A method, device and system for assessing concentration for amblyopia training

Abstract

The present invention discloses a method, device and system for evaluating concentration for amblyopia training, which obtains first sample data by expanding the current training data of the user to be evaluated through a generative adversarial network, extracts features from the sample data through iterations of several vortex parameter sets after performance sorting, and adjusts the optimal vortex parameters and the worst vortex attenuation of the parameter set until the first feature data is output, then inputs the first feature data into the encoder network of the feature dimensionality reduction model for encoding processing, and iteratively adjusts the encoder parameters according to the reverse reinforcement learning mechanism, and adjusts the encoding process according to the decoder performance feedback until the first dimensionality reduction data is output, and then evaluates the dimensionality reduction data through a classifier to output an evaluation result. By expanding the sample data of real-time data, the quality and diversity of the sample data are improved; by using the vortex optimization algorithm that simulates the vortex phenomenon, the dependence of the model accuracy on the gradient is reduced.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a concentration assessment method, device and system for amblyopia training. Background Art

[0002] With the rapid development of science and technology and the accelerated pace of people's lives, visual health issues are receiving increasing attention. Amblyopia, as a common visual impairment, affects the quality of life of many people. Amblyopia usually occurs during the critical period of children's growth. If effective training and treatment are not carried out, it may lead to permanent vision loss. Therefore, the development of effective amblyopia training methods and evaluation methods is of great significance for the prevention and treatment of amblyopia. In the treatment process of amblyopia, the assessment of concentration is a key link. The effect of training depends largely on the concentration of participants during the training process. Participants with high concentration tend to benefit better from the training and thus improve their vision more effectively.

[0003] However, existing methods for assessing concentration have certain limitations, such as single assessment means, incomplete data collection and processing, need to improve model training efficiency and stability, and inaccurate feature extraction and dimensionality reduction methods. These problems not only affect the accuracy and efficiency of the assessment, but also limit the further optimization and application of amblyopia training methods. Summary of the invention

[0004] The present invention provides a concentration assessment method, device and system for amblyopia training, so as to improve the accuracy and efficiency of concentration assessment, thereby achieving the technical effect of further optimizing the amblyopia training method.

[0005] In order to solve the above technical problems, an embodiment of the present invention provides a method for evaluating concentration for amblyopia training, comprising the following steps:

[0006] Acquire current training data of the amblyopic user to be evaluated, and input the current training data into a preset generative adversarial network for sample data expansion, and output first sample data;

[0007] Inputting the first sample data into a feature extraction model, iteratively extracting features from the first sample data through a plurality of vortex parameter sets ranked by performance, and performing optimal vortex parameter adjustment and worst vortex attenuation on the plurality of vortex parameter sets until the iteration is completed and the first feature data is output;

[0008] The first feature data is input into the encoder of the feature dimensionality reduction model for encoding processing, and the parameters of the encoder are iteratively adjusted according to the inverse reinforcement learning mechanism. At the same time, the encoding process is adjusted according to the performance feedback of the decoder until the iteration is completed and the first dimensionality reduction data is output. The first dimensionality reduction data is input into the classifier for concentration evaluation, and the evaluation result is output.

[0009] The concentration evaluation method provided by the present invention obtains the current training data of the user to be evaluated, that is, the real-time training data of the amblyopic user, and inputs it into a generative adversarial network for sample data expansion, and uses the expanded data as the first sample data for subsequent concentration evaluation. By performing sample data expansion on the collected real-time data, insufficient sample data can be used to avoid insufficient generalization ability of the system evaluation model, thereby affecting the model evaluation accuracy. At the same time, the quality and diversity of the sample data and the stability of the model are improved.

[0010] After obtaining the first sample data, the system inputs the sample data into the feature extraction model, and extracts features from the sample data in turn using several vortex parameter sets ranked by performance in the model. During the feature extraction process, the optimal vortex in the above vortex parameter sets is adjusted and the worst vortex is attenuated, and the feature extraction and vortex adjustment process is continuously iterated until the iteration is completed and the corresponding first feature data is output. The system optimizes the parameters in the feature extraction model by simulating the vortex optimization algorithm of the vortex phenomenon, simulating the formation, evolution and dissipation process of the vortex, reducing the model accuracy's dependence on the gradient, solving the problems of gradient vanishing, gradient explosion and search-oriented local optimal solution that may exist in the model in the prior art, and improving the training efficiency and stability of the model.

[0011] In order to further improve the precision and accuracy of model evaluation, after obtaining the first feature data, the system also inputs the first feature data into the encoder of the feature dimensionality reduction model for encoding processing, and iteratively adjusts the encoder parameters through the inverse reinforcement learning mechanism during the encoding process, further optimizing the feature extraction of the encoding process. At the same time, the encoding process is adjusted according to the performance feedback of the decoder, and the encoder is adjusted through an interactive optimization mechanism to ensure that the decoded data retains the information in the feature data that is most relevant to the concentration evaluation as much as possible, that is, the key information for the concentration evaluation, further improving the accuracy and efficiency of the concentration evaluation model. After completing the feature dimensionality reduction, the system inputs the reduced dimensionality data into the classifier for evaluation, and then outputs the concentration evaluation result.

[0012] As a preferred example, the current training data is input into a preset generative adversarial network for sample data expansion, and the first sample data is output, specifically:

[0013] Input the current training data into the generator network in the generative adversarial network for forward propagation, and output generated sample data and the current training data;

[0014] Input the current training data and the generated sample data into the discriminator network for sample discrimination, output and calculate network loss data and stability loss data according to the discrimination result data, and then adjust the generator network parameters according to the network loss data and the stability loss data;

[0015] At the same time, the network loss data and the stability loss data are evaluated, and then the learning strategies and parameters of the generator network and the discriminator network are adjusted according to the evaluation results, and the discrimination process is iteratively generated until the iteration is completed to output the first sample data.

[0016] In order to improve the comprehensiveness and accuracy of the model's assessment of concentration, when the system collects the real-time training data of the person to be assessed, it will also perform data expansion on the real-time data through a generative adversarial network, so as to improve the quality and richness of the sample data. The system of the present invention inputs the generated sample data output by the generator network and the current training data as input data into the discriminator network for discrimination, and adjusts the network parameters of the generator network according to the discrimination result of the discriminator, so as to improve the quality of the generated samples in a competitive learning manner.

[0017] At the same time, compared with the existing generative adversarial network, the loss function used by the generative adversarial network provided by the system of the present invention includes not only the network loss function but also the stability loss function. The system evaluates the stability of the generated samples through the stable loss function to ensure that the data generated during the training process is more diverse and closer to the actual distribution of the training data. The generator network is adjusted through the feedback of the discriminator network, and the cycle consistency check and adaptive adjustment mechanism are used to guide the generator network to generate data closer to the real sample, thereby improving the stability, generalization and accuracy of the model evaluation.

[0018] As a preferred example, the iterative extraction of features from the first sample data by the several vortex parameter sets ranked by performance, and the optimal vortex parameter adjustment and the worst vortex attenuation of the several vortex parameter sets are performed until the iteration is completed and the first feature data is output, specifically including:

[0019] Performing initial feature extraction on the first sample data, obtaining and outputting initial feature data, retrieving and performing performance evaluation calculation on the initial feature data through the performance loss function of the plurality of vortex parameter sets, and then sorting the plurality of vortex parameter sets according to the calculated loss function values;

[0020] Adjusting the positions and intensities of the plurality of vortex parameter sets according to the corresponding rankings of the plurality of vortex parameter sets after sorting and the distance between every two parameter sets, and performing merging operations and splitting operations on the plurality of vortex parameter sets according to preset merging thresholds and splitting thresholds;

[0021] The performance of the several vortex parameter sets is evaluated according to a preset performance algorithm, and the parameters of the optimal vortex determined by the evaluation are adjusted through local search. At the same time, the worst vortex determined by the evaluation is attenuated through life value adjustment, and the feature extraction process is iterated until the iteration is completed and the first feature data is output.

