Eye muscle regulation ability training method and visual training graph generation method thereof

Abstract

The invention relates to the technical field of vision intervention, in particular to an eye muscle regulation ability training method and a visual training graph generation method thereof.According to the eye muscle regulation ability training method, feature information of a trainee and distance information of training are obtained, and a preset image generation network model is used for generating a corresponding visual training graph; the training device is used for training the eye muscle adjusting ability of a trainee. According to the method, the initial distance point location is matched based on the eyesight data of the trainee, the corresponding vision training graph is generated according to the current distance point location and the feature data of the trainee, the training intensity is dynamically adjusted, a personalized training scheme is achieved, the eye muscle adjusting ability is helped to be improved, and the training effect is improved. The problem of poor effect or overtraining caused by inappropriate training intensity is avoided, and the technical problems that in an existing eye muscle training method, due to insufficient individual pertinence, the training effect is difficult to quantify, and training scientificity and systematicness are insufficient are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of vision intervention, and in particular relates to a method for training eye muscle adjustment ability and a method for generating a visual training diagram thereof. Background Art

[0002] In recent years, the myopia rate among adolescents has continued to rise and has become a global public health issue. High myopia not only affects learning and quality of life, but may also cause a series of serious eye diseases, such as retinal detachment and macular degeneration. Early intervention and scientific vision training can help quickly improve the control ability of the eye muscles and achieve rapid relaxation of the ciliary muscles, thereby better controlling the development of myopia. However, there is currently a lack of effective predictive methods to assess the myopia risk and training effects of adolescents, which makes the implementation and optimization of vision training face many challenges. In addition, families and schools also have certain deficiencies in knowledge about myopia prevention and control, and are unable to provide sufficient support and guidance.

[0003] Vision training, through a series of eye exercises and visual stimulation, aims to enhance the ability of eye muscles to adjust and improve blood circulation in the eyes, thereby preventing and correcting myopia. Although this method is theoretically feasible, it has encountered many difficulties in practice.

[0004] The existing eye muscle training methods generally have the following problems: 1. Lack of individual targeting: Traditional training programs often adopt a unified standard and fail to fully consider individual differences, such as different people's initial vision levels, eye health conditions, and differences in muscle adjustment abilities of different ages and genders. 2. The training effect is difficult to quantify: There is a lack of effective monitoring tools and technical means to track the training effects in real time, which makes it difficult to accurately evaluate the actual effectiveness of the training. 3. Lack of scientificity and systematicity: Existing training methods mostly rely on experience and intuition, lack of systematic theoretical support and scientific design, making the training process less rigorous and difficult to guarantee results. These problems exist mainly because traditional training methods ignore individual needs and lack sufficient technical support for accurate effect evaluation. To solve these problems, it is necessary to comprehensively use modern scientific and technological means, combined with professional ophthalmic knowledge, to develop a training system that has both scientific basis and can meet personalized needs.

[0005] Therefore, how to design an eye vision training program to better meet the training needs of personalized training methods based on individual characteristics, monitoring of training status, and improving the adjustment ability of eye muscles has become an urgent problem to be solved. Summary of the Invention

[0006] In response to the above-mentioned deficiencies in the existing technologies, the present invention provides a method for training eye muscle adjustment ability, which can dynamically adjust the training intensity according to individual vision, effectively monitor the training status, and help improve the eye muscle adjustment ability, thereby solving the technical problems in existing eye muscle training methods such as insufficient individual targeting, difficulty in quantifying training effects, and lack of scientific and systematic training.

[0007] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0008] In the first aspect, the present invention provides a method for training eye muscle adjustment ability, which obtains the characteristic information of the trainee and the training distance information, and uses a preset image generation network model to generate a corresponding visual training image for training the trainee's eye muscle adjustment ability.

[0009] The method specifically comprises the following steps:

[0010] S1. Preset distance points for training the trainee; there are multiple distance points, each at a different distance from the observation target;

[0011] S2. Acquire characteristic information of the trainee, wherein the characteristic information of the trainee includes the trainee's vision data, match the initial distance points based on the trainee's vision data, and determine the initial training distance;

[0012] S3, based on the trainee's feature information and the training distance, using a preset image generation network model to generate a corresponding visual training image;

[0013] S4. Using the generated visual training chart, the trainee is trained at corresponding distance points in the order of the weak-vision eye, the strong-vision eye, and both eyes; the weak-vision eye is the eye with the lower naked eye vision of the two eyes, and the strong-vision eye is the eye with the higher naked eye vision of the two eyes. If the trainee's naked eye vision of both eyes is the same, one eye is used as the weak-vision eye and the other eye is used as the strong-vision eye;

[0014] During the training process, after the total gaze time of the weak-vision eye, the strong-vision eye, and both eyes at the distance point reaches the preset time t, the trainee is asked to move to an adjacent distance point farther from the observation target and return to S3 until the training of the distance point farthest from the observation target is completed or the total training time reaches the preset total time T0, and the training ends.

[0015] Compared with the prior art, the present invention has the following beneficial effects:

[0016] 1. Personalized Training Intensity Adjustment. This method matches the initial distance points based on the trainee's vision data, generates a corresponding visual training map based on the current distance points and the trainee's characteristic data, and dynamically adjusts the training intensity to achieve a personalized training plan. Compared with traditional, uniformly standardized training methods, this method can more accurately meet the needs of different individuals and avoid problems such as poor results or overtraining due to inappropriate training intensity.

[0017] 2. Real-time monitoring and feedback mechanism. During training, the system collects eye data from the trainee's training eye and determines whether it is in a fixed gaze state. This allows for real-time monitoring of the training process. This not only ensures the effectiveness of training but also allows for timely adjustments to training strategies to address any potential issues. In contrast, traditional training methods lack this real-time feedback mechanism, making it difficult to guarantee training quality.

[0018] 3. Comprehensive coverage of binocular vision. This method not only focuses on training one eye (the weak-vision eye and the strong-vision eye), but also emphasizes the coordinated training of both eyes. This comprehensive training method helps to comprehensively improve the adjustment ability of eye muscles and promote the coordinated development of binocular vision. Traditional training methods often focus on monocular training and ignore the importance of binocular coordination.

[0019] 4. A scientific design that gradually increases difficulty. By setting multiple distance points at different distances and conducting training from near to far, this method adheres to the principle of gradually increasing training difficulty. This helps trainees gradually adapt to higher visual challenges and enhance the adjustment ability of eye muscles. Compared to a static training environment, this method can better stimulate the potential of eye muscles and improve training results.

[0020] In summary, this method dynamically adjusts training intensity based on individual vision, effectively monitors training status, and significantly improves eye muscle regulation. Through personalized training intensity adjustment, real-time monitoring and feedback mechanisms, comprehensive coverage of binocular vision, and a scientific design that gradually increases difficulty, this method overcomes the limitations of traditional eye muscle training methods, such as uncertain results and slow training speed, providing a more scientific, systematic, and effective solution. This approach not only helps improve the trainee's vision but also prevents and slows the progression of myopia, possessing significant application value and broad prospects for widespread adoption.

[0021] Preferably, in step S2, the specific method of matching the initial distance point based on the trainee's vision data is:

[0022] S201, based on the visual acuity value V of the weak-sighted eye in the trainee's visual acuity data a , the trainer's initial estimated training distance L is obtained as followst0 :

[0023] L t0 =5V a -19.4;

[0024] Initial estimated training distance L t0 The unit is meter;

[0025] S202: Find the distance value to the observed target and the initial estimated training distance L in the preset distance point. t0 The closest distance point is used as the initial distance point for matching.

[0026] This setup, by directly linking visual acuity and training distance through a mathematical model, provides a scientifically calculated optimal starting training distance for each individual with varying vision levels. This approach ensures a highly personalized and precise training plan, allowing each individual to begin training at the distance that works best for them.

[0027] Preferably, the image generation network model includes:

[0028] A database module is used to store various graphic elements and their multimodal features used to generate visual training images, as well as all historical user training graphic element interaction data and feature data of each historical user. The types of graphic elements include image background, observation targets, and distractors, each category containing several graphic elements. The multimodal features of the graphic elements include image features and text features. The historical user training graphic element interaction data is used to record the interaction relationships between various graphic elements corresponding to suitable visual training images used by different users in historical training.

[0029] A graph conversion module is used to convert the historical user training graph element interaction data into a user-graph element interaction graph G = (D, E); wherein each historical user and each graph element constitutes a node in the graph, and the node set is marked as D; the interaction relationship between the historical user node and the graph element node constitutes an edge in the graph, and the edge set is marked as E; and based on the user-graph element interaction graph, an interaction matrix of user nodes to graph element nodes is constructed. a ui Represents the interaction relationship between the historical user node u and the graph element node i, a ui =1 indicates that there is an interactive relationship between the historical user node u and the graph element node i, a ui = 0 means there is no interaction relationship between the historical user node u and the graph element node i, N U 、N I Represent the number of historical users and the number of graph elements respectively;

[0030] The feature extraction module includes a multi-layer graph convolutional network model and a Transformer model, which is used to process the feature data of historical users and the data of the user-graph element interaction graph using the multi-layer graph convolutional network model to obtain the user interaction matching feature f between the feature data of historical users and the graph elements. u And the modal fusion embedding feature extraction process is performed on each graph element through the Transformer model to obtain the modal fusion embedding feature f of each graph element i ;

[0031] Contrastive learning module, used to match features f based on historical user feature data and user interactions between graph elements u and the modal fusion embedding feature f of the graph elements i , calculate the matching value fs of different graph elements relative to historical user features ui , and based on the matching value fs ui Constructing a loss function, iteratively training the multi-layer graph convolutional network model and the Transformer model, and updating and optimizing the parameters of the multi-layer graph convolutional network model and the Transformer model through backpropagation to obtain a trained and optimized model;

[0032] The graph element matching module is used to obtain the feature data of the user to be matched, input it into the trained and optimized multi-layer graph convolutional network model, obtain the user interaction matching features between the feature vector of the user to be matched and the graph elements, and calculate the matching degree value of each different element relative to the user to be matched with the trained and optimized Transformer model. According to the matching degree value, a number of candidate graph elements are selected from each type of graph element to form a subset of matching candidate graph elements;

[0033] The visual training image generation module is used to select image elements from the matching candidate image element subset, and based on a preset image generation program, combine the selected image elements according to the trainee's vision data and the training distance to generate a new visual training image.

