Multispectral dynamic visual function training system and method

Abstract

The scheme belongs to the field of vision training, and specifically relates to a multispectral dynamic visual function training system and method. The multispectral dynamic visual function training system comprises an image acquisition module, an intelligent analysis module, a control module, a training execution module and a multispectral light source module; the image acquisition module is arranged at a corresponding position of the eye and is used for acquiring image information near the eyeball of a user, personal information of the user and a vision diagnosis result in real time, the image information comprising an eyeball movement trajectory, pupil size and morphology, corneal edge contour and eye white elasticity state; the intelligent analysis module is combined with the image information and the acquisition time of the image information. Through acquisition analysis regulation and control of a closed loop, the scheme solves the problem that the prior art lacks a real-time analysis mechanism based on eyeball images, the adaptability of a training scheme to dynamic characteristics of the eye is weak, and precise regulation and control of functions such as light nutrition supplementation and choroid stimulation cannot be achieved, and also achieves the effects of precise training, prevention and protection, synergistic improvement and improved compliance.

Description

TECHNICAL FIELD

[0001] The present solution belongs to the field of visual training, and specifically relates to a multi-spectral dynamic visual function training system and method. BACKGROUND

[0002] Current visual function training technology has significant limitations in individualized adaptation and dynamic optimization, making it difficult to meet the precise training needs of children, adolescents, and patients with amblyopia and astigmatism.

[0003] As disclosed in the document of Chinese patent application CN109984924A, a method for improving astigmatism vision training is implemented by combining multiple steps such as vision evaluation, fumigation care, and eye muscle training. However, it relies on manual assessment of eye habits and vision test results, lacks real-time capture of eye dynamic characteristics, and cannot adjust the training program according to subtle parameters such as eye movement and pupil changes, limiting the degree of individualization. Although the twelve zodiac eye training modules are targeted, the training parameters are fixed and cannot adapt to individual differences in corneal shape and eye axis development.

[0004] Chinese patent application CN111743741A addresses axial refractive disparity amblyopia and uses a combination of occlusion, prescription glasses, and optical suppression to develop a training plan based on eye axis length and vision differences. However, data collection relies on static measurements such as eye axis length and does not involve real-time image analysis of the eye, making it difficult to dynamically monitor eye axis changes and accommodation function, resulting in delayed program adjustments that do not keep pace with eye development, especially in terms of the precision of activating visual cells in the amblyopic eye.

[0005] Based on the above technology, Chinese patent application CN104865712A discloses a child intelligent health glasses that monitors eye behavior through distance and light sensors and implements intervention by combining electrooculography. However, it only focuses on eye posture and fatigue warning and does not collect images near the eye, analyze corneal flatness, or support the targeted activation of visual cells in amblyopic patients and the smoothing of the cornea in astigmatism patients.

[0006] That is, existing visual training technology mainly relies on individualized guidance from professional teams based on experience, and lacks individualized adaptation of training parameters based on image features of eye structures such as choroid thickness and ciliary muscle state. Light nutrition supplementation and training duration lack individual adaptation basis.

[0007] In summary, existing technology lacks real-time image analysis mechanisms based on eye images, and the training program has weak adaptability to eye dynamic characteristics, making it difficult to precisely control functions such as light nutrition supplementation and choroid stimulation. There is an urgent need for an innovative system that integrates image collection, intelligent analysis, and multi-modal training. SUMMARY

[0008] The purpose of the present scheme is to provide a multi-spectral dynamic visual function training system and method to solve the significant limitations of current visual function training techniques in terms of personalized adaptation and dynamic optimization, and to meet the precise training needs of children, adolescents, and patients with amblyopia and astigmatism.

[0009] In order to achieve the above purpose, the present scheme provides a multi-spectral dynamic visual function training system, comprising:

[0010] An image acquisition module is arranged at a corresponding position of the eye for real-time acquisition of image information near the eyeball of the user, wherein the image information includes eyeball movement trajectory, pupil size and shape, corneal edge profile, and eye white elasticity state;

[0011] An intelligent analysis module combines the image information and the acquisition time of the image information, analyzes the pupil light reaction speed according to the eyeball movement trajectory and the pupil size and shape, analyzes the distribution of irregular areas of the cornea according to the corneal edge profile and the eye white elasticity state, integrates the pupil light reaction speed and the distribution of irregular areas of the cornea into eyeball characteristic parameters, and stores the eyeball characteristic parameters in association with the corresponding image information and the acquisition time of the image information; obtains personal information and vision diagnosis results of the user, and establishes an individual eye model based on the eyeball characteristic parameters, the personal information, and the vision diagnosis results;