[0022] In order to solve the problem of model accuracy being too dependent on gradients, the system will optimize the parameters in the neural network by simulating the formation, evolution, and dissipation of vortices when extracting features from sample data, and provides a vortex-based feature extraction method. In fluid dynamics, vortices can increase the chaos and complexity of the fluid, thereby guiding the search process in the parameter space toward a better solution. By simulating this natural process, the vortex optimization neural network can find the parameter configuration that can minimize the neural network loss function, thereby reducing the dependence of model parameter optimization on model gradients.

[0023] As a preferred example, the iterative adjustment of the encoder parameters according to the inverse reinforcement learning mechanism and the adjustment of the encoding process according to the performance feedback of the decoder until the iterative output of the first dimension reduction data is completed specifically include:

[0024] Encoding the first feature data, and calling a corresponding reward function according to the correlation between the first feature data and the concentration evaluation index, adjusting the parameters of the encoder by using a reverse learning mechanism during the encoding process, and outputting the first encoded data;

[0025] The first encoded data is input into a decoder so that the decoder performs feature reconstruction based on the first encoded data and outputs first reconstructed data. The correlation between the first feature data and the first reconstructed data is then used as performance feedback of the decoder to adjust the encoding process, and the encoding and decoding process is iterated until the iteration is completed and the first reduced dimensionality data is output.

[0026] In order to further optimize the extracted feature data, when the system performs feature dimensionality reduction processing on the feature data, compared with the traditional autoencoder that realizes data dimensionality reduction and reconstruction through encoding and decoding processes, the system provided by the present invention adopts an inverse reinforcement learning mechanism to further optimize the feature extraction in the encoding process. The reward function called is related to the feature data and the concentration evaluation index, so that it pays more attention to the features in the amblyopia training data that are closely related to the concentration evaluation.

[0027] At the same time, the encoding process of the encoder is adjusted according to the correlation between the reconstructed data and the feature data output by the decoder as performance feedback, and the encoder is adjusted through an interactive optimization mechanism to ensure that the decoded data retains the most critical information for concentration assessment in the original data as much as possible, further improving the accuracy and evaluation precision of the output reduced dimensionality data for concentration assessment.

[0028] As a preferred example, the first dimension reduction data is input into a classifier to perform concentration evaluation, and the evaluation result is output, specifically:

[0029] Initializing the twin network of the classifier according to the feature dimension of the first dimension reduction data, and setting initial parameters of the submodular function optimization algorithm;

[0030] The first dimension reduction data is divided into a positive sample set and a negative sample set, and input into two support vector machines of the twin network for model training, and the submodular function optimization algorithm is used to iteratively update the parameters of the support vector machine and the weight of the loss function until convergence;

[0031] After convergence, the classifier output data is output as the evaluation result.

[0032] In order to improve the evaluation accuracy, when the system uses a classifier to perform classification and evaluation on the reduced-dimensional data, compared with the traditional support vector machine algorithm when processing large-scale nonlinear separable data, it may encounter optimization difficulties. The system provided by the present invention adopts a submodular function optimization strategy, which can improve the convergence speed of the algorithm while ensuring that a more global optimal solution is found.

[0033] At the same time, since the system of the present invention adopts a twin network structure, which includes two parallel support vector machines, it can learn the differences between features more accurately, thereby improving classification accuracy and further improving evaluation accuracy.

[0034] Accordingly, an embodiment of the present invention further provides a concentration assessment device for amblyopia training, the concentration assessment device comprising a sample collection module, a feature extraction module, a feature dimension reduction module and a result output module;

[0035] The sample acquisition module is used to obtain current training data of the amblyopic user to be evaluated, and input the current training data into a preset generative adversarial network to perform sample data expansion, and output first sample data;

[0036] The feature extraction module is used for inputting the first sample data into the feature extraction model, iteratively extracting features from the first feature data through a plurality of vortex parameter sets ranked by performance, and performing optimal vortex parameter adjustment and worst vortex attenuation on the plurality of vortex parameter sets, until the iteration is completed and the first feature data is output;

[0037] The feature dimension reduction module is used to input the first feature data into the encoder of the feature dimension reduction model for encoding processing, and iteratively adjust the parameters of the encoder according to the inverse reinforcement learning mechanism, and adjust the encoding process according to the performance feedback of the decoder until the iteration is completed and the first dimension reduction data is output;

[0038] The result output module is used to input the first dimension reduction data into a classifier to perform concentration evaluation and output the evaluation result.

[0039] As a preferred example, the sample acquisition module inputs the current training data into a preset generative adversarial network to perform sample data expansion, and outputs first sample data, specifically:

[0040] Input the current training data into the generator network in the generative adversarial network for forward propagation, and output generated sample data and the current training data;

[0041] Input the current training data and the generated sample data into the discriminator network for sample discrimination, output and calculate network loss data and stability loss data according to the discrimination result data, and then adjust the generator network parameters according to the network loss data and the stability loss data;

[0042] At the same time, the network loss data and the stability loss data are evaluated, and then the learning strategies and parameters of the generator network and the discriminator network are adjusted according to the evaluation results, and the discrimination process is iteratively generated until the iteration is completed to output the first sample data.

[0043] As a preferred example, the result output module inputs the first dimension reduction data into a classifier to perform concentration evaluation, and outputs an evaluation result, specifically:

[0044] Initializing the twin network of the classifier according to the feature dimension of the first dimension reduction data, and setting initial parameters of the submodular function optimization algorithm;

[0045] The first dimension reduction data is divided into a positive sample set and a negative sample set, and input into two support vector machines of the twin network for model training, and the submodular function optimization algorithm is used to iteratively update the parameters of the support vector machine and the weight of the loss function until convergence;

[0046] After convergence, the classifier output data is output as the evaluation result.

[0047] Correspondingly, an embodiment of the present invention also provides a concentration assessment system for amblyopia training, the concentration assessment system comprising a memory, a processor, and a computer program stored in the memory and running on the processor, and the processor implements any of the above-mentioned concentration assessment methods for amblyopia training when processing the computer program.

[0048] Correspondingly, an embodiment of the present invention further provides a storage medium, on which a computer program is stored. The computer program is called and executed by a processor to implement a concentration assessment method for amblyopia training as described in any one of the above items. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 : A flow chart of an embodiment of a method for assessing concentration for amblyopia training provided by the present invention;

[0050] Figure 2 : A structural schematic diagram of an embodiment of a concentration assessment device for amblyopia training provided by the present invention;

[0051] Figure 3 : A schematic flow chart of an embodiment of a vortex optimization algorithm for simulating vortex phenomena provided by the present invention. DETAILED DESCRIPTION

[0052] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0053] Embodiment 1

[0054] Please refer to Figure 1 , is a flow chart of an embodiment of a method for assessing concentration for amblyopia training provided by the present invention, including steps 101 to 103, each of which is specifically as follows:

[0055] Step 101: obtaining current training data of a visually impaired user to be evaluated, and inputting the current training data into a preset generative adversarial network for sample data expansion, and outputting first sample data.