[0034] Such a setting can dynamically select image elements based on the user's vision data (such as the difference between amblyopic eyes and strong eyes) to accurately match the training intensity. For example, it can generate targets with higher contrast or brighter colors for amblyopic eyes to increase their recognition difficulty and improve their adjustment ability in a targeted manner. In addition, the newly generated visual training images can also better adapt to the visual training needs of different distances. For example, distant points generate larger and clearer targets (by adjusting the RGB background brightness and target color contrast) to simulate real distant viewing scenes, while close points generate complex and detailed targets (such as small fonts or low-contrast patterns) to enhance close-range adjustment capabilities. Therefore, with the help of the image generation network model, it is possible to optimize the selection of image elements for generating visual training images based on the characteristic information of different individual trainees, and then generate visual training images for each training session based on the preset image generation program according to the distance from the observed target image during specific training. Since the image generation network model is trained based on the historical users' usage and interaction with visual training graphs, the model constructs the interaction relationship between historical users and graph elements through the user-graph element interaction graph. A multi-layer graph convolutional network model is used to perform multi-dimensional deep extraction of the interaction features between historical users and the graph elements they interacted with. The attention mechanism of the Transformer model is used to perform multi-dimensional deep extraction of the deep features of various graph elements. With the help of the contrastive learning mechanism provided by the contrastive learning module, the connection between the two aspects of deep features is established. Therefore, the trained image generation network model can match and generate appropriate visual training graphs based on the individual feature information of the trainee, thereby improving the eye muscle adjustment training effect for the individual trainee.

[0035] Specifically, the feature data of historical users is obtained by embedding the feature information of historical users, which is expressed as X u ={x u,1 ,x u,2 ,…,x u,n ,…,x u,N}, x u,n Represents the word embedding value of the nth user feature of historical user u, where N represents the total number of user features;

[0036] The image features of the image elements are obtained by processing the image of the image elements using the pre-trained Resnet model. For the image features of any type of image element i in the image background, observation target or interference item, Expressed as:

[0037] v i An image representing graphic element i;

[0038] The text features of the image elements are obtained by processing the text information of the image elements using the pre-trained Bert model. For any type of image element i in the image background, observation target or interference item, the text features of the image element i are obtained by processing the text information of the image element using the pre-trained Bert model. Expressed as:

[0039] t i Represents the text information of graph element i.

[0040] Specifically, in the feature extraction module, the feature data of historical users and the data of user-graph element interaction graph of the multi-layer graph convolutional network model are processed to obtain the user interaction matching feature f u The specific method is:

[0041] For any historical user u, obtain its feature data X u , and obtain the image features and text features of each graph element interacted by the historical user u from the historical user training graph element interaction data, and respectively form the interaction image feature sequence of the historical user u and interactive text feature sequences

[0042]

[0043] in, and They represent the image features and text features of the jth graph element interacted by historical user u, j = 1, 2, …, J, where J represents the number of graph elements interacted by historical user u;

[0044] Initialize the feature vector input to the multi-layer graph convolutional network model Concat(·) is a concatenation function; the hierarchical update of the lth layer is then performed as follows:

[0045]

[0046] in, represents the unit embedding matrix of the interaction matrix A, I is the identity matrix; W represents the degree matrix of the node set D composed of each graph element of each historical user in the user-graph element interaction graph G; (l) is the trainable parameter of the lth layer in the multi-layer graph convolutional neural network model, l=1,2,…,L, L represents the total number of layers of the multi-layer graph convolutional neural network model; Softmax(·) represents the normalized exponential function operation; the output of the last layer of the multi-layer graph convolutional neural network model is used as the user interaction matching feature f of the historical user u u ,Right now

[0047] Through the above method, user interaction matching features of different historical users are obtained respectively.

[0048] Specifically, in the feature extraction module, the modal fusion embedding feature extraction process is performed on each graph element through the Transformer model to obtain the modal fusion embedding feature f of each graph element. i The specific method is:

[0049] For any graph element i, its image features and text features Unified as graph element features Right now or Graph element features for graph element i Perform multi-head attention mechanism processing to obtain the multi-head self-attention feature vector of graph element i

[0050]

[0051] in, Graph element features The corresponding query vector, key vector, and value vector, m∈{v,t}; are the linear transformation matrices of query, key, and value in the multi-head self-attention mechanism respectively; d k is the dimension of the attention head; Softmax(·) represents the normalized exponential function operation; Attention(·) represents the attention mechanism operation;

[0052] are the trainable weight vectors corresponding to the query, key, and value in the z-th attention head, respectively; z = 1, 2, …, Z, where Z represents the number of attention heads in the multi-head self-attention mechanism; Concat(·) is the concatenation function;

[0053] Then, the cosine similarity between the multi-head self-attention feature vectors of different graph elements is calculated

[0054]

[0055] in, are the multi-head self-attention feature vectors of graph elements i and j respectively, m∈{v,t}; ||·|| represents modular calculation;

[0056] For graph element i, the first K other graph elements whose cosine similarity values ​​are arranged from large to small constitute the K nearest neighbor graph element set N i-K , and then calculate the image modality embedding weight of image element i by the following formula and text modality embedding weights

[0057]

[0058] in, are the image modality multi-head self-attention feature vector and text modality multi-head self-attention feature vector of graph element j respectively; are the cosine similarities between the image modality multi-head self-attention feature vectors of image element i and image element j, is the cosine similarity between the text modality multi-head self-attention feature vectors of image element i and image element j;

[0059] Then the modal fusion embedding feature f of graph element i is obtained by the following formula: i :

[0060]

[0061] Through the above method, the modal fusion embedding features of different graph elements are obtained respectively.

[0062] Specifically, in the feature extraction module and the comparative learning module, for any graph element i, the matching value fs of the graph element i relative to the historical user u is calculated as follows: ui :

[0063] fs ui =(f u ) T f i ;

[0064] Among them, f u is the user interaction matching feature corresponding to the historical user u, f i is the modal fusion embedding feature of graph element i; in this way, the matching value of each different graph element relative to different historical users is calculated respectively;

[0065] In the contrastive learning module, the constructed loss function is expressed as:

[0066]

[0067] Where D is the node set of each historical user and each graph element; The modal fusion embedding features of the graph elements that represent the interaction relationship with the historical user u, The modality fusion embedding features representing graph elements that have no interaction relationship with historical user u.

[0068] Specifically, the specific method for calculating the matching value of each element relative to the user to be matched is:

[0069] According to the user u to be matched t Feature data X ut , match it with the user-graph element interaction graph G, and determine the user u to be matched t For the K1 matched historical users, the user interaction matching features of the K1 matched historical users are calculated using the trained and optimized multi-layer graph convolutional network model. The optimized Transformer model is then trained to calculate the matching values ​​of each different element relative to the K1 matched historical users, which are used as the matching values ​​of each different element relative to the user to be matched.

[0070] Then, the matching degree values ​​of each type of graphic element are counted separately, and the top K2 graphic elements with the highest matching degree values ​​in the corresponding graphic element type are screened out from each type of graphic element. K2 candidate graphic elements of the picture background type, K2 candidate graphic elements of the observation target type, and K2 candidate graphic elements of the interference item type are obtained to form a subset of matching candidate graphic elements.

[0071] Preferably, in step S4, the total gaze duration is calculated in the following manner:

[0072] By collecting the trainee's eye images, identifying the trainee's eye movement state and pupil state, and judging the gaze state based on the eye movement state and pupil state; if the trainee's eye remains motionless and the pupil state remains unchanged for a duration that reaches the preset minimum gaze duration t0, the trainee is determined to be in a gaze state, and the trainee's total gaze duration t is calculated. g =t0+t x , t x Indicates the time the trainee continues to stare after being judged to be in the staring state.

[0073] This setup allows for a comprehensive assessment of gaze state by simultaneously collecting data on eye movement, gaze point location, and pupil response (rather than a single indicator). This effectively reduces misjudgments. For example, relying solely on gaze point location could lead to misjudgment of gaze as "gaze" due to slight head movements. However, combining pupil constriction (a physiological marker of accommodation) with eye movement stability can filter out false gazes (such as compensatory head fixation). Furthermore, the true state can be captured dynamically. Pupil constriction is a direct physiological signal of ciliary muscle contraction, and combined with gaze stability, it can more accurately reflect the actual accommodation state of the eye muscles. Furthermore, a minimum gaze duration, t0, is introduced. This requires the trainee's eyes to remain motionless and their pupil state unchanged for a predetermined minimum gaze duration, t0 (e.g., set to 2 seconds), before gaze is considered valid. This ensures that the ciliary and extraocular muscles are fully activated in a stable state, preventing ineffective training due to brief gazes. This ensures the stability and effectiveness of gaze state monitoring and ensures that eye muscle training achieves the desired effect at each training stage. In addition, the use of eye tracking technology to monitor the trainee's gaze stability and reaction time in real time can provide instant feedback on the training effect. If it is found that the trainee has difficulty maintaining a stable gaze or the reaction time is too long, it may indicate that the current state is too difficult or too low for the trainee. The training strategy or difficulty can be adjusted in time to ensure the effectiveness of the training and maximize the training results. In addition, based on the specific analysis results of the eye data, the coach or system can dynamically adjust the training plan according to the actual performance of each trainee. For example, for those trainees who show poor gaze stability at a specific distance point, the training time at that distance point can be appropriately extended or additional auxiliary exercises can be provided to better meet individual needs.