[0012] A control module generates a personalized training scheme based on the individual eye model, wherein the personalized training scheme includes a training picture, and each training parameter in the training scheme is updated based on the image information;

[0013] A multi-spectral light source module is used to emit light waves in the visible light wavelength range according to each training parameter, wherein green light of a preset green light wavelength is used to stimulate cone cells, and red light of a preset red light wavelength is used to stimulate choroidal blood circulation; the intensity and switching frequency of the light waves are dynamically adjusted according to the eyeball characteristic parameters;

[0014] A training execution module includes a contrast adjustment unit and a cyclic stimulation unit; the contrast adjustment unit acquires the distribution of irregular areas of the cornea, and adjusts the contrast of the training picture according to the distribution of irregular areas of the cornea to simulate the visual environment required for myopia prevention and control; the cyclic stimulation unit adjusts the red light irradiation wavelength of the multi-spectral light source module through the eyeball characteristic parameters to promote eye water circulation and blood circulation, and thicken the choroid; and the training execution module sends the adjusted training parameters to the multi-spectral light source module;

[0015] The principle and technical effect of the scheme are that the scheme realizes precise visual function training through acquisition, analysis and closed-loop control. The image acquisition module captures the eye structure (including the corneal edge profile and the eye white elasticity state) and dynamic characteristics (including the eye movement trajectory, the pupil size and shape) in real time to provide original data support; the intelligent analysis module converts these image data into quantitative parameters (such as corneal irregularity), combines user personal information and vision diagnosis results to build an individual eye model, and lays the foundation for customization of thousands of people with different faces; the multi-spectrum light source module outputs specific light waves according to the individual model, i.e. the preset green light activates the cone cells, and the preset red light precisely stimulates the choroid to promote blood circulation; the training execution module dynamically adjusts the training intensity (such as the focusing speed of ciliary muscle training and the gradient change of contrast adjustment) combined with the quantitative parameters, and the control module is real-time linked with each module to ensure that the training parameters are real-time adapted to the eye state (such as the eye axis development and the accommodation function change), forming a complete closed loop. The above-mentioned way directly solves the core problem of poor adaptability of the prior art scheme. The scheme customizes the individual eye model, and for the irregular area of the cornea of the astigmatism patient, the smoothness of the cornea can be improved by specific light waves and contrast adjustment.

[0016] The scheme enhances the accuracy of light nutrition supplementation through the linkage of multi-spectrum light waves and training units, and can significantly improve the activation efficiency of visual cells for amblyopia patients. The scheme dynamically optimizes the training scheme according to the eye state through real-time data feedback, and the eye axis difference reduction speed of the axial amblyopia patient is obviously improved compared with the fixed module training. The scheme integrates choroid stimulation and ciliary muscle training to promote eye circulation while relieving fatigue, shortens the effective training time of the user per day, and still maintains the same improvement effect.

[0017] Secondly, the eye white elasticity state monitoring of the image acquisition module and the long-term data tracking of the intelligent analysis module can predict potential risks such as dry eye syndrome and corneal damage through the eye white elasticity change trend, and adjust the parameters in advance through the training execution module. For example, for a teenager user with continuously decreasing eye white elasticity, after analyzing the data, the contrast of the training picture will be automatically reduced to reduce the stimulation of the cornea, and the red light irradiation time will be increased (for example, from 2 minutes to 3 minutes) to promote tear secretion, which not only ensures the training effect, but also prevents eye surface damage in advance, realizing the dual value of therapeutic training combined with preventive protection.

[0018] Furthermore, the red light of the multi-spectrum light source is coordinated with the green light to accurately analyze the mixed vision problem (such as amblyopia combined with astigmatism) according to the individual eye model, and can simultaneously act on different pathological links. For example, for patients with amblyopia combined with moderate astigmatism, the choroid is thickened by red light stimulation to improve amblyopia, and at the same time, the irregular area of the cornea is irradiated by green light in a directional manner, and the contrast is adjusted, so that the retinal light sensing ability and the corneal refractive state are simultaneously optimized in a single training process, thereby shortening the improvement period of the mixed vision problem, which is a synergistic effect that cannot be achieved by a single light source or a fixed scheme.