[0056] The concentration assessment method provided by an embodiment of the present invention obtains the current training data of the user to be evaluated, that is, the real-time training data of the amblyopic user, and inputs it into a generative adversarial network for sample data expansion, and uses the expanded data as the first sample data for subsequent concentration assessment. By performing sample data expansion on the collected real-time data, insufficient sample data can be used to avoid insufficient generalization ability of the system evaluation model, thereby affecting the model evaluation accuracy. At the same time, the quality and diversity of the sample data and the stability of the model are improved.

[0057] In this embodiment, the real-time training data collected by the system comes from participants undergoing amblyopia training, i.e., the subjects to be evaluated. The subjects undergo training by wearing eye tracking devices and physiological signal monitoring devices (such as heart rate monitors and skin electrical activity sensors), and the system collects data through the above devices. The data format includes time series data and discrete event markers, which are stored in vector form. The attributes of the real-time training data collected in this embodiment specifically include:

[0058] Eye movement data (a x ): Eye tracking equipment was used to capture participants’ eye movements and record eye position, movement speed, and pupil size.

[0059] Heart rate data (a y ): Heart rate variability was collected using a heart rate belt to assess participants’ emotional state and concentration.

[0060] Electrodermal activity (a z ): The participants' emotional arousal levels were recorded through galvanic skin sensors, reflecting the physiological changes during concentration.

[0061] a x1 : Eye movement speed.

[0062] a x2 : Pupil size.

[0063] a x3 : Number of eye closures.

[0064] a y1 : Heart rate.

[0065] a y2 : Heart rate variability.

[0066] a z1 : Galvanic skin response level.

[0067] a v : Response time to visual stimuli in the training task.

[0068] a a : Response time to audio stimulus.

[0069] a s : Number of successes of the participant on the training task.

[0070] a f : Number of failures of the participant on the training task.

[0071] The above data attributes are only the data attributes illustrated in this embodiment. This embodiment does not further limit the specific types and quantities of the data attributes. During the specific implementation of the concentration assessment method provided in this embodiment, that is, in the actual application of this embodiment, the attributes of the data are usually more than 10 attributes, and may reach dozens or even hundreds.

[0072] In addition, when the system trains the evaluation model, it will also annotate the collected data manually. In this embodiment, the annotation categories include "high concentration", "moderate concentration", "low concentration" and "non-concentration".

[0073] In this embodiment, the current training data is input into a preset generative adversarial network for sample data expansion, and the first sample data is output, specifically:

[0074] Input the current training data into the generator network in the generative adversarial network for forward propagation, and output generated sample data and the current training data;

[0075] Input the current training data and the generated sample data into the discriminator network for sample discrimination, output and calculate network loss data and stability loss data according to the discrimination result data, and then adjust the generator network parameters according to the network loss data and the stability loss data;

[0076] At the same time, the network loss data and the stability loss data are evaluated, and then the learning strategies and parameters of the generator network and the discriminator network are adjusted according to the evaluation results, and the discrimination process is iteratively generated until the iteration is completed to output the first sample data.

[0077] In order to improve the comprehensiveness and accuracy of the model's assessment of concentration, when the system collects the real-time training data of the person to be assessed, it will also perform data expansion on the real-time data through a generative adversarial network, so as to improve the quality and richness of the sample data. The system of the present invention inputs the generated sample data output by the generator network and the current training data as input data into the discriminator network for discrimination, and adjusts the network parameters of the generator network according to the discrimination result of the discriminator, so as to improve the quality of the generated samples in a competitive learning manner.

[0078] At the same time, compared with the existing generative adversarial network, the loss function used by the generative adversarial network provided by the system of the present invention includes not only the network loss function but also the stability loss function. The system evaluates the stability of the generated samples through the stable loss function to ensure that the data generated during the training process is more diverse and closer to the actual distribution of the training data. The generator network is adjusted through the feedback of the discriminator network, and the cycle consistency check and adaptive adjustment mechanism are used to guide the generator network to generate data closer to the real sample, thereby improving the stability, generalization and accuracy of the model evaluation.

[0079] In this embodiment, since the collection, acquisition, labeling and preprocessing of training data are time-consuming and labor-intensive, insufficient training samples can easily lead to poor generalization ability of the model and affect the accuracy of the model. Therefore, this embodiment provides a method for data expansion of collected real-time training data, which is a generative adversarial network algorithm based on steady-state rules. The network includes a generator network and a discriminator network, and the two networks improve the quality of generated samples through a competitive learning process. At the same time, compared to traditional generative adversarial networks, the embodiment of the present invention adopts a new loss function, namely the stability loss function, which is used to evaluate the stability of generated samples to ensure that the data generated during the training process is more diverse and close to the real distribution.

[0080] When training the steady-state rule-based generative adversarial network algorithm provided in this embodiment, the first step is to initialize the parameters of the generator G and the discriminator D respectively. Specifically, let the parameter of the generator be b G , the parameter of the discriminator is b D ; Among them, the generator is responsible for generating data samples from random noise, and the discriminator's task is to distinguish the generated samples from the real samples.

[0081] In the second step, the generator receives the training data as input data and generates data samples through forward propagation. At the same time, the system will guide the generator to generate high-quality and diverse data by learning the stability loss function. Specifically, the generator G receives the training data, that is, the random noise b z , the number of samples generated by the network

[0082] data sample b x can be expressed as:

[0083] b x =G(b z ; b G )

[0084] Among them, b z is a random noise vector, b G The generator parameters.

[0085] In the third step, the discriminator receives real samples, i.e., training data, and fake samples, i.e., data samples, generated by the generator, and learns how to accurately distinguish between the two types of samples through training. At the same time, the output of the discriminator will also be used to calculate the loss function, which includes the traditional generative adversarial network loss and the loss function proposed in this embodiment.

[0086] Specifically, the discriminator D receives the fake sample b from the generator. x and the real sample b x ,

[0087] Output the judgment result b D (b x ) and b D (b x ), which can be expressed as:

[0088] b y =D(b x ; b D )

[0089] b y =D(b x ; b D )

[0090] Among them, b y is the discriminator’s discrimination result for the real sample, b y is the discriminant result of the generated sample. Since the discriminator aims to maximize the accuracy of judging the real sample and the generated sample, its loss function L D The definition is as follows:

[0091]

[0092] Among them, p data is the real data distribution, p G is the generated data distribution.

[0093] The fourth step is to adjust the network parameters of the generator according to the discriminator's discrimination result. The goal is to reduce the discriminator's ability to correctly distinguish true and false samples. The system provided by the present invention uses a cyclic consistency check and an adaptive adjustment mechanism to guide the generator to generate data that is closer to the real sample. Since the training goal of the generator is to minimize the discriminator's ability to correctly identify the generated samples, its loss function L G The definition is as follows:

[0094]

[0095] At the same time, this embodiment also provides a stability loss function L stable , which is used to increase the stability and diversity of generated samples. Its calculation method can be expressed as:

[0096]

[0097] Among them, λ e is the stability loss weight, and ||||1 represents the L1 norm, which aims to minimize the difference between the generated samples and the samples after the feedback loop through D and G. In this embodiment, λ e is set to 0.1.