[0074] In a second aspect, the present invention provides a method for generating a visual training image, characterized in that the method uses characteristic information of a trainee and training distance information as inputs of a preset image generation network model to generate a visual training image for training the trainee's eye muscle adjustment ability; the image generation network model includes:

[0075] A database module is used to store various graphic elements and their multimodal features used to generate visual training images, as well as all historical user training graphic element interaction data and feature data of each historical user. The types of graphic elements include image background, observation targets, and distractors, each category containing several graphic elements. The multimodal features of the graphic elements include image features and text features. The historical user training graphic element interaction data is used to record the interaction relationships between various graphic elements corresponding to suitable visual training images used by different users in historical training.

[0076] A graph conversion module is used to convert the historical user training graph element interaction data into a user-graph element interaction graph G = (D, E), where each historical user and each graph element constitutes a node in the graph, the node set is marked as D, and the interaction relationship between the historical user node and the graph element node constitutes an edge in the graph, the edge set is marked as E; and based on the user-graph element interaction graph, an interaction matrix of user nodes to graph element nodes is constructed. a ui Represents the interaction relationship between the historical user node u and the graph element node i, a ui =1 indicates that there is an interactive relationship between the historical user node u and the graph element node i, a ui = 0 means there is no interaction relationship between the historical user node u and the graph element node i, N U 、N I Represent the number of historical users and the number of graph elements respectively;

[0077] The feature extraction module includes a multi-layer graph convolutional network model and a Transformer model, which is used to process the feature data of historical users and the data of the user-graph element interaction graph using the multi-layer graph convolutional network model to obtain the user interaction matching feature f between the feature data of historical users and the graph elements. u And the modal fusion embedding feature extraction process is performed on each graph element through the Transformer model to obtain the modal fusion embedding feature f of each graph element i ;

[0078] Contrastive learning module, used to match features f based on historical user feature data and user interactions between graph elements u and the modal fusion embedding feature f of the graph elements i , calculate the matching value fs of different graph elements relative to historical user features ui , and based on the matching value fs ui Constructing a loss function, iteratively training the multi-layer graph convolutional network model and the Transformer model, and updating and optimizing the parameters of the multi-layer graph convolutional network model and the Transformer model through backpropagation to obtain a trained and optimized model;

[0079] The graph element matching module is used to obtain the feature data of the user to be matched, input it into the trained and optimized multi-layer graph convolutional network model, obtain the user interaction matching features between the feature vector of the user to be matched and the graph elements, and calculate the matching degree value of each different element relative to the user to be matched with the trained and optimized Transformer model. According to the matching degree value, a number of candidate graph elements are selected from each type of graph element to form a subset of matching candidate graph elements;

[0080] The visual training image generation module is used to select image elements from the matching candidate image element subset, and based on a preset image generation program, combine the selected image elements according to the trainee's vision data and the training distance to generate a new visual training image.

[0081] This visual training image generation method can be compiled into a computer program using programming techniques and stored on a storage medium. The computer program stored on the storage medium can then be retrieved and executed by a processor to execute the visual training image generation method. The technical advantages of this method have been described above and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] In order to make the purpose, technical solutions and advantages of the invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings, in which:

[0083] Figure 1 This is a schematic diagram of the architecture of the image generation network model in the solution of the present invention.

[0084] Figure 2 This is an example of a visual training diagram in the solution of the present invention.

[0085] Figure 3 The figure is a flow chart of the eye muscle adjustment ability training method of the present invention. DETAILED DESCRIPTION

[0086] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the invention claimed for protection, but only represents selected embodiments of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0087] The present invention will be described in further detail below with reference to the accompanying drawings.

[0088] Considering the impact of individual differences on vision training, common problems with existing eye muscle training programs, such as the limited training parameters and rigid training intensity, are the primary technical issues that need to be addressed. To address this, the present invention proposes a method for training eye muscle accommodation. This method obtains characteristic information of the trainee and training distance information, and uses a preset image generation network model to generate corresponding visual training images for training the trainee's eye muscle accommodation ability.

[0089] In order to better reflect the technical features of the present invention, the image generation network model used in the present invention is described in detail below.

[0090] 1. Image Generation Network Model

[0091] The image generation network model architecture used in this invention is as follows Figure 1 As shown, it includes a database module, a graph conversion module, a feature extraction module, a contrast learning module, a graph element matching module and a visual training graph generation module.

[0092] Among them, the database module is used to store various types of graphic elements and their multimodal features for generating visual training images, and is also used to store all historical user training graphic element interaction data and feature data of each historical user; the types of graphic elements include picture background, observation targets and interference items, and each type contains several graphic elements; the multimodal features of the graphic elements include image features and text features; the historical user training graphic element interaction data is used to record the interaction relationship between various types of graphic elements corresponding to suitable visual training images used by different users in historical training.

[0093] The graph conversion module is used to convert the historical user training graph element interaction data to generate a user-graph element interaction graph G = (D, E), where each historical user and each graph element constitutes a node in the graph, the node set is marked as D, and the interaction relationship between the historical user node and the graph element node constitutes an edge in the graph, the edge set is marked as E; and based on the user-graph element interaction graph, an interaction matrix of user nodes to graph element nodes is constructed. a ui Represents the interaction relationship between the historical user node u and the graph element node i, a ui =1 indicates that there is an interactive relationship between the historical user node u and the graph element node i, a ui = 0 means there is no interaction relationship between the historical user node u and the graph element node i, N U 、N I Represent the number of historical users and the number of graph elements respectively.

[0094] The feature extraction module includes a multi-layer graph convolutional network model and a Transformer model, which is used to process the feature data of historical users and the data of the user-graph element interaction graph using the multi-layer graph convolutional network model to obtain the user interaction matching feature f between the feature data of historical users and the graph elements. u And the modal fusion embedding feature extraction process is performed on each graph element through the Transformer model to obtain the modal fusion embedding feature f of each graph element i .

[0095] The contrastive learning module is used to match features f based on historical user feature data and user interactions between graph elements. u and the modal fusion embedding feature f of the graph elements i , calculate the matching value fs of different graph elements relative to historical user features ui , and based on the matching value fs ui A loss function is constructed, the multi-layer graph convolutional network model and the Transformer model are iteratively trained, and the parameters of the multi-layer graph convolutional network model and the Transformer model are updated and optimized through back propagation to obtain a trained and optimized model.

[0096] The graph element matching module is used to obtain the feature data of the user to be matched, input it into the trained and optimized multi-layer graph convolutional network model, obtain the user interaction matching features between the feature vector of the user to be matched and the graph elements, and combine it with the trained and optimized Transformer model to calculate the matching value of each different element relative to the user to be matched. According to the matching value, several candidate graph elements are selected from each type of graph element to form a matching candidate graph element subset.

[0097] The visual training image generation module is used to select image elements from the matching candidate image element subset, and based on a preset image generation program, the selected image elements are combined and processed according to the trainee's vision data and the training distance to generate a new visual training image.

[0098] It can be seen that the image generation network model achieves highly personalized and intelligent visual training image generation through multi-module collaborative design. Its technical advantages are mainly reflected in the following aspects: First, the database module integrates multimodal graph element features (images and text) and user interaction history data, providing a rich feature expression basis for the model; second, the graph conversion module structures the user-graph element interaction relationship into a graph and constructs an interaction matrix, effectively capturing the complex association pattern between users and graph elements; the feature extraction module combines the multi-layer graph convolutional network and the Transformer model to extract user interaction matching features and element deep fusion features from graph structure data and sequence modal features respectively, realizing efficient fusion of multi-dimensional features; the contrastive learning module establishes the correlation between user features and graph element features through matching degree calculation and loss function optimization, thereby improving the model's adaptability to personalized needs; finally, the graph element matching module screens alternative elements based on the optimized model, and dynamically generates training images based on the trainee's vision data and distance parameters, ensuring that each generated visual training image is highly matched with individual characteristics in terms of element selection and layout rules. This series of technical designs enables the model to accurately capture user needs, dynamically optimize the combination of graphic elements, and significantly improve the accuracy and effectiveness of individual eye muscle adjustment training.

[0099] The present invention, with the help of an image generation network model, can optimize and select the graphic elements for generating visual training images in a targeted manner according to the characteristic information of different trainees, and then generate the visual training images for each training session based on the distance from the observed target image during specific training and based on a preset image generation program. Since the image generation network model is trained based on the use and interaction of historical users' visual training images, the model constructs the interaction relationship between historical users and graphic elements through the user-graphic element interaction graph, and uses a multi-layer graph convolutional network model to perform multi-dimensional deep extraction of the interaction features between historical users and the graphic elements they interact with. The attention mechanism of the Transformer model is used to perform multi-dimensional deep extraction of the deep features of various graphic elements, and with the help of the contrast learning mechanism provided by the contrast learning module, the connection between the two aspects of deep features is established. Therefore, the trained image generation network model can match and generate appropriate visual training images according to the characteristic information of the individual trainees, thereby improving the eye muscle adjustment training effect for the individual trainees.

[0100] Below, each module in the image generation network model adopted by the invention is described respectively.

