characterized in that the step S2 further includes: S201. Each candidate facial expression recognition network includes N sequentially connected stage sequences, and each stage sequence includes l sequentially connected layers; S202. Select the channel expansion rate of each stage sequence of the current candidate facial expression recognition network; S203. Randomly select one of standard convolution, moving reverse bottleneck convolution, hybrid depth convolution and skip connection as the layer operation block. S204. When the standard convolution is selected, randomly select a convolution kernel size from the three convolution kernels {3×3, 5×5, 7×7}, and then enter step S208; S205. When the moving reverse bottleneck convolution is selected, randomly select a convolution kernel size from the three convolution kernels {3×3, 5×5, 7×7}, and then enter step S207; When the hybrid depth convolution is selected, the number of convolution kernels of different scales to be mixed is randomly selected from three groups of {2, 3, 4}, and then step S207 is entered; S206. When the skip connection is selected, enter step S208; S207. Generate a random number in [0, 1]. When the random number is greater than or equal to 0.5, the attention module is added to the operation block. When the random number is less than 0.5, the attention module is not added, and then step S208 is entered; S208. Determine whether all layers of each stage in the current candidate facial expression recognition network have been configured with operation blocks. If so, the generation of the current candidate facial expression recognition network is completed and step S209 is entered. Otherwise, return to step S203; S209. Determine whether the parameter amount of the current candidate facial expression recognition network meets the preset parameter amount. If so, proceed to step S210. Otherwise, delete the current candidate facial expression recognition network and return to step S202; S210. Determine whether the generated candidate facial expression recognition network is the same as the current candidate facial expression recognition network. If so, delete the current candidate facial expression recognition network and return to step S202. Otherwise, proceed to step S211; S211. Determine whether the number of generated candidate facial expression recognition networks meets the preset number of networks. If so, complete the generation of candidate networks. Otherwise, return to step S202. 8. A facial expression recognition neural network architecture search device, characterized by including: Search space building module, used to build a search space for facial expression recognition; The candidate network generation module is used to use random search to generate several facial expression recognition candidate networks that meet the preset parameters; The zero-shot agent selection module is used to select a preset number of zero-shot agents that meet the preset conditions in the NAS benchmark data set; Facial expression recognition network performance evaluation module, used to use the selected zero-shot agent to select the best candidate network from several facial expression recognition candidate networks as the final facial expression recognition network; The facial expression recognition module is used to select a facial expression image data set to train the facial expression recognition neural network. After training, it is used to recognize the facial expression image to be recognized. 9. Application method of facial expression recognition network. The facial expression recognition network is searched using the facial expression recognition neural network architecture search method described in any one of claims 1 to 7, and is characterized by including: Select a facial expression image data set to train the facial expression recognition network; Obtain a facial expression image to be recognized and preprocess the facial expression image; Input the preprocessed facial expression image into the trained facial expression recognition network for recognition, and obtain the recognition result. 10. The application method of the facial expression recognition network according to claim 9, characterized in that the expression images in the facial expression image data set include seven emotions: anger, disgust, fear, happiness, sadness, surprise and natural.