[0098] In this embodiment, after the system feeds back the discrimination result, the generator generates sample data in the following manner:

[0099] b x =G(D(b x );b G )

[0100] Among them, D(b x ) is actually the generated sample b x The process of evaluating, the output of which is a probability

[0101] Value, indicating b x is the probability of a true sample. And, this probability value is further used by the generator G to generate a new sample b x , then the calculation process of D is further expanded as follows:

[0102] D(b x )=Sig(W D b x +b D )

[0103] Among them, Sig() is the Sigmoid activation function, W D and b D are the weight and bias of the discriminator respectively.

[0104] At the same time, the embodiment of the present invention also combines the cycle consistency loss L cycle And adaptive adjustment mechanism, cycle consistency loss ensures that the samples generated from random noise remain consistent after being regenerated. The calculation method can be expressed as:

[0105]

[0106] Among them, the adaptive adjustment mechanism is to dynamically adjust the parameters b of the generator G by evaluating the discriminator D's discrimination results on the samples generated by the generator G. G . For D(G(b z The specific calculation process of )) is as follows:

[0107] D(G(b z ))=Sig(W D ·G(bz )+b D )

[0108] Among them, G(b z ) represents the sample generated by the generator for the first time, D(G(b z )) represents the probability values ​​of these samples after being evaluated by the discriminator, and these values ​​will be used as input for the generator to generate samples again.

[0109] The fifth step is to iteratively repeat the above steps 2 to 4, and continuously optimize the parameters of the generator and discriminator through multiple iterations. At the same time, the system will also evaluate the quality and diversity of the generated data after each iteration, and adjust the learning strategy and parameters according to the evaluation results. Specifically, through the iterative process, the parameters of G and D are continuously updated to minimize L G and maximize L D , while considering L stable and L cycle The impact on the quality of generated samples and the way to update parameters can be expressed as:

[0110]

[0111] Among them, η G and η D are the learning rates of the generator and discriminator, α and β are the weights of stability loss and cycle consistency loss, respectively. In this embodiment, the learning rates of the generator and discriminator are set to 0.0002 and 0.0001, respectively.

[0112] Step 6: When the preset number of iterations is reached, the training process ends. At this point, the generator is able to generate high-quality data.

[0113] Step 102: Input the first sample data into a feature extraction model, perform feature extraction on the first sample data through iteratively using several vortex parameter sets ranked by performance, and perform optimal vortex parameter adjustment and worst vortex attenuation on the several vortex parameter sets until the iteration is completed and the first feature data is output.

[0114] After obtaining the first sample data, the system inputs the sample data into the feature extraction model, and extracts features from the sample data in turn using several vortex parameter sets ranked by performance in the model. During the feature extraction process, the optimal vortex in the above vortex parameter sets is adjusted and the worst vortex is attenuated, and the feature extraction and vortex adjustment process is continuously iterated until the iteration is completed and the corresponding first feature data is output. The system optimizes the parameters in the feature extraction model by simulating the vortex optimization algorithm of the vortex phenomenon, simulating the formation, evolution and dissipation process of the vortex, reducing the model accuracy's dependence on the gradient, solving the problems of gradient vanishing, gradient explosion and search-oriented local optimal solution that may exist in the model in the prior art, and improving the training efficiency and stability of the model.

[0115] Specifically, the embodiment of the present invention performs feature extraction on the first sample data through iteration of several vortex parameter sets after performance sorting, and performs optimal vortex parameter adjustment and worst vortex attenuation on the several vortex parameter sets until the iteration is completed and the first feature data is output, which specifically includes:

[0116] Performing initial feature extraction on the first sample data, obtaining and outputting initial feature data, retrieving and performing performance evaluation calculation on the initial feature data through the performance loss function of the plurality of vortex parameter sets, and then sorting the plurality of vortex parameter sets according to the calculated loss function values;

[0117] Adjusting the positions and intensities of the plurality of vortex parameter sets according to the corresponding rankings of the plurality of vortex parameter sets after sorting and the distance between every two parameter sets, and performing merging operations and splitting operations on the plurality of vortex parameter sets according to preset merging thresholds and splitting thresholds;

[0118] The performance of the several vortex parameter sets is evaluated according to a preset performance algorithm, and the parameters of the optimal vortex determined by the evaluation are adjusted through local search. At the same time, the worst vortex determined by the evaluation is attenuated through life value adjustment, and the feature extraction process is iterated until the iteration is completed and the first feature data is output.

[0119] In order to solve the problem of model accuracy being too dependent on gradients, the system will optimize the parameters in the neural network by simulating the formation, evolution, and dissipation of vortices when extracting features from sample data, and provides a vortex-based feature extraction method. In fluid dynamics, vortices can increase the chaos and complexity of the fluid, thereby guiding the search process in the parameter space toward a better solution. By simulating this natural process, the vortex optimization neural network can find the parameter configuration that can minimize the neural network loss function, thereby reducing the dependence of model parameter optimization on model gradients.

[0120] In this embodiment, see Figure 3 , Figure 3 A schematic diagram of a flow chart of an embodiment of a vortex optimization algorithm for simulating vortex phenomena provided by the present invention. Figure 3 As shown, the feature extraction model provided in this embodiment uses a 3-layer fully connected neural network to extract features from sample data. At the same time, in order to avoid the problem of gradient disappearance, gradient explosion or falling into a local optimal solution, and the resulting impact on the stability of training and the performance of the model, this embodiment also provides a vortex optimization neural network algorithm, which optimizes the parameters (weight w and bias b) in the neural network by simulating the formation, evolution and dissipation process of the vortex. In fluid dynamics, vortices can increase the chaos and complexity of the fluid, thereby guiding the search process in the parameter space towards a better solution. By simulating this natural process, the vortex optimization neural network aims to find a parameter configuration that can minimize the neural network loss function, without relying on the gradient-based neural network parameter optimization method.

[0121] like Figure 3 As shown, the training process of the vortex optimization neural network algorithm described in this embodiment is as follows:

[0122] The first step is to randomly generate a set of vortices in the parameter space, namely Figure 3 The parameter set is also the vortex parameter set in this embodiment. Each vortex represents a possible parameter configuration (including weight w and bias b). Specifically: define the neural network parameter vector P i =(w i ,b i ), where i is the vortex index. During initialization, the parameter vector P of each vortex i It is randomly generated in the parameter space and conforms to the uniform distribution. If the total number of initial vortices is set to N, the initialization process can be expressed as:

[0123]

[0124] in, Represents uniform distribution, a and b are the lower and upper bounds of the parameter space, which are preset manually.

[0125] The second step is to calculate the performance of the parameter configuration represented by each vortex on the training set and sort the vortices according to the performance. Specifically: The performance evaluation uses the loss function L(P i ) is used to complete the performance evaluation through the preset softmax function. After the evaluation is completed, the vortices are sorted according to the value of the loss function:

[0126] L(P1)≤L(P2)≤…≤L(P N )

[0127] In addition, the loss function can also be cross entropy loss.

[0128] The third step is to dynamically update the vortexes. According to the ranking of the vortices and the distance between them, the position and strength of the vortices are dynamically adjusted to simulate the interaction and evolution of the vortices. Specifically: In the vortex dynamic update stage, the interaction of the vortices is simulated to adjust their positions in the parameter space, and the distance between vortices i and j is set to d ij =∥P i -P j ∥, then the update amount of the i-th vortex ΔP i It can be expressed as:

[0129]

[0130] Among them, γ is a hyperparameter that adjusts the influence of distance, F(dij) is a distance-based adjustment function, λb is an adaptive vortex factor, and ξi is an adjustment factor based on environmental feedback.