[0101] 1.1 Database Module

[0102] The database module is the core data storage module of the image generation network model. It stores all available types of image elements used to generate visual training images, as well as the multimodal features corresponding to each image element, the feature data of all historical users, and all historical user training image element interaction data.

[0103] There are three main types of graphic elements used to generate visual training images, including picture backgrounds, observation targets, and interference items. Among them, picture backgrounds are used as background image elements of visual training images, and can be monochrome background images, two-color gradient background images, multi-color splicing background images, etc. Observation targets are target objects that trainees need to look at and identify during vision training. They can be text, characters, or object graphics of different colors (such as shapes, animal pictures, object pictures, etc.). They need to have clearer recognition features, but also require a certain degree of visual concentration to be identified; interference items are certain interference factors in visual training images. They can be text, characters, graphics, textures, color blocks, etc. of different colors. They are used to increase the difficulty of trainees in identifying observation targets to a certain extent, promote the adjustment of their eye muscles, and adjust the training intensity together with the picture background and observation targets.

[0104] Each graphic element has its own multimodal information, including image modal data and text modal information. The image modal data of a graphic element is the RGB image of the graphic element itself, which is stored in the image data format. The text modal information of a graphic element is text description information used to illustrate the characteristics of the graphic element, such as the type label of the graphic element (belonging to the image background, observation target or interference item), image resolution, image average brightness value, color composition description of the image background, font or content description of the observation target, font or content description of the interference item, etc.; the text modal information of the graphic element is stored in the text data format. The multimodal information of the graphic element is used to construct the multimodal features corresponding to each graphic element and is stored in the database module.

[0105] Historical users also have their own characteristic information, including the trainee's gender, age group, vision data, eye usage time per unit time period, blood pressure test values ​​(relevant research has shown that high blood pressure may have a certain impact on intraocular pressure and vision, so blood pressure test values ​​can also be included as user vision-related characteristic information for consideration), etc. Historical users are users who have previously been trainees. Their characteristic information is also stored in text data format and is used to construct historical user characteristic data and stored in the database module.

[0106] In specific application implementations, the text modal information of graphic elements and the characteristic information of historical users can be entered through the information entry method commonly used in the existing technology, and then stored in text data format; for example, keyboard or touch screen input can be used for entry, options can be set for selection operation entry, and voice entry can be used, with the help of large language models (such as ChatGPT, DeepSeek, Tongyi, etc.) to perform semantic recognition processing into text information for entry.

[0107] After obtaining the multimodal information of graph elements and the characteristic information of users, they need to be processed through different pre-trained models to convert the multimodal features of graph elements and the characteristic data of historical users and store them in the database module.

[0108] Specifically, the feature data of historical users is obtained by embedding the feature information of historical users, which can be expressed as X u ={x u,1 ,x u,2 ,…,x u,n ,…,x u,N}, x u,n Represents the word embedding value of the nth user feature of historical user u, and N represents the total number of user features.

[0109] The image features of the image elements are obtained by processing the image of the image elements using the pre-trained Resnet model. For any type of image feature of an element i in the image background, observation target or interference item, It can be expressed as:

[0110] v i An image representing graphic element i;

[0111] The text features of the image elements are obtained by processing the text information of the image elements using the pre-trained Bert model. For any type of image element i in the image background, observation target or interference item, the text features of the image element i are obtained by processing the text information of the image elements using the pre-trained Bert model. It can be expressed as:

[0112] t i Represents the text information of graph element i.

[0113] The purpose of this conversion process is to enable the multimodal features of the graph elements and the feature data of historical users to be correctly identified and feature extracted by the neural network model in subsequent processing.

[0114] 1.2 Graph Conversion Module

[0115] The graph conversion module is used to convert and generate a user-graph element interaction graph based on the historical user training graph element interaction data. In specific applications, a user-graph element interaction graph can be constructed based on a Bayesian network, which can be expressed as G = (D, E); wherein each historical user and each graph element constitutes a node in the graph, and the node set is marked as D; the interaction relationship between the historical user node and the graph element node constitutes the edge in the graph. If there is a correlation or interaction relationship between the historical user nodes, the graph element nodes, and the historical user nodes and the graph element nodes, there will be an edge between the nodes; otherwise, there will be no edge, and the edge set is marked as E.

[0116] Based on the user-graph element interaction graph G = (D, E), we can construct the interaction matrix of user nodes to graph element nodes, which can be expressed as a ui Represents the interaction relationship between the historical user node u and the graph element node i, a ui =1 indicates that there is an interactive relationship between the historical user node u and the graph element node i, a ui = 0 means there is no interaction relationship between the historical user node u and the graph element node i, N U 、N I Represent the number of historical users and the number of graph elements respectively.

[0117] The user-graph element interaction graph G=(D, E) can be continuously updated based on the trainer's training data. The graph elements in the vision training graph that have better training effects on the trainer are marked with interaction relationships, and the trainer's feature information is converted into feature data of a new historical user. A record of the interaction relationship between the graph elements used in the vision training graph with better training effects and the new historical user is established. With the help of the graph conversion module, the user-graph element interaction graph can be continuously updated.

[0118] 1.3 Feature Extraction Module

[0119] A multi-layer graph convolutional network model and a Transformer model are set up in the feature extraction module. The purpose is to use the multi-layer graph convolutional network model to perform multi-dimensional deep extraction of the interaction features between historical users and their graph elements, and use the attention mechanism of the Transformer model to perform multi-dimensional deep extraction of the deep features of various graph elements, so as to facilitate the subsequent use of the two aspects of deep features to establish matching associations.

[0120] 1.3.1 Multi-layer graph convolutional network model

[0121] In the specific implementation, the multi-layer graph convolutional network model can adopt the LightGCN network model. The LightGCN network model is a lightweight GCN network (Graph Convolutional Network) with simplified architecture and improved performance. It can be used to extract the deep features between the feature data of historical users and the graph elements, and construct the user interaction matching feature f u , and uses these deep features to establish implicit associations between historical user feature data and the multimodal features of graph elements. In specific applications, multi-layer graph convolutional network models typically use 3 layers (or a maximum of 4 layers), which is sufficient to achieve optimal feature extraction performance. Further increasing the number of layers will not significantly improve the performance, but will increase the computational workload of model training and processing.

[0122] Specifically, for any historical user u, obtain its feature data X u , and obtain the image features and text features of each graph element interacted by the historical user u from the historical user training graph element interaction data, and respectively form the interaction image feature sequence of the historical user u and interactive text feature sequences

[0123]

[0124] in, and They represent the image features and text features of the jth graph element interacted by historical user u, j = 1, 2, …, J, and J represents the number of graph elements interacted by historical user u.

[0125] Initialize the feature vector input to the multi-layer graph convolutional network model Concat(·) is a concatenation function; the hierarchical update of the lth layer is then performed as follows:

[0126]

[0127] in, represents the unit embedding matrix of the interaction matrix A, I is the identity matrix; W represents the degree matrix of the node set D composed of each graph element of each historical user in the user-graph element interaction graph G; (l) is the trainable parameter of the lth layer in the multi-layer graph convolutional neural network model, l=1,2,…,L, L represents the total number of layers of the multi-layer graph convolutional neural network model; Softmax(·) represents the normalized exponential function operation; the output of the last layer of the multi-layer graph convolutional neural network model is used as the user interaction matching feature f of the historical user u u ,Right now

[0128] Through the above method, the user interaction matching features f of different historical users can be obtained respectively. u ,u=1,2,…,N U , N U Indicates the number of historical users.

[0129] 1.3.2 Transformer Model

[0130] The Transformer model is used to extract the modal fusion embedding features of each graph element. The multi-head attention mechanism of the Transformer model is mainly used to deeply extract the characteristics of the graph elements themselves and construct the modal fusion embedding feature f of the graph elements. i , in order to be used in combination with the subsequent contrast learning mechanism to explore the characteristics of the image elements' own characteristics for the trainees' visual training.

[0131] Specifically, for any image element i, for the convenience of expression, its image feature and text features Unified as graph element features Right now or

[0132] Graph element features for graph element i Perform multi-head attention mechanism processing to obtain the multi-head self-attention feature vector of graph element i

[0133]

[0134] in, Graph element features The corresponding query vector, key vector, and value vector, m∈{v,t}; are the linear transformation matrices of query, key, and value in the multi-head self-attention mechanism respectively; d k is the dimension of the attention head; Softmax(·) represents the normalized exponential function operation; Attention(·) represents the attention mechanism operation; are the trainable weight vectors corresponding to the query, key, and value in the z-th attention head respectively; z = 1, 2, …, Z, where Z represents the number of attention heads in the multi-head self-attention mechanism; Concat(·) is the concatenation function.

[0135] Then, the cosine similarity between the multi-head self-attention feature vectors of different graph elements is calculated

[0136]

[0137] in, are the multi-head self-attention feature vectors of graph elements i and j respectively, m∈{v,t}; ||·|| represents modular calculation.

[0138] Through the above processing, we can get the multi-head self-attention feature vector of any image element i in the image feature mode And in the text modality multi-head self-attention feature vector And the cosine similarity between the image element i and any other image element j in the image feature modality multi-head self-attention feature vector Cosine similarity between multi-head self-attention feature vectors in text modality

[0139] Next, for the graph element i, the first K other graph elements whose cosine similarity values ​​are arranged from large to small are taken to form the K nearest neighbor graph element set N i-K , and then calculate the image modality embedding weight of image element i by the following formula and text modality embedding weights

[0140]

[0141] in, are the image modality multi-head self-attention feature vector and text modality multi-head self-attention feature vector of graph element j respectively; are the cosine similarities between the image modality multi-head self-attention feature vectors of image element i and image element j, is the cosine similarity between the text modality multi-head self-attention feature vectors of image element i and image element j.