[0019] In addition, the image acquisition module captures dynamic features related to user attention (such as blink frequency, eye movement stability) in real time, and the intelligent analysis module adjusts the training mode in conjunction with the control module and the training execution module if it identifies that the user (especially children) is distracted (such as chaotic eye movement and abnormally high blink frequency). For example, when it is found through eye movement trajectory analysis that a child user is distracted, the training intensity is automatically reduced (such as slowing down the contrast adjustment gradient), and at the same time, the green light with a wavelength of 550 nm is temporarily switched (this wavelength is more easily perceived by children's vision, and the color attraction is stronger than that of the conventional training light source). On the other hand, the static training picture is converted into a dynamic scene (such as a slowly moving visual target), and the contrast adjustment gradient is slowed down to reduce the training intensity. By combining color attraction and dynamic interaction, the traditional fixed training scene is replaced, the boredom is reduced, the user's resistance due to lack of concentration is avoided, and the long-term training compliance is ultimately improved, which is especially suitable for children and other user groups with higher requirements for interestingness.

[0020] In summary, the present scheme solves the problem of lack of real-time analysis mechanism based on eye image in the prior art, weak adaptability of training scheme to eye dynamic features, and inability to achieve precise regulation of light nutrition supplementation and choroid stimulation by using a closed loop of acquisition, analysis and regulation. In addition, the present scheme also achieves the effects of precise training, prevention and protection, synergistic improvement, and improved compliance.

[0021] Further, the intelligent analysis module is also configured to obtain an individual eye model before training as an initial model, obtain an individual eye model at the end of training as a second model, obtain an interval time length between the initial model and the second model as a training time length, and construct a training improvement model based on the training parameters, the initial model, and the second model in the training process. The intelligent analysis module obtains an individual eye model before the next training after the end of training as a third model, obtains a time length between the second model and the third model as a decay time length, and constructs a decay model based on the second model, the third model, and the decay time length. The intelligent analysis module sets a training parameter threshold according to user personal information and vision diagnosis results, wherein the threshold is adjusted differently according to the eye axis development stage and the astigmatism degree range of users of different ages.

[0022] The intelligent analysis module identifies the initial model, the second model, the third model and the corresponding collection time according to the collection time of the image information and the eyeball characteristic parameter, inputs the initial model and the corresponding training parameter into the training improvement model, outputs different training durations and the corresponding second model as the training prediction result under the constraint of the training parameter threshold, and then inputs the second model into the attenuation model to output different attenuation durations and the corresponding third model as the attenuation result of the training prediction result; the training prediction result and the attenuation result of the training prediction result are integrated into the training prediction result and sent to the user terminal, and the training duration fed back by the user terminal is received as the current training duration, and the control module adjusts the personalized training scheme according to the current training duration.

[0023] The present scheme can not only predict the visual improvement result corresponding to different training durations based on the initial model and the training parameter (such as the parents of child users can know in advance that training for 20 minutes can improve the activation efficiency of cone cells by about 15%), but also predict the change of vision during the interval between two training (i.e. before and after the attenuation duration) through the attenuation model (such as the student group can predict that the choroid thickness may attenuate about 0.02mm from the interval of 5 days after the weekend training to the next week training), helping users to actively plan the training period; the differentiated adjustment of the training parameter threshold (such as the eye axis of children aged 6-12 develops fast, and the threshold is relaxed by about 20% compared with adults) not only avoids overtraining but also ensures the effect; the user terminal feedback mechanism makes the training scheme more suitable for actual needs (such as the white-collar selects 15-minute efficient training and confirms, and the present scheme automatically optimizes the red light irradiation intensity). Among them, the attenuation model of the present scheme is also linked with the daily schedule of the user, for example, for middle school students, the attenuation model predicts the visual attenuation trend from the weekend training to the next Wednesday by combining the eye use intensity of the students from Monday to Friday, and increases the red light irradiation duration of the next training by 30 seconds to weaken or even offset the attenuation. This kind of dynamic adjustment deeply bound with the change of vision and the schedule breaks through the limitation of traditional training data, realizes the whole cycle management of prediction, intervention and adaptation.

[0024] Further, the intelligent analysis module is also used for multi-dimensional classification according to the user personal information and the visual diagnosis result, statistical analysis of the attenuation model and the training improvement model corresponding to each category, extraction of the universal user characteristics in the same type of users according to the distribution of the number of users in different categories; mark the high-quality characteristics with better training recovery effect than the universal user characteristics, and the weak characteristics with worse recovery effect than the universal user characteristics; for the users who use the multi-spectrum dynamic visual function training system for the first time, the personal information and the visual characteristics are extracted and matched with the universal user characteristics, the high-quality characteristics and the weak characteristics to locate the category to which the user belongs, and the attenuation model and the training improvement model of the corresponding category are bound as the initial template with the user personal information based on the positioning result, and the reinforcement parameters corresponding to the high-quality characteristics or the compensation parameters corresponding to the weak characteristics are synchronously integrated.