[0131] Furthermore, for the distance-based adjustment function F(dij) in the above formula, its specific calculation method is to use an exponential decay model to reflect the characteristic that the interaction force between vortices weakens with increasing distance. The specific expression is as follows:

[0132] F(dij)=αqexp(-βqdij)

[0133] Among them, αq and βq are hyperparameters that control the amplitude and decay rate of the force. In this embodiment, the parameter αq is set to 0.5 and βq is set to 0.1. The dij in the above formula represents the distance between vortex i and vortex j, which is calculated based on the Euclidean distance of the parameter vectors represented by the two vortices. The calculation expression is as follows:

[0134]

[0135] Where D is the dimension of the parameter vector, Pik and Pjk are the parameter values ​​of vortex i and vortex j in the kth dimension respectively.

[0136] Furthermore, for the adaptive vortex factor λb in the vortex update formula, its specific calculation method is to use the adaptive vortex guided learning method to calculate it. First, the information entropy H of the vortex distribution is defined as:

[0137]

[0138] Among them, pi is the proportion of the i-th vortex in all vortices, which is calculated by the inverse of the vortex loss function and can be expressed as:

[0139]

[0140] According to the change of information entropy H, the adaptive vortex factor λb is used to adjust the vortex search strategy, and its calculation method can be expressed as:

[0141]

[0142] Among them, κ is a preset maximum regulation strength, H max is the maximum information entropy value that can be achieved theoretically, indicating a completely uniform distribution state, μ i is the momentum adjustment factor, M i is the average momentum vector of vortex i in the last T iterations. Due to the momentum adjustment factor μ i and the average momentum vector M i It is obtained by time series momentum adjustment. The time series momentum adjustment aims to adjust the exploration direction of the vortex by analyzing its historical movement trend in the parameter space. Therefore, the system of this embodiment will consider the movement direction and speed of the vortex in the historical iteration, calculate its momentum, and then use this momentum information to guide future parameter updates. Specifically, the average momentum vector M of vortex i in the last T iterations is calculated. i It can be expressed as:

[0143]

[0144] Among them, P i (t) is the parameter vector of vortex i after the tth iteration, P i (t-1) is the parameter vector after the t-1th iteration. And the momentum adjustment factor μ is defined i It is used to measure the size and consistency of momentum and can be expressed as:

[0145] μ i =∥M i ∥·cos(θ i )

[0146] Among them, θ i It is M i and P i (t) -P i (t-1) ) is used to measure the consistency of the latest moving direction with the average momentum direction, and the smaller the angle, the more consistent the momentum.

[0147] Furthermore, for the adjustment factor ξ based on environmental feedback in the vortex update formula i The calculation method can be expressed as:

[0148] ξ i =Sig(E(P i ))

[0149] Among them, Sig() is the Sigmoid mapping function, which maps the environmental feedback to the (0, 1) interval. i ) is the environmental feedback function, which is used to evaluate the characteristics of the environment around vortex i. The calculation method can be expressed as:

[0150]

[0151] in, is the Hessian matrix of the loss function at the vortex i position (used to evaluate local concavity), is the magnitude of the gradient (used to assess steepness in parameter space).

[0152] The fourth step is to simulate the merging and splitting of vortices in nature to promote the understanding of spatial diversity and explore new parameter areas. Specifically, a merging threshold T is preset. m and a splitting threshold T s , when the distance between the two vortices is less than T m When the strength of a vortex exceeds T s When , it will split into two vortices. The merging operation can be expressed as:

[0153]

[0154] Among them, L(P i ) and L(P j ) is the loss function value corresponding to vortex i and vortex j. Correspondingly, the splitting operation can be expressed as:

[0155] P i1 =P i +δ q

[0156] P i2 =P i -δ q

[0157] Among them, δ q It is a random perturbation vector based on the parameter vector of vortex i, which is used to generate the parameters of two new vortices. For the random perturbation vector δ in the above formula, it is generated based on the local gradient information of the current vortex position to increase the randomness and exploration of the search process. The calculation method can be expressed as:

[0158]

[0159] Among them, q and ε qis a hyperparameter that controls the gradient direction and random perturbation intensity, and g represents the vortex position P i The gradient information calculated at The mean is 0 and the covariance matrix is In this embodiment, ζ q is set to 1.0, ε q is set to 0.1.

[0160] The fifth step is to select the best performing vortex and perform local search to fine-tune its parameters to simulate the local absorption and energy concentration of the vortex in the fluid. Specifically, for the best performing vortex, perform local search to further fine-tune its parameters, which is achieved through a domain-based search strategy. In the domain-based search strategy, the local search method can be expressed as:

[0161]

[0162] Among them, η m is the learning rate, is 0 as the mean, Σ loc is the Gaussian distribution of the covariance matrix, which is used to generate random perturbations of the parameters. loc , which can be dynamically adjusted according to the historical moving direction and speed of the vortex in the parameter space to adapt to different search stages and characteristics. Its update strategy can be expressed as:

[0163]

[0164] in, is the adjusted covariance matrix, is the covariance matrix before adjustment, ρ n is a learning rate parameter used to control the influence of historical information on the current search. In this embodiment, ρ n is set to 0.2.

[0165] The sixth step is to perform vortex attenuation, simulating the dissipation process of vortex energy in nature, gradually eliminating vortices with poor performance, and releasing resources to newly generated or better performing vortices. Specifically, the attenuation process can be achieved by adjusting the life value of the vortex. Vortices with a life value below a certain threshold will be eliminated. The life value of vortex i is set to L i , the decay rule is:

[0166]

[0167] in, is the health value after decay, δ L It is the amount of health reduction.

[0168] Step 7: Repeat the iterations of Steps 2 to 6 until the maximum number of iterations is reached, and the iterations are completed. At the same time, after each round of iterations, the state of the vortices is updated and their performance is re-evaluated to ensure that the algorithm progresses toward the optimal solution. In this embodiment, the maximum number of iterations is set to 1000.

[0169] Step 103: Input the first feature data into the encoder of the feature dimensionality reduction model for encoding, and iteratively adjust the parameters of the encoder according to the inverse reinforcement learning mechanism. At the same time, adjust the encoding process according to the performance feedback of the decoder until the iteration is completed and the first dimensionality reduction data is output. Input the first dimensionality reduction data into the classifier for concentration evaluation, and output the evaluation result.

[0170] In order to further improve the precision and accuracy of model evaluation, after obtaining the first feature data, the system also inputs the first feature data into the encoder of the feature dimensionality reduction model for encoding processing, and iteratively adjusts the encoder parameters through the inverse reinforcement learning mechanism during the encoding process, further optimizing the feature extraction of the encoding process. At the same time, the encoding process is adjusted according to the performance feedback of the decoder, and the encoder is adjusted through an interactive optimization mechanism to ensure that the decoded data retains the information in the feature data that is most relevant to the concentration evaluation as much as possible, that is, the key information for the concentration evaluation, further improving the accuracy and efficiency of the concentration evaluation model. After completing the feature dimensionality reduction, the system inputs the reduced dimensionality data into the classifier for evaluation, and then outputs the concentration evaluation result.