[0142] Then the modal fusion embedding feature f of graph element i is obtained by the following formula: i :

[0143]

[0144] Through the above method, the modal fusion embedding features f of different graph elements can be obtained respectively. i , i=1,2,…,N I , N I Indicates the number of graph elements.

[0145] 1.4 Comparative Learning Module

[0146] The contrastive learning module uses the contrastive learning mechanism to first match the feature f based on the historical user feature data and the user interaction between the graph elements. u and the modal fusion embedding feature f of the graph elements i, calculate the matching value fs of different graph elements relative to historical user features ui , and based on the matching value fs ui , a loss function is constructed with the help of positive and negative sample pairs, and the multi-layer graph convolutional network model and the Transformer model are iteratively trained using a contrastive learning mechanism, and the parameters of the multi-layer graph convolutional network model and the Transformer model are updated and optimized through back propagation.

[0147] Specifically, for any graph element i, the matching value fs of the graph element i relative to the historical user u is calculated as follows: ui :

[0148] fs ui =(f u ) T f i ;

[0149] Among them, f u is the user interaction matching feature corresponding to the historical user u, f i is the modal fusion embedding feature of graph element i.

[0150] In this way, the matching degree values ​​fs for different historical users can be calculated separately. ui ,u=1,2,…,N U , i=1,2,…,N I , N U 、N I Represent the number of historical users and the number of graph elements respectively.

[0151] Based on the matching value fs ui The constructed loss function is expressed as:

[0152]

[0153] Where D is the node set of each historical user and each graph element; The modal fusion embedding features of the graph elements that represent the interaction relationship with the historical user u, i.e., the positive samples; The modality fusion embedding features of graph elements that do not have an interactive relationship with historical user u are negative samples.

[0154] The contrastive learning mechanism uses a loss function to measure the loss of similar and heterogeneous samples, optimizing model parameters by maximizing the similarity of positive sample pairs and minimizing the similarity of negative sample pairs. During contrastive learning, for positive sample pairs, the smaller the distance, the smaller the loss value, while for negative sample pairs, the larger the distance, the smaller the loss value. The optimization results of the contrastive learning module are used to adjust the weights and associated parameters of the multi-layer graph convolutional network model and the Transformer model, more accurately establishing the matching degree between historical users with different characteristics and the graph elements required for their appropriate visual training graphs. By constructing positive and negative sample pairs, calculating sample distances, optimizing the loss function, and optimizing and adjusting the deep features focused on by the matching mechanism, the matching performance of the multi-layer graph convolutional network model and the Transformer model is improved.

[0155] 1.5 Graph Element Matching Module

[0156] The graph element matching module is used to obtain the feature data of the user to be matched, input it into the trained and optimized multi-layer graph convolutional network model, obtain the user interaction matching features between the feature vector of the user to be matched and the graph elements, and combine it with the trained and optimized Transformer model to calculate the matching value of each different element relative to the user to be matched. According to the matching value, several candidate graph elements are screened from each type of graph element to form a matching candidate graph element subset, from which the visual training graph generation module selects graph elements to generate a visual training graph suitable for the user to be matched.

[0157] The users to be matched here refer to new trainees. For new trainees, we also need to collect their gender, age group, vision data, eye usage time per unit time, blood pressure test value and other information as feature information. Through the embedding word embedding process, we can get the corresponding feature data, which can be expressed as X ut ={x ut,1 ,x ut,2 ,…,x ut,n ,…,x ut,N}, x ut,n Indicates the user u to be matched t The word embedding value of the nth user feature, N represents the total number of user features.

[0158] In a specific implementation, the specific method in which the graph element matching module calculates the matching value of each different element relative to the user to be matched is:

[0159] According to the user u to be matched t Feature data X ut , match it with the user-graph element interaction graph G, and determine the user u to be matched tFor the K1 matched historical users, the user interaction matching features of the K1 matched historical users are calculated using the trained and optimized multi-layer graph convolutional network model, and the trained and optimized Transformer model is used to calculate the matching values ​​of each different element relative to the K1 matched historical users, which are used as the matching values ​​of each different element relative to the user to be matched.

[0160] Then, the matching degree values ​​of each type of graphic element are counted separately, and the top K2 graphic elements with the highest matching degree values ​​in the corresponding graphic element type are screened out from each type of graphic element. K2 candidate graphic elements of the picture background type, K2 candidate graphic elements of the observation target type, and K2 candidate graphic elements of the interference item type are obtained to form a subset of matching candidate graphic elements.

[0161] During the above processing, the values ​​of K1 and K2 can be set according to the actual scale of the number of historical users, the scale of the number of graphic elements, and the required number of graphic elements of the matching alternative graphic element subset to be constructed; generally speaking, if the scale of the number of historical users and the scale of the number of graphic elements are large, the required number of graphic elements of the matching alternative graphic element subset to be constructed is relatively large, so that the system can generate different visual training images for new trainees in more eye training conditions with the same conditions, then the values ​​of K1 and K2 can be set to be larger; and vice versa.

[0162] 1.6 Visual Training Image Generation Module

[0163] The visual training image generation module is used to select image elements from the matching candidate image element subset, and based on a preset image generation program, the selected image elements are combined and processed according to the trainee's vision data and the training distance to generate a new visual training image.

[0164] In specific applications, the visual training image generation module can be set to randomly select image background, observation target and interference image elements from the matching candidate image element subset to generate a new visual training image; therefore, the more image elements included in the matching candidate image element subset (the larger the value of K2), the more different visual training images can be generated (there can be more K2 3 different combinations).

[0165] The rules for generating visual training images using picture background, observation target and interference items can be implemented through preset program programming. Among them, it is necessary to consider the layout relationship between the three picture elements of picture background, observation target and interference items, and combine the trainee's vision data and training distance to set the display size, brightness and other parameters of the observation target and interference items to adapt them to the required eye training effect.

[0166] For example, as an implementation method of a preset image generation program, after selecting three image elements, namely, picture background, observation target and interference items, from a subset of matching alternative image elements, the observation target is first placed in the central area of ​​the picture background, and the trainee's vision data and the training distance are used as measurement parameters. The display size and brightness of the observation target are determined according to a preset measurement algorithm. Then, the interference item elements are distributed in the surrounding area of ​​the observation target, and the trainee's vision data and the training distance are used as measurement parameters. The display size and brightness of the interference items themselves, as well as the interval distance between the interference items and the observation target are determined according to a preset measurement algorithm. Finally, they are combined into a visual training image. Figure 2 This figure shows an example of a visual training image generated using a pre-set image generation program. The image has a blue background, with the white image occupying a large central area as the observation target, and the orange and orange-red Chinese and English characters arranged around the observation target as distractors.

[0167] It should be noted that, in the design of the visual training diagram of the present invention, interference items are specially set. By adding interference items, the visual training diagram can better simulate the visual environment with complex object backgrounds in real life, thereby helping the trainee to improve their visual discrimination ability when facing cluttered backgrounds or similar objects. This method is particularly important for vision training that requires distinguishing subtle differences. In addition, due to the presence of interference items, the trainee needs to focus more on finding the target object, which helps to train their ability to eliminate interference from irrelevant information, thereby strengthening the concentration and persistence of attention. In addition, analyzing and generating trial training diagrams based on the content, color, position and other features of the interference items can make the training materials closer to the visual challenges of the real world. Using such visual training diagrams for vision training can help improve the trainee's visual processing ability and reaction speed in complex or changing environments.

[0168] 1.7 Overview of Image Generation Network Models

[0169] With the help of the image generation network model of the present invention, it is possible to dynamically select image elements based on the trainee's vision data and related characteristics. For example, a target with higher contrast or brighter color can be generated for amblyopic eyes to enhance their recognition difficulty and improve their adjustment ability in a targeted manner. At the same time, it can also adapt to the visual needs of training at different distances and perform algorithmic adjustments on the image elements. For example, a larger and clearer target is generated for a distant point (by adjusting the RGB background brightness and target color contrast) to simulate a real distant view scene, while a target with complex details (such as small fonts or low-contrast patterns) is generated for a close point to enhance the close adjustment ability. Thus, a targeted visual training map is generated, and then the training intensity is dynamically adjusted according to the trainee's personalized characteristics, avoiding the training efficiency bottleneck caused by fixed parameters and overcoming the problem of fixed images and single parameters in traditional visual training. By dynamically generating different visual training maps in this way, the personalized and scientific gradient design of the trainee's eye muscle adjustment training can be significantly improved.

[0170] From the perspective of technical implementation, the method of generating a visual training image for training the trainee's eye muscle adjustment ability by using the trainee's characteristic information and the training distance information as the input of the above-mentioned preset image generation network model can be compiled into a computer program with the help of programming technology, stored in a storage medium, and the computer program stored in the storage medium is called and executed by the processor to run the above-mentioned visual training image generation method.

[0171] The storage medium may be any combination of one or more computer-readable media. The computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium may include: an electrical connection having one or more wires, a portable computer disk, a hard disk, a random access memory, a read-only memory, an erasable programmable read-only memory, an optical fiber, a portable compact disk read-only memory, an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0172] The code for the computer program for performing the operations of the present invention can be written in one or more programming languages, or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer can be connected to the user's computer via any type of network, including a local area network or a wide area network, or can be connected to an external computer.

[0173] 2. Eye muscle adjustment ability training method process

[0174] After solving the problems of single training parameters and fixed training intensity in the existing eye muscle training program, the next step is to solve the scientific nature of the training program design and the monitorability of the training status. Therefore, it is necessary to design a rational training method process. The process of the eye muscle adjustment ability training method proposed by the present invention is as follows: Figure 3 As shown, the specific steps include:

[0175] S1. Preset distance points for training the trainee; there are multiple distance points, and the distance between each distance point and the observation target is different.