[0025] Marking the good and weak features and integrating the corresponding parameters, the initial template has the ability of individualized reinforcement (such as improving the light stimulation efficiency for good features) or compensation (such as relaxing the training parameter threshold for weak features) on the basis of group adaptation, balancing universality and individual differences. At the same time, the mechanism of group feature positioning and individual parameter binding not only breaks through the slow limitations of traditional individual data iteration, but also avoids the roughness of the simple group scheme. Among them, the potential changes of individuals can be predictively guided through the group characteristics, for example, for a young child with astigmatism, the classification shows that the universal feature of the 3-6 year old astigmatism group is that the irregular area of the cornea is slow to repair, and the vision characteristics of the child match the good feature of the pupil light reaction speed better than the group average. The initial template is improved by the training model of the group, and the good feature corresponding to the reinforcement parameter (such as shortening the green wavelength switching interval) is integrated, so that the corneal repair and pupil response can be guided to optimize synchronously in the first training, avoiding the poor initial training effect caused by the limited patience of the child, and realizing the early activation of the potential advantages of the group rules to the individual.

[0026] Further, the intelligent analysis module takes the correlation between the training duration and the decay duration into the statistical dimension when extracting the universal user characteristics; the user characteristics are marked in combination with the eye feature parameters in the image information; for the first-time user, if there is a difference between the eye feature parameters after feature matching and the universal characteristics, in addition to matching the corresponding decay model, the decay duration in the training prediction result is adjusted, and the training execution module is adjusted to adjust the training parameters.

[0027] Taking the correlation between the training duration and the decay duration into the universal feature statistics enhances the adaptability of the model to the time dimension, making the training scheme more suitable for the actual training interval rules of the user; combining the eye feature parameters to mark the user characteristics realizes the fusion of static group rules and dynamic eye state, and improves the accuracy of feature recognition; for the first-time user, when there is a difference in feature matching, the decay duration and the training parameters are adjusted synchronously to realize the dual adjustment of model adaptation and parameter optimization, and accelerate the effect landing of the first-time training. Among them, the scheme can also realize the cross-dimension collaborative correction of the time correlation rule and the eye dynamic parameter, for example: for a first-time adult amblyopia user, the extracted universal characteristics show that the user trains 2 times a week and each time for 15 minutes, and the decay rate is relatively stable, but if the eye feature parameters (such as ciliary muscle activity) of the user are lower than the average of the universal characteristics, the decay duration is adjusted from 3 days to 2 days, and the training execution module is also linked to increase the gradient change rate of the contrast adjustment, which not only reduces the influence of decay through time rules, but also optimizes the ciliary muscle training intensity with the help of dynamic parameters, so that the visual cell activation efficiency of single training is improved, breaking through the limitations of traditional individual differences only by prolonging the training duration, and realizing the synergistic effect of time management and dynamic training.

[0028] Further, the intelligent analysis module is also used for comparing the attenuation model, the training improvement model and the positioning result of the current user with the universal user features of the corresponding category, generating current state information containing a training effect deviation value and sending the current state information to the user terminal; obtaining the individual eye model and the training parameter calculation state change trend of the user according to a preset period; combining a preset emotional support information library, matching the accompanying guide information from the emotional support information library according to the state change trend and the training effect deviation value, and synchronously sending the accompanying guide information to the user terminal.

[0029] By comparing the user model with the universal features to generate the state information, the user can intuitively master the deviation of the training effect and the group level (for example, the corneal irregularity improvement rate is lower than that of the same user by 12%), and the visualization of the training progress is realized. The state change trend is tracked according to a period, and the user can find the long-term regularity (for example, the attenuation rate slows down for three consecutive weeks). The emotional support information library is combined to match the accompanying guide information, the emotional intervention is added outside the technical guidance, and the user adherence is improved. The training effect deviation value can also be accurately linked with the emotional guidance. For example, for a 10-year-old child with amblyopia, the detection shows that the activation efficiency of the cone cells in the training improvement model of the child continuously lags behind the universal value by 8% (that is, the deviation value is stable), and instead of directly pushing the general encouraging words, the accompanying information that the sensitive period of the cone cells of the amblyopia repair is usually 4-6 weeks is matched from the emotional library. The data shows that the activation efficiency has been improved by 5% compared with the first week, and the child will enter the rapid improvement period if the child continues to adhere to the training. The accompanying information not only explains the lagging reason with professional data, but also strengthens the confidence through the phased progress. This scheme has professional authority and emotional empathy, and effectively reduces the probability that the child gives up the training due to the short-term effect being not obvious.