[0171] Specifically, the embodiment of the invention iteratively adjusts the parameters of the encoder according to the inverse reinforcement learning mechanism, and adjusts the encoding process according to the performance feedback of the decoder until the iterative output of the first dimensionality reduction data is completed, specifically including:

[0172] Encoding the first feature data, and calling a corresponding reward function according to the correlation between the first feature data and the concentration evaluation index, adjusting the parameters of the encoder by using a reverse learning mechanism during the encoding process, and outputting the first encoded data;

[0173] The first encoded data is input into a decoder so that the decoder performs feature reconstruction based on the first encoded data and outputs first reconstructed data. The correlation between the first feature data and the first reconstructed data is then used as performance feedback of the decoder to adjust the encoding process, and the encoding and decoding process is iterated until the iteration is completed and the first reduced dimensionality data is output.

[0174] In order to further optimize the extracted feature data, when the system performs feature dimensionality reduction processing on the feature data, compared with the traditional autoencoder that realizes data dimensionality reduction and reconstruction through encoding and decoding processes, the system provided by the present invention adopts an inverse reinforcement learning mechanism to further optimize the feature extraction in the encoding process. The reward function called is related to the feature data and the concentration evaluation index, so that it pays more attention to the features in the amblyopia training data that are closely related to the concentration evaluation.

[0175] At the same time, the encoding process of the encoder is adjusted according to the correlation between the reconstructed data and the feature data output by the decoder as performance feedback, and the encoder is adjusted through an interactive optimization mechanism to ensure that the decoded data retains the most critical information for concentration assessment in the original data as much as possible, further improving the accuracy and evaluation precision of the output reduced dimensionality data for concentration assessment.

[0176] In this embodiment, compared with the traditional autoencoder that realizes data dimension reduction and reconstruction through encoding and decoding processes, the system of this embodiment adopts the reverse reinforcement learning mechanism to further optimize the feature extraction in the encoding process, so that it pays more attention to the features closely related to the concentration assessment in the amblyopia training data. At the same time, during the training process of the autoencoder, the system will also adjust the encoder through an interactive optimization mechanism to ensure that the decoded data retains the most critical information for concentration assessment in the original data as much as possible.

[0177] Specifically, the training process of the autoencoder neural network structure based on inverse reinforcement learning described in this embodiment is as follows:

[0178] The first step is to initialize the encoder and decoder parameters of the autoencoder. The system provided in this embodiment initializes the encoder and decoder parameters in a random manner to provide a starting point for the training process. Let the encoder weight be eW enc , the encoder bias is eb enc , the decoder weight is eW dec , the decoder bias is eb dec .

[0179] In the second step, the feature-extracted data is input to the encoder, and the data is processed by a multi-layer neural network, each layer of which is designed to capture high-level features in the data. Specifically, the input data ex is converted to the encoded representation ez by the encoder, which can be expressed as: ez=Re(eW enc ·ex+eb enc )

[0180] Among them, Re() is the ReLU activation function.

[0181] The third step is to use the inverse reinforcement learning mechanism during the encoding process to adjust the encoder parameters according to a predetermined reward function, which is based on the correlation between the encoded features and the concentration evaluation to ensure that the extracted features are most useful for the final concentration evaluation. Specifically, the reward function R(ez) is defined based on the correlation between the encoded representation ez and the target evaluation index, and the encoder parameters are updated using gradient ascent to maximize the reward. The parameter update method can be expressed as:

[0182]

[0183]

[0184] Among them, α c is the learning rate. In this embodiment, α c Set to 0.001. In this embodiment, the calculation method of the reward function R(ez) can be expressed as:

[0185] R(ez)=Corr(ez,y target )

[0186] Among them, Corr represents the correlation function, which is used to calculate the target variable y of the encoding representation ez and the concentration evaluation target The correlation coefficient between target is the label of the data. The gradient term in the parameter update formula and The calculation method is based on the partial derivative of the reward function with respect to the parameters, which can be expressed as:

[0187]

[0188]

[0189] The fourth step is to input the encoded data into the decoder so that the decoder attempts to reconstruct the original data based on the encoded features and compare the reconstructed data with the original data to evaluate the integrity of feature preservation. Specifically, the encoded representation ez is converted back to the reconstructed data ex by the decoder ′ , which can be expressed as:

[0190] ex ′ =Sig(eW dec ez+eb dec )

[0191] The fifth step is to adjust the encoder parameters based on the performance feedback of the decoding process. If the decoded data performs poorly in retaining key attention features, the encoding process is adjusted through an interactive optimization mechanism to better capture and retain these key features. Specifically, the interactive optimization adjustment method is based on the reconstruction error Lrecon Adjust the encoder parameters, which can be expressed as:

[0192]

[0193] Among them, ||||2 is the symbol of L2 norm. Since the parameter adjustment adopts the gradient descent method, it can be expressed as:

[0194]

[0195] Where β is the learning rate, which is used to control the magnitude of parameter updates based on the reconstruction error.

[0196] Step 6: Repeat the above steps 2 to 5 until the preset number of iterations is reached, that is, the iteration is completed. In this embodiment, the number of iterations is set to 500.

[0197] Furthermore, in this embodiment, the first dimension reduction data is input into a classifier to perform concentration evaluation, and an evaluation result is output, specifically:

[0198] Initializing the twin network of the classifier according to the feature dimension of the first dimension reduction data, and setting initial parameters of the submodular function optimization algorithm;

[0199] The first dimension reduction data is divided into a positive sample set and a negative sample set, and input into two support vector machines of the twin network for model training, and the submodular function optimization algorithm is used to iteratively update the parameters of the support vector machine and the weight of the loss function until convergence;

[0200] After convergence, the classifier output data is output as the evaluation result.

[0201] In order to improve the evaluation accuracy, when the system uses a classifier to perform classification and evaluation on the reduced-dimensional data, compared with the traditional support vector machine algorithm when processing large-scale nonlinear separable data, it may encounter optimization difficulties. The system provided by the present invention adopts a submodular function optimization strategy, which can improve the convergence speed of the algorithm while ensuring that a more global optimal solution is found.

[0202] In this embodiment, since the traditional support vector machine algorithm may encounter optimization difficulties when processing large-scale nonlinear separable data, the embodiment of the present invention adopts a twin support vector machine algorithm based on submodular function optimization. Through the submodular function optimization strategy, the convergence speed of the algorithm can be improved while ensuring that a more global optimal solution is found. The twin network structure includes two parallel support vector machines, which can more accurately learn the differences between features, thereby improving classification accuracy.

[0203] Therefore, the training process of the twin support vector machine algorithm based on submodular function optimization described in this embodiment is as follows:

[0204] First, according to the feature dimension of the input data, the parameters of the twin support vector machine are initialized, including selecting a suitable kernel function to map the original features to a high-dimensional space and setting the initial parameters of the submodular optimization algorithm. Specifically, it includes: setting the parameters in the twin support vector machine, where pθ represents the model parameter and pλ represents the regularization term coefficient, which is used to control the complexity of the model to avoid overfitting. The initial setting is:

[0205] pθ0=0

[0206] pλ=0.1

[0207] For the input feature set {x i ,y i}, where i = 1, 2, ..., n, x i is the input feature vector, y i ∈{-1,1} is the corresponding label, and the objective function pF(pθ) is defined as:

[0208]

[0209] Among them, pb is the bias term, is the regularization term, The loss term optimized for the submodular function.