[0176] In the solution of the present invention, the visual training image can be displayed through a display (such as an electronic display screen). The trainee needs to observe the visual training image from a certain distance and perform eye muscle adjustment training by focusing on the observation target in the image.

[0177] In specific implementation, 11 distance points can be set, corresponding to the initial distance points of different visual acuity values. The naked eye visual acuity values ​​corresponding to each distance point and their estimated distance decibels relative to the visual training chart are shown in the following table. In this way, different training visual distance difficulties can be matched to the visual acuity levels of different trainees, ensuring that each trainee can start training from a distance suitable for their current visual condition, avoiding the problem of poor training results due to starting points that are too difficult or too easy. In addition, as the training process progresses and the visual condition improves, the trainee's starting position for each subsequent training session can gradually transition from a closer distance point corresponding to lower visual acuity to a farther distance point corresponding to higher visual acuity, which is more conducive to gradually enhancing the eye muscle adjustment ability and promoting a steady improvement in the effect of eye muscle training. Furthermore, the 11 distance points are designed to cover the common visual acuity range (from 4.0 to 5.0). If the naked eye visual acuity is lower than 4.0, which is also quite common during training, the solution is to directly select the first point and make fine adjustments based on the actual situation. Because the first point is already very close, the eye with the worst vision can be fine-tuned by about 20 cm at most before training can begin. This way, whether you have high myopia or near-normal vision, you can find the appropriate training starting point and progression path, ensuring the applicability and inclusiveness of the training program.

[0178]

[0179] This mathematical model directly links visual acuity and training distance, providing a scientifically calculated optimal starting distance for each person with varying vision levels. This approach ensures a highly personalized and precise training plan, allowing each person to begin training at the distance that best suits them.

[0180] S2. Acquire characteristic information of the trainee, wherein the characteristic information of the trainee includes the trainee's vision data, match the initial distance points based on the trainee's vision data, and determine the initial training distance.

[0181] In this step, the specific method of matching the initial distance points is:

[0182] S201, based on the visual acuity value V of the weak-sighted eye in the trainee's visual acuity data a , the trainer's initial estimated training distance L is obtained as follows t0 :

[0183] L t0 =5V a -19.4;

[0184] If the visual acuity of the low-vision eye is less than or equal to 4.0, then V a =4.0; Initial estimated training distance L t0The unit is meter;

[0185] S202: Find the distance value to the observed target and the initial estimated training distance L in the preset distance point. t0 The closest distance point is used as the initial distance point for matching.

[0186] In specific implementation, the visual acuity value V of the trainee's weak eye can be used as the calculation input. a , to match the initial distance point L t0 The weaker eye refers to the eye with the lower naked eye vision. This ensures that each trainee begins training at a distance that best suits their current vision. Matching the weaker eye to the starting distance avoids overloading the weaker eye or preventing it from seeing the target clearly due to a starting distance that is too far, thus providing each trainee with a personalized starting point for training. By prioritizing the needs of the weaker eye, more focused training can be provided on the weaker eye, helping to bridge the visual gap between the two eyes and promote balanced development of binocular visual function. This is particularly important for preventing further deterioration of the weaker eye and improving overall vision. In addition, choosing an initial training distance suitable for the weaker eye not only improves the effectiveness of training, but also reduces the risk of visual fatigue or other potential eye problems caused by excessive challenge. This makes the training process safer and more reliable, and promotes long-term persistence.

[0187] S3. Based on the trainee's feature information and the training distance, a preset image generation network model is used to generate a corresponding visual training image.

[0188] The preceding content has already provided a detailed introduction to the image generation network model of the present invention. In this step, the image generation network model enables targeted optimization and selection of image elements for generating visual training images based on the individual characteristics of the trainees. Furthermore, based on the distance from the target image during the specific training session, a pre-set image generation program is used to generate the visual training images for each training session. Since the image generation network model is trained based on the historical users' use and interaction with visual training graphs, the model constructs the interaction relationship between historical users and graph elements through the user-graph element interaction graph. A multi-layer graph convolutional network model is used to perform multi-dimensional deep extraction of the interaction features between historical users and the graph elements they interacted with. The attention mechanism of the Transformer model is used to perform multi-dimensional deep extraction of the deep features of various graph elements. With the help of the contrastive learning mechanism provided by the contrastive learning module, the connection between the two aspects of deep features is established. Therefore, the trained image generation network model can match and generate appropriate visual training graphs according to the individual feature information of the trainee, dynamically adjust the training intensity, avoid the training efficiency bottleneck caused by fixed parameters, overcome the problems of fixed images and single parameters in traditional visual training, and improve the eye muscle adjustment training effect for individual trainees.

[0189] S4. Use the generated visual training chart to train the trainee at corresponding distance points in the order of the weak-vision eye, the strong-vision eye, and both eyes; the weak-vision eye is the eye with lower naked eye vision of the two eyes, and the strong-vision eye is the eye with stronger naked eye vision of the two eyes. If the naked eye vision of the trainee's two eyes is the same, one eye can be used as the weak-vision eye and the other eye as the strong-vision eye.

[0190] During the training process, after the total gaze time of the weak-vision eye, the strong-vision eye, and both eyes at the distance point reaches the preset time t, the trainee is asked to move to an adjacent distance point farther from the observation target and return to S3 until the training of the distance point farthest from the observation target is completed or the total training time reaches the preset total time T0, and the training ends.

[0191] During this training step, after completing training at a distance point, the trainee can determine whether the visual training image meets the expected requirements based on their feedback and the actual vision test results after training. If so, the elements in the visual training image can be labeled with interaction relationships corresponding to the trainee's characteristic information. This is used to continuously update the user-element interaction map stored in the map conversion module of the image generation network model, ensuring that the quality of the generated visual training image can be continuously optimized, thereby improving the effectiveness and relevance of the training. This mechanism introduces a feedback loop that enables the image generation network model to dynamically adjust and improve based on real-time data (i.e., the actual performance of the visual training image), enhancing the system's adaptability to different scenarios.

[0192] In specific implementation, the method for judging whether the visual training chart meets the preset requirements is that if the trainee completes the training at the corresponding distance point using the visual training chart, and the training time used for the low vision eye is within the preset optimal training time range, then the preset requirements are met.

[0193] In this way, by measuring the time required for the low vision eye to complete training using a specific visual training image, the effectiveness of the visual training image for the target individual can be accurately evaluated. If the training time falls within the optimal training duration range, it means that the design of the visual training image can effectively stimulate the eye muscles for adjustment, and it is neither too simple nor too difficult. Within the optimal training duration range, it means that the intensity of the visual training image is appropriate, which can effectively exercise the eye muscles without causing excessive fatigue. This not only improves the effect of a single training session, but also maintains a good training state for a long time and promotes vision improvement. Conversely, if the total duration is too short, it means that the difficulty is too low. If the total duration is too long, it means that the difficulty is too high and it is difficult to ensure a sustained training effect. Using a visual training image that meets the above preset requirements to optimize the training of the image generation network model can ensure the effectiveness of the dynamic adjustment and improvement of the image generation network model.

[0194] In addition, in this step, the total gaze duration is calculated as follows: by collecting the trainee's eye image, identifying the trainee's eye movement state and pupil state, and judging the gaze state based on the eye movement state and pupil state; if the trainee's eye remains motionless and the pupil state remains unchanged for a duration that reaches a preset minimum gaze duration t0, the trainee is determined to be in a gaze state, and the trainee's total gaze duration t is calculated. g =t0+t x , t x Indicates the time the trainee continues to stare after being judged to be in the staring state.

[0195] In specific implementations, the device used to capture the trainee's eye images can be an eye tracker or other instrument. The technology for identifying eye movement and pupil status using captured eye images is relatively mature and widely used in many fields (e.g., concentration testing, driver fatigue monitoring, and other technical scenarios). This mature technology can also be used in the present invention.

[0196] By simultaneously collecting eye movement and pupil reaction data (rather than a single indicator), the gaze state can be comprehensively determined. This can effectively reduce misjudgments of whether the trainee is in a gaze state. For example, if the judgment is based solely on eye movement status, the trainee's eye direction may not move, but may not yet focus on the visual training chart, and may be mistakenly judged as a gaze state. If the judgment is based solely on pupil status, the trainee's pupil visual distance may not be adjusted, but the visual observation point may be moving and not fixed, and may be mistakenly judged as a gaze state. Pupil constriction is a direct physiological signal of ciliary muscle contraction. Therefore, combining eye movement status and pupil status to determine gaze stability in the gaze state can more accurately reflect the actual adjustment state of the eye muscles. Furthermore, a minimum gaze duration t0 is introduced. Effective gaze is only considered effective if the participant's eyes remain still and their pupils remain unchanged for a predetermined minimum gaze duration t0 (e.g., 2 seconds). This ensures that the ciliary and extraocular muscles are fully activated in a stable state, preventing ineffective training caused by brief gazes. This ensures stable and effective gaze monitoring and ensures that eye muscle training achieves the desired effect at each training stage. Furthermore, eye tracking technology is used to monitor the participant's gaze stability and reaction time in real time, providing immediate feedback on training effectiveness. If a participant struggles to maintain a stable gaze or takes too long to react, this may indicate that the current state is too challenging for the participant. Training strategies or difficulty levels can be adjusted promptly to ensure effectiveness and maximize training outcomes. Furthermore, based on the specific analysis of eye data, coaches or the system can dynamically adjust training plans based on each participant's actual performance. For example, for participants who exhibit poor gaze stability at a specific distance, training time at that distance can be appropriately extended or additional supplementary exercises can be provided to better meet individual needs.

[0197] In specific implementation, before the trainee is trained at the corresponding distance point, a distance correction analysis process can also be performed. The process includes: letting the trainee's low-vision eye look at the observation target at the distance point, and collecting the trainee's eye data to determine whether the trainee can see the observation target clearly within a preset time. If so, continue the subsequent process of the training point; if not, let the trainee move forward a preset adjustment distance and re-perform the above-mentioned judgment process; the adjustment distance is less than the distance between two adjacent distance points.

[0198] This method allows for precise distance adjustment based on individual differences by having the participant's weaker eye focus on a target at a specific distance. Whether the target can be clearly seen within a preset timeframe determines whether the distance needs to be adjusted. Compared to fixed training distances, this personalized distance adjustment better matches each participant's actual visual ability, helping to improve training effectiveness and comfort. It also ensures appropriate initial training conditions. Conducting a distance correction analysis before formal training ensures that participants begin training at an appropriately challenging starting point. If the current distance is found to be too challenging for the participant, the distance is shortened to make the task achievable, thus avoiding frustration or ineffective training caused by starting too difficult. If the target is initially judged to be unsightly, the participant is reassessed by moving forward a preset adjustment distance that is less than the distance between adjacent distance points. This process allows participants to gradually approach the training distance that is most suitable for them. This method balances training challenge with feasibility, helping to gradually improve eye muscle accommodation. This approach to distance correction not only improves the adaptability of training programs but also reduces potential eye damage from excessive strain or discomfort. It provides a gentler and more effective way to discover the optimal training distance for an individual, thereby promoting healthy vision development.

[0199] This method matches initial distance points based on the trainee's vision data, generates a corresponding visual training map based on the current distance points and the trainee's characteristic data, and dynamically adjusts the training intensity, achieving a personalized training plan. Compared to traditional, standardized training methods, this method more accurately meets the needs of different individuals, avoiding poor results or overtraining due to inappropriate training intensity. Furthermore, during training, eye data from the trainee's current training eye is collected to determine whether it is in a gaze state, allowing real-time monitoring of the training process. This approach not only ensures the effectiveness of training but also allows for timely adjustments to training strategies to address any potential issues. In contrast, traditional training methods lack such real-time feedback mechanisms, making it difficult to ensure training quality. Furthermore, this method not only focuses on training a single eye (either the weak or strong eye) but also emphasizes the coordinated development of both eyes. This comprehensive training approach helps comprehensively improve the regulatory capacity of eye muscles and promotes the coordinated development of binocular vision. Traditional training methods, on the other hand, often focus on monocular training, neglecting the importance of binocular coordination. Furthermore, by setting up multiple distance points and training from near to far, this method adheres to the principle of gradually increasing training difficulty. This helps trainees gradually adapt to higher visual challenges and strengthen the eye muscles' ability to adjust. Compared to a static training environment, this method can better stimulate the potential of eye muscles and improve training effectiveness.

[0200] Practical application shows that this method has a significant training effect. This is mainly reflected in the clear training effect. For poor vision caused by simple myopia, astigmatism, amblyopia, and other non-organic eye diseases, the implementation effect has achieved a 100% rapid improvement in vision, and no cases of insignificant effect after training have been found. In addition, the training improvement speed is fast. Under normal circumstances, after the first training, the trainees' vision improves by an average of more than 0.3. After about a week of training, more than 90% of the trainees achieve a vision of more than 5.0 after training. In addition, this method is not only effective for myopia, but also has significant effects on astigmatism and amblyopia. Practice has shown that it is generally applicable to congenital amblyopia, and it is fast and does not rebound.

[0201] In general, compared with the prior art, the present invention has the following beneficial effects:

[0202] 1. Personalized Training Intensity Adjustment. This method matches the initial distance points based on the trainee's vision data, generates a corresponding visual training map based on the current distance points and the trainee's characteristic data, and dynamically adjusts the training intensity to achieve a personalized training plan. Compared with traditional, uniformly standardized training methods, this method can more accurately meet the needs of different individuals and avoid problems such as poor results or overtraining due to inappropriate training intensity.

[0203] 2. Real-time monitoring and feedback mechanism. During training, the system collects eye data from the trainee's training eye and determines whether it is in a fixed gaze state. This allows for real-time monitoring of the training process. This not only ensures the effectiveness of training but also allows for timely adjustments to training strategies to address any potential issues. In contrast, traditional training methods lack this real-time feedback mechanism, making it difficult to guarantee training quality.

[0204] 3. Comprehensive coverage of binocular vision. This method not only focuses on training one eye (the weak-vision eye and the strong-vision eye), but also emphasizes the coordinated training of both eyes. This comprehensive training method helps to comprehensively improve the adjustment ability of eye muscles and promote the coordinated development of binocular vision. Traditional training methods often focus on monocular training and ignore the importance of binocular coordination.

[0205] 4. A scientific design that gradually increases difficulty. By setting multiple distance points at different distances and conducting training from near to far, this method adheres to the principle of gradually increasing training difficulty. This helps trainees gradually adapt to higher visual challenges and enhance the adjustment ability of eye muscles. Compared to a static training environment, this method can better stimulate the potential of eye muscles and improve training results.

[0206] In summary, this method can dynamically adjust training intensity based on individual vision, effectively monitor training status, and significantly improve eye muscle regulation. Through personalized training intensity adjustment, real-time monitoring and feedback mechanisms, comprehensive coverage of binocular vision, and a scientific design that gradually increases difficulty, this method overcomes the limitations of traditional eye muscle training methods, such as uncertain training results and slow training speed, caused by insufficient individual targeting, difficulty quantifying training results, and a lack of scientific and systematic training. It provides a more scientific, systematic, and effective solution. This not only helps improve the trainee's eye muscle regulation ability and aids in the restoration of vision, but also prevents or slows the progression of myopia. It has important application value and broad prospects for promotion.

[0207] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the technical solutions. Those skilled in the art should understand that modifications or equivalent replacements of the technical solutions of the present invention that do not depart from the purpose and scope of the technical solutions of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A method for training eye muscle regulation ability, characterized in that: By obtaining the trainee's characteristic information and training distance information, a preset image generation network model is used to generate the corresponding visual training image, which is used to train the trainee's eye muscle adjustment ability.

2. The eye muscle adjustment ability training method according to claim 1, wherein: The method specifically comprises the following steps: S1. Preset distance points for training the trainee; there are multiple distance points, each at a different distance from the observation target; S2. Acquire characteristic information of the trainee, wherein the characteristic information of the trainee includes the trainee's vision data, match the initial distance points based on the trainee's vision data, and determine the initial training distance; S3, based on the trainee's feature information and the training distance, using a preset image generation network model to generate a corresponding visual training image; S4. Using the generated visual training chart, the trainee is trained at corresponding distance points in the order of the weak-vision eye, the strong-vision eye, and both eyes; the weak-vision eye is the eye with the lower naked eye vision of the two eyes, and the strong-vision eye is the eye with the higher naked eye vision of the two eyes. If the trainee's naked eye vision of both eyes is the same, one eye is used as the weak-vision eye and the other eye is used as the strong-vision eye; During the training process, after the total gaze time of the weak-vision eye, the strong-vision eye, and both eyes at the distance point reaches the preset time t, the trainee is asked to move to an adjacent distance point farther from the observation target and return to S3 until the training of the distance point farthest from the observation target is completed or the total training time reaches the preset total time T0, and the training ends.

3. The eye muscle adjustment ability training method according to claim 2, wherein: In step S2, the specific method of matching the initial distance point based on the trainee's vision data is as follows: S201, based on the visual acuity value V of the weak-sighted eye in the trainee's visual acuity data a , the trainer's initial estimated training distance L is obtained as follows t0 : L t0 =5V a -19.4; If the visual acuity of the low-vision eye is less than or equal to 4.0, then V a =4.0; Initial estimated training distance L t0 The unit is meter; S202: Find the distance value to the observed target and the initial estimated training distance L in the preset distance point. t0 The closest distance point is used as the initial distance point for matching.

4. The eye muscle adjustment ability training method according to claim 1, wherein: The image generation network model includes: A database module is used to store various graphic elements and their multimodal features used to generate visual training images, as well as all historical user training graphic element interaction data and feature data of each historical user. The types of graphic elements include image background, observation targets, and distractors, each category containing several graphic elements. The multimodal features of the graphic elements include image features and text features. The historical user training graphic element interaction data is used to record the interaction relationships between various graphic elements corresponding to suitable visual training images used by different users in historical training. A graph conversion module is used to convert the historical user training graph element interaction data into a user-graph element interaction graph G = (D, E); wherein each historical user and each graph element constitutes a node in the graph, and the node set is marked as D; the interaction relationship between the historical user node and the graph element node constitutes an edge in the graph, and the edge set is marked as E; and based on the user-graph element interaction graph, an interaction matrix of user nodes to graph element nodes is constructed. a ui Represents the interaction relationship between the historical user node u and the graph element node i, a ui =1 indicates that there is an interactive relationship between the historical user node u and the graph element node i, a ui = 0 means there is no interaction relationship between the historical user node u and the graph element node i, N U 、N I Represent the number of historical users and the number of graph elements respectively; The feature extraction module includes a multi-layer graph convolutional network model and a Transformer model, which is used to process the feature data of historical users and the data of the user-graph element interaction graph using the multi-layer graph convolutional network model to obtain the user interaction matching feature f between the feature data of historical users and the graph elements. u And the modal fusion embedding feature extraction process is performed on each graph element through the Transformer model to obtain the modal fusion embedding feature f of each graph element i ; Contrastive learning module, used to match features f based on historical user feature data and user interactions between graph elements u and the modal fusion embedding feature f of the graph elements i , calculate the matching value fs of different graph elements relative to historical user features ui , and based on the matching value fs ui Constructing a loss function, iteratively training the multi-layer graph convolutional network model and the Transformer model, and updating and optimizing the parameters of the multi-layer graph convolutional network model and the Transformer model through backpropagation to obtain a trained and optimized model; The graph element matching module is used to obtain the feature data of the user to be matched, input it into the trained and optimized multi-layer graph convolutional network model, obtain the user interaction matching features between the feature vector of the user to be matched and the graph elements, and calculate the matching value of each different element relative to the user to be matched with the trained and optimized Transformer model. Based on the matching value, a number of candidate graph elements are selected from each type of graph element to form a subset of matching candidate graph elements; The visual training image generation module is used to select image elements from the matching candidate image element subset, and based on a preset image generation program, combine the selected image elements according to the trainee's vision data and the training distance to generate a new visual training image.