[0210] Secondly, the input data is divided into positive sample sets and negative sample sets, and input into the two support vector machines of the twin network respectively. The submodular function optimization strategy is adopted to iteratively update the parameters of the support vector machine until convergence. Specifically, the parameters pθ and pb are updated using the gradient descent method. During the iteration process, for each parameter, the update rule is:

[0211]

[0212]

[0213] Where pη is the learning rate, and are the gradients of pF with respect to pθ and pb at the tth iteration. In this embodiment, the learning rate pη is set to 0.01. And for the gradient of the loss function and The specific calculation method is:

[0214]

[0215]

[0216] Finally, in each iteration, the weight of the loss function is automatically adjusted according to the performance of the current model on the training set. Specifically, a threshold pε is determined. When the difference between the objective function values ​​of two consecutive iterations is less than the threshold, the iteration is stopped. It can be expressed as:

[0217] |pF(pθ t+1 )-pF(pθ t )| <pε

[0218] In this embodiment, the termination condition threshold pε is set to 1e-4.

[0219] Through the above method, the concentration of the person to be evaluated during the amblyopia training process can be evaluated based on the real-time training data collected from the person to be evaluated. The evaluation results output by this embodiment include "high concentration", "moderate concentration", "low concentration" and "non-concentration".

[0220] In order to better illustrate the working principle and step flow of the concentration assessment method, device and system for amblyopia training of the present invention, it is possible but not limited to refer to the relevant records above.

[0221] Accordingly, see Figure 2 , Figure 2 This is a schematic diagram of the structure of an embodiment of a concentration assessment device for amblyopia training provided by the present invention. Figure 2 As shown, the concentration evaluation device includes a sample collection module 201, a feature extraction module 202, a feature dimension reduction module 203 and a result output module 204.

[0222] The sample collection module 201 is used to obtain current training data of the amblyopic user to be evaluated, and input the current training data into a preset generative adversarial network to perform sample data expansion, and output first sample data.

[0223] Furthermore, the sample collection module 201 inputs the current training data into a preset generative adversarial network to perform sample data expansion, and outputs first sample data, specifically:

[0224] Input the current training data into the generator network in the generative adversarial network for forward propagation, and output generated sample data and the current training data;

[0225] The current training data and the generated sample data are input into the discriminator network for sample discrimination, and the network loss data and the stability loss data are output and calculated according to the discrimination result data, and then the generator network parameters are adjusted according to the network loss data and the stability loss data; at the same time, the network loss data and the stability loss data are evaluated, and then the learning strategies and parameters of the generator network and the discriminator network are adjusted according to the evaluation results, and the discrimination process is iteratively generated until the iteration is completed and the first sample data is output.

[0226] The feature extraction module 202 is used to input the first sample data into a feature extraction model, perform feature extraction on the first feature data through iteratively using several vortex parameter sets ranked by performance, and perform optimal vortex parameter adjustment and worst vortex attenuation on the several vortex parameter sets until the iteration is completed and the first feature data is output.

[0227] Further, the feature extraction module 202 iteratively extracts features from the first sample data through the several vortex parameter sets after performance sorting, and simultaneously performs optimal vortex parameter adjustment and worst vortex attenuation on the several vortex parameter sets until the iteration is completed and the first feature data is output, specifically including:

[0228] Performing initial feature extraction on the first sample data, obtaining and outputting initial feature data, retrieving and performing performance evaluation calculation on the initial feature data through the performance loss function of the plurality of vortex parameter sets, and then sorting the plurality of vortex parameter sets according to the calculated loss function values;

[0229] The positions and intensities of the several vortex parameter sets are adjusted according to their corresponding rankings after sorting and the distance between every two parameter sets, and the several vortex parameter sets are merged and split according to the preset merging threshold and splitting threshold; the performance of the several vortex parameter sets is evaluated according to the preset performance algorithm, and the parameters of the optimal vortex determined by the evaluation are adjusted through local search, and the worst vortex determined by the evaluation is attenuated through life value adjustment, and the feature extraction process is iterated until the iteration is completed and the first feature data is output.

[0230] The feature dimensionality reduction module 203 is used to input the first feature data into the encoder of the feature dimensionality reduction model for encoding processing, and iteratively adjust the parameters of the encoder according to the inverse reinforcement learning mechanism, and adjust the encoding process according to the performance feedback of the decoder until the iteration is completed and the first dimensionality reduction data is output.

[0231] Furthermore, the feature dimension reduction module 203 iteratively adjusts the parameters of the encoder according to the inverse reinforcement learning mechanism, and adjusts the encoding process according to the performance feedback of the decoder until the iteration is completed and the first dimension reduction data is output, which specifically includes:

[0232] The first feature data is encoded, and the corresponding reward function is called according to the correlation between the first feature data and the concentration evaluation index. The parameters of the encoder are adjusted by a reverse learning mechanism during the encoding process, and the first encoded data is output; the first encoded data is input into the decoder, so that the decoder performs feature reconstruction according to the first encoded data, and outputs first reconstructed data, and then the correlation between the first feature data and the first reconstructed data is used as the performance feedback of the decoder to adjust the encoding process, and the encoding and decoding process is iterated until the iteration is completed and the first dimensionality reduction data is output.

[0233] The result output module 204 is used to input the first dimension reduction data into a classifier to perform concentration evaluation and output an evaluation result.

[0234] Furthermore, the result output module 204 inputs the first dimension reduction data into a classifier to perform concentration evaluation, and outputs an evaluation result, specifically:

[0235] Initializing the twin network of the classifier according to the feature dimension of the first dimension reduction data, and setting initial parameters of the submodular function optimization algorithm;

[0236] The first dimension reduction data is divided into a positive sample set and a negative sample set and input into the two support vector machines of the twin network for model training, and the submodular function optimization algorithm is used to iteratively update the parameters of the support vector machine and the weight of the loss function until convergence; after convergence, the classifier output data is output as the evaluation result.

[0237] Correspondingly, an embodiment of the present invention also provides a concentration assessment system for amblyopia training, the concentration assessment system comprising a memory, a processor, and a computer program stored in the memory and running on the processor, and the processor implements any of the above-mentioned concentration assessment methods for amblyopia training when processing the computer program.

[0238] Correspondingly, an embodiment of the present invention further provides a storage medium, on which a computer program is stored. The computer program is called and executed by a processor to implement a concentration assessment method for amblyopia training as described in any one of the above items.

[0239] In summary, the embodiments of the present invention provide a method, device and system for evaluating concentration for amblyopia training, which obtains first sample data by expanding the current training data of the user to be evaluated through a generative adversarial network, extracts features from the sample data through iterative extraction of several vortex parameter sets after performance sorting, and adjusts the optimal vortex parameters and the worst vortex attenuation of the parameter set until the first feature data is output, then inputs the first feature data into the encoder network of the feature dimensionality reduction model for encoding, and iteratively adjusts the encoder parameters according to the inverse reinforcement learning mechanism, and adjusts the encoding process according to the decoder performance feedback until the first dimensionality reduction data is output, and then evaluates the dimensionality reduction data through a classifier to output an evaluation result. By expanding the sample data of real-time data, the quality and diversity of the sample data are improved; by simulating the vortex optimization algorithm of the vortex phenomenon, the dependence of the model accuracy on the gradient is reduced.

[0240] The specific embodiments described above further illustrate the purpose, technical solutions and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. It is particularly pointed out that for those skilled in the art, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for assessing concentration for amblyopia training, characterized in that: The following steps are involved: Acquire current training data of the amblyopic user to be evaluated, and input the current training data into a preset generative adversarial network for sample data expansion, and output first sample data; A feature extraction model is constructed, the first sample data is input into the feature extraction model, initial feature extraction is performed on the first sample data, initial feature data is obtained and output, and the initial feature data is iteratively optimized through training of the feature extraction module to output first feature data; wherein the feature extraction model adopts a vortex optimization neural network, and the parameters in the feature extraction model are optimized by simulating the formation, evolution and dissipation process of the vortex, and the training steps of the feature extraction model are as follows: Randomly generating a set of vortices through the parameter space, wherein each vortex represents a parameter configuration, including weights and biases, to define a parameter vector of the neural network; Calculating the performance of the parameter configuration represented by each vortex on the training set, and sorting the vortices according to the performance, that is, using a loss function to perform performance evaluation calculation, and then sorting the vortices according to the calculated loss function value; The positions and intensities of the vortices are adjusted according to the corresponding rankings of the sorted vortices and the distance between every two vortices, so as to dynamically update the vortices, and at the same time, the vortices are merged and split according to the preset merging threshold and splitting threshold; The performance of the vortex is evaluated according to a preset performance algorithm, and the parameters of the optimal vortex determined by the evaluation are adjusted through local search, and the worst vortex determined by the evaluation is attenuated through life value adjustment, and the feature extraction process is iterated until the iteration is completed and the first feature data is output; wherein the life value is calculated and adjusted according to the attenuation rule; The first feature data is input into the encoder of the feature dimensionality reduction model for encoding processing, and the parameters of the encoder are iteratively adjusted according to the inverse reinforcement learning mechanism. At the same time, the encoding process is adjusted according to the performance feedback of the decoder until the iteration is completed and the first dimensionality reduction data is output. The first dimensionality reduction data is input into the classifier for concentration evaluation, and the evaluation result is output.

2. A method for assessing concentration for amblyopia training according to claim 1, characterized in that: The step of inputting the current training data into a preset generative adversarial network to perform sample data expansion and outputting first sample data is specifically as follows: Input the current training data into the generator network in the generative adversarial network for forward propagation, and output generated sample data and the current training data; Input the current training data and the generated sample data into the discriminator network for sample discrimination, output and calculate network loss data and stability loss data according to the discrimination result data, and then adjust the generator network parameters according to the network loss data and the stability loss data; At the same time, the network loss data and the stability loss data are evaluated, and then the learning strategies and parameters of the generator network and the discriminator network are adjusted according to the evaluation results, and the discrimination process is iteratively generated until the iteration is completed to output the first sample data.

3. A method for assessing concentration for amblyopia training according to claim 1, characterized in that: The iterative adjustment of the encoder parameters according to the inverse reinforcement learning mechanism and the adjustment of the encoding process according to the performance feedback of the decoder until the iterative output of the first dimension reduction data is completed specifically include: Encoding the first feature data, and calling a corresponding reward function according to the correlation between the first feature data and the concentration evaluation index, adjusting the parameters of the encoder by using a reverse learning mechanism during the encoding process, and outputting the first encoded data; The first encoded data is input into a decoder so that the decoder performs feature reconstruction based on the first encoded data and outputs first reconstructed data. The correlation between the first feature data and the first reconstructed data is then used as performance feedback of the decoder to adjust the encoding process, and the encoding and decoding process is iterated until the iteration is completed and the first reduced dimensionality data is output.

4. The method for assessing concentration for amblyopia training according to claim 1, characterized in that: The first dimension reduction data is input into a classifier to perform concentration evaluation, and the evaluation result is output, specifically: Initializing the twin network of the classifier according to the feature dimension of the first dimension reduction data, and setting initial parameters of the submodular function optimization algorithm; The first dimension reduction data is divided into a positive sample set and a negative sample set, and input into two support vector machines of the twin network for model training, and the submodular function optimization algorithm is used to iteratively update the parameters of the support vector machine and the weight of the loss function until convergence; After convergence, the classifier output data is output as the evaluation result.

5. A concentration assessment device for amblyopia training, characterized in that: The concentration assessment device includes a sample collection module, a feature extraction module, a feature dimension reduction module and a result output module; The sample acquisition module is used to obtain current training data of the amblyopic user to be evaluated, and input the current training data into a preset generative adversarial network to perform sample data expansion, and output first sample data; The feature extraction module is used to construct a feature extraction model, input the first sample data into the feature extraction model, perform initial feature extraction on the first sample data, obtain and output initial feature data, and iteratively optimize the initial feature data through training of the feature extraction module to output first feature data; wherein, the feature extraction model adopts a vortex optimization neural network, and optimizes the parameters in the feature extraction model by simulating the formation, evolution and dissipation process of the vortex, and the training steps of the feature extraction model are as follows: Randomly generating a set of vortices through the parameter space, wherein each vortex represents a parameter configuration, including weights and biases, to define a parameter vector of the neural network; Calculating the performance of the parameter configuration represented by each vortex on the training set, and sorting the vortices according to the performance, that is, using a loss function to perform performance evaluation calculation, and then sorting the vortices according to the calculated loss function value; The positions and intensities of the vortices are adjusted according to the corresponding rankings of the sorted vortices and the distance between every two vortices, so as to dynamically update the vortices, and at the same time, the vortices are merged and split according to the preset merging threshold and splitting threshold; The performance of the vortex is evaluated according to a preset performance algorithm, and the parameters of the optimal vortex determined by the evaluation are adjusted through local search, and the worst vortex determined by the evaluation is attenuated through life value adjustment, and the feature extraction process is iterated until the iteration is completed and the first feature data is output; wherein the life value is calculated and adjusted according to the attenuation rule; The feature dimension reduction module is used to input the first feature data into the encoder of the feature dimension reduction model for encoding processing, and iteratively adjust the parameters of the encoder according to the inverse reinforcement learning mechanism, and adjust the encoding process according to the performance feedback of the decoder until the iteration is completed and the first dimension reduction data is output; The result output module is used to input the first dimension reduction data into a classifier to perform concentration evaluation and output the evaluation result.

6. The concentration assessment device for amblyopia training according to claim 5, characterized in that: The sample collection module inputs the current training data into a preset generative adversarial network to perform sample data expansion, and outputs first sample data, specifically: Input the current training data into the generator network in the generative adversarial network for forward propagation, and output generated sample data and the current training data; Input the current training data and the generated sample data into the discriminator network for sample discrimination, output and calculate network loss data and stability loss data according to the discrimination result data, and then adjust the generator network parameters according to the network loss data and the stability loss data; At the same time, the network loss data and the stability loss data are evaluated, and then the learning strategies and parameters of the generator network and the discriminator network are adjusted according to the evaluation results, and the discrimination process is iteratively generated until the iteration is completed to output the first sample data.

7. The concentration assessment device for amblyopia training according to claim 5, characterized in that: The result output module inputs the first dimension reduction data into a classifier to perform concentration evaluation, and outputs the evaluation result, specifically: Initializing the twin network of the classifier according to the feature dimension of the first dimension reduction data, and setting initial parameters of the submodular function optimization algorithm; The first dimension reduction data is divided into a positive sample set and a negative sample set, and input into two support vector machines of the twin network for model training, and the submodular function optimization algorithm is used to iteratively update the parameters of the support vector machine and the weight of the loss function until convergence; After convergence, the classifier output data is output as the evaluation result.

8. A concentration assessment system for amblyopia training, characterized in that: The concentration assessment system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor processes the computer program, it implements a concentration assessment method for amblyopia training as described in any one of claims 1 to 4.

9. A storage medium, characterized in that: The storage medium stores a computer program, which is called and executed by a processor to implement a concentration assessment method for amblyopia training as described in any one of claims 1 to 4 above.

Images