5. The eye muscle adjustment ability training method according to claim 4, characterized in that: The feature data of the historical user is obtained by embedding the feature information of the historical user, and is represented by X u ={x u,1 ,x u,2 ,…,x u,n ,…,x u,N }, x u,n Represents the word embedding value of the nth user feature of historical user u, where N represents the total number of user features; The image features of the image elements are obtained by processing the image of the image elements using the pre-trained Resnet model. For the image features of any type of image element i in the image background, observation target or interference item, Expressed as: v i An image representing graphic element i; The text features of the image elements are obtained by processing the text information of the image elements using the pre-trained Bert model. For any type of image element i in the image background, observation target or interference item, the text features of the image element i are obtained by processing the text information of the image element using the pre-trained Bert model. Expressed as: t i Represents the text information of graph element i.

6. The eye muscle adjustment ability training method according to claim 4, characterized in that: In the feature extraction module, the feature data of historical users and the data of user-graph element interaction graph of the multi-layer graph convolutional network model are processed to obtain the user interaction matching feature f u The specific method is: For any historical user u, obtain its feature data X u , and obtain the image features and text features of each graph element interacted by the historical user u from the historical user training graph element interaction data, and respectively form the interaction image feature sequence of the historical user u and interactive text feature sequences in, and They represent the image features and text features of the jth graph element interacted by historical user u, j = 1, 2, …, J, where J represents the number of graph elements interacted by historical user u; Initialize the feature vector input to the multi-layer graph convolutional network model Concat(·) is a concatenation function; the hierarchical update of the lth layer is then performed as follows: in, represents the unit embedding matrix of the interaction matrix A, I is the identity matrix; W represents the degree matrix of the node set D composed of each graph element of each historical user in the user-graph element interaction graph G; (l) is the trainable parameter of the lth layer in the multi-layer graph convolutional neural network model, l=1,2,…,L, L represents the total number of layers of the multi-layer graph convolutional neural network model; Softmax(·) represents the normalized exponential function operation; the output of the last layer of the multi-layer graph convolutional neural network model is used as the user interaction matching feature f of the historical user u u ,Right now Through the above method, the user interaction matching features of different historical users are obtained respectively; In the feature extraction module, the modal fusion embedding feature extraction process is performed on each graph element through the Transformer model to obtain the modal fusion embedding feature f of each graph element. i The specific method is: For any graph element i, its image features and text features Unified as graph element features m∈{v,t}, that is or Graph element features for graph element i Perform multi-head attention mechanism processing to obtain the multi-head self-attention feature vector of graph element i in, V i m Graph element features The corresponding query vector, key vector, and value vector, m∈{v,t}; are the linear transformation matrices of query, key, and value in the multi-head self-attention mechanism respectively; d k is the dimension of the attention head; Softmax(·) represents the normalized exponential function operation; Attention(·) represents the attention mechanism operation; are the trainable weight vectors corresponding to the query, key, and value in the z-th attention head, respectively; z = 1, 2, …, Z, where Z represents the number of attention heads in the multi-head self-attention mechanism; Concat(·) is the concatenation function; Then, the cosine similarity between the multi-head self-attention feature vectors of different graph elements is calculated in, are the multi-head self-attention feature vectors of graph elements i and j respectively, m∈{v,t}; ||·|| represents modular calculation; For graph element i, the first K other graph elements whose cosine similarity values ​​are arranged from large to small constitute the K nearest neighbor graph element set N i-K , and then calculate the image modality embedding weight of image element i by the following formula and text modality embedding weights in, are the image modality multi-head self-attention feature vector and text modality multi-head self-attention feature vector of graph element j respectively; are the cosine similarities between the image modality multi-head self-attention feature vectors of image element i and image element j, is the cosine similarity between the text modality multi-head self-attention feature vectors of image element i and image element j; Then the modal fusion embedding feature f of graph element i is obtained by the following formula: i : Through the above method, the modal fusion embedding features of different graph elements are obtained respectively.

7. The eye muscle adjustment ability training method according to claim 4, characterized in that: In the feature extraction module and the comparative learning module, for any graph element i, the matching value fs of the graph element i relative to the historical user u is calculated as follows: ui : fs ui =(f u ) T f i ; Among them, f u is the user interaction matching feature corresponding to the historical user u, f i is the modal fusion embedding feature of graph element i; in this way, the matching value of each different graph element relative to different historical users is calculated respectively; In the contrastive learning module, the constructed loss function is expressed as: Where D is the node set of each historical user and each graph element; f i + The modal fusion embedding features of the graph elements that represent the interaction relationship with the historical user u, The modality fusion embedding features representing graph elements that have no interaction relationship with historical user u.

8. The eye muscle adjustment ability training method according to claim 4, wherein: In the graph element matching module, the specific method for calculating the matching value of each different element relative to the user to be matched is: According to the user u to be matched t Feature data X ut , match it with the user-graph element interaction graph G, and determine the user u to be matched t For the K1 matched historical users, the user interaction matching features of the K1 matched historical users are calculated using the trained and optimized multi-layer graph convolutional network model. The optimized Transformer model is then trained to calculate the matching values ​​of each different element relative to the K1 matched historical users, which are used as the matching values ​​of each different element relative to the user to be matched. Then, the matching degree values ​​of each type of graphic element are counted separately, and the top K2 graphic elements with the highest matching degree values ​​in the corresponding graphic element type are screened out from each type of graphic element. K2 candidate graphic elements of the picture background type, K2 candidate graphic elements of the observation target type, and K2 candidate graphic elements of the interference item type are obtained to form a subset of matching candidate graphic elements.

9. The eye muscle adjustment ability training method according to claim 2, wherein: In step S4, the total gaze duration is calculated as follows: By collecting the trainee's eye images, identifying the trainee's eye movement state and pupil state, and judging the gaze state based on the eye movement state and pupil state; if the trainee's eye remains motionless and the pupil state remains unchanged for a duration that reaches the preset minimum gaze duration t0, the trainee is determined to be in a gaze state, and the trainee's total gaze duration t is calculated. g =t0+t x , t x Indicates the time the trainee continues to stare after being judged to be in the staring state.

10. A method for generating a visual training image, characterized in that: The trainee's characteristic information and the training distance information are used as inputs to a preset image generation network model to generate a visual training image for training the trainee's eye muscle adjustment ability; The image generation network model includes: A database module is used to store various graphic elements and their multimodal features used to generate visual training images, as well as all historical user training graphic element interaction data and feature data of each historical user. The types of graphic elements include image background, observation targets, and distractors, each category containing several graphic elements. The multimodal features of the graphic elements include image features and text features. The historical user training graphic element interaction data is used to record the interaction relationships between various graphic elements corresponding to suitable visual training images used by different users in historical training. A graph conversion module is used to convert the historical user training graph element interaction data into a user-graph element interaction graph G = (D, E), where each historical user and each graph element constitutes a node in the graph, the node set is marked as D, and the interaction relationship between the historical user node and the graph element node constitutes an edge in the graph, the edge set is marked as E; and based on the user-graph element interaction graph, an interaction matrix of user nodes to graph element nodes is constructed. a ui Represents the interaction relationship between the historical user node u and the graph element node i, a ui =1 indicates that there is an interactive relationship between the historical user node u and the graph element node i, a ui = 0 means there is no interaction relationship between the historical user node u and the graph element node i, N U 、N I Represent the number of historical users and the number of graph elements respectively; The feature extraction module includes a multi-layer graph convolutional network model and a Transformer model, which is used to process the feature data of historical users and the data of the user-graph element interaction graph using the multi-layer graph convolutional network model to obtain the user interaction matching feature f between the feature data of historical users and the graph elements. u And the modal fusion embedding feature extraction process is performed on each graph element through the Transformer model to obtain the modal fusion embedding feature f of each graph element i ; Contrastive learning module, used to match features f based on historical user feature data and user interactions between graph elements u and the modal fusion embedding feature f of the graph elements i , calculate the matching value fs of different graph elements relative to historical user features ui , and based on the matching value fs ui Constructing a loss function, iteratively training the multi-layer graph convolutional network model and the Transformer model, and updating and optimizing the parameters of the multi-layer graph convolutional network model and the Transformer model through backpropagation to obtain a trained and optimized model; The graph element matching module is used to obtain the feature data of the user to be matched, input it into the trained and optimized multi-layer graph convolutional network model, obtain the user interaction matching features between the feature vector of the user to be matched and the graph elements, and calculate the matching value of each different element relative to the user to be matched with the trained and optimized Transformer model. Based on the matching value, a number of candidate graph elements are selected from each type of graph element to form a subset of matching candidate graph elements; The visual training image generation module is used to select image elements from the matching candidate image element subset, and based on a preset image generation program, combine the selected image elements according to the trainee's vision data and the training distance to generate a new visual training image.

Images