[0030] Further, after the intelligent analysis module generates the training effect deviation value and the state change trend, if the training effect deviation value is in a significant deviation interval and the trend is deviation expansion for two consecutive preset periods, the attenuation model is automatically called to optimize the attenuation model, the basic parameters of the attenuation model are dynamically adjusted based on the changes of the eye feature parameters of the current user, and the light stimulation parameter weight in the training improvement model is simultaneously optimized, so that the adaptability of the attenuation model and the training improvement model is iterated in real time according to the state change trend.

[0031] The scheme realizes automatic starting of model optimization, reduces manual intervention lag, through trigger conditions of significant deviation and trend expansion in two consecutive periods; improves individual adaptation accuracy based on dynamic adjustment of user eye feature parameters (such as corneal irregularity fluctuation) to basic parameters of the decay model; and avoids adaptation imbalance caused by single model optimization by simultaneously optimizing training to improve the light stimulation parameter weight (such as red light wavelength adjustment proportion) of the model. Among them, the deviation judgment at one time acts on two models, not only quickly corrects the current deviation, but also accumulates the correlation between eye features and double model parameters, which can be reused by similar users, realizes the transformation of individual optimization experience to group adaptation ability, breaks through the limitation of single model independent optimization, and significantly improves the self-adaptation and generalization ability of the scheme.

[0032] Further, when acquiring the current state information in a period, the intelligent analysis module compares the trend change of the group formed by similar users with the state change trend corresponding to the individual user, and when the deviation rate between the state change trend corresponding to the individual user and the group trend exceeds the preset deviation threshold, the training parameters of the individual user are supplemented to the similar universal characteristics as samples, and the correlation of the decay model is optimized in reverse, realizing the closed loop of individual optimization, group feature iteration and model universality improvement.

[0033] Comparing the state change trend of the individual and the group can accurately identify individual-specific characteristics and avoid the disconnection of group universal characteristics; the effective training parameters of the individual with deviation exceeding the threshold are supplemented to the universal feature library, so that the group characteristics are dynamically iterated with actual cases, and the coverage of diversified individuals is enhanced; the correlation of the decay model (such as the correlation coefficient logic of the training duration and the decay duration) is optimized in reverse, promoting the model from being based on historical data to adapting to real-time trends, forming a positive closed loop of individual optimization, group feature updating and model universality improvement.

[0034] The group characteristics of the traditional system are mostly static, while the individual effective experience becomes a group live sample in the scheme. For example, in a group of 13-18-year-old high myopia, the repair response of a user to 650nm red light is much higher than that of the group, with a trend deviation rate of 18% (exceeding the threshold of 15%). The scheme supplements its 650nm red light irradiation for 18 minutes to the group characteristics, and optimizes the correlation algorithm of red light wavelength and decay duration in the decay model. After updating, the decay duration prediction error of this user decreases from 15% to 8%, and the first training of similar new users can adapt to this parameter, realizing the transformation of individual experience into group adaptation ability, and breaking through the limitation of the separation of group and individual needs.

[0035] Further, after updating the parameters of the decay model or the training improvement model, the intelligent analysis module dynamically adjusts the training parameter threshold according to the model change rate, and corrects the threshold by weighted integration of the decay coefficient change rate, so that the change rate correction threshold and the decay model or the training improvement model iteration are coordinated, realizing single threshold adjustment adapting to decay characteristics and stimulation effect.

[0036] Further, the intelligent analysis module adjusts the training parameter threshold value in layers after the population universal feature update or individual model optimization, and the training parameter threshold value includes a training duration threshold value, a red light power threshold value, and a pulse frequency threshold value.

[0037] The training parameter threshold value is dynamically adjusted by the change rate of the decay model or the training improvement model, the real-time cooperation of the training parameter threshold value and the double model iteration is realized, the fixed threshold value is avoided to be out of touch with the dynamic model, the single threshold value adjustment can simultaneously adapt the decay characteristics (such as the decay duration change) and the stimulation effect (such as the light wavelength adaptability), and the one adjustment and double adaptation are achieved. The hierarchical adjustment logic differentiates the core parameters (such as the contrast threshold value) and the auxiliary parameters (such as the training interval), ensures the high-precision matching of the key parameters and the model optimization, and simplifies the auxiliary parameter adjustment to improve the efficiency. The closed-loop linkage of the training parameter threshold value adjustment and the model iteration, for example, when the decay characteristics accelerate, the change can be adapted and the green light stimulation effectiveness is ensured; when the population universal feature is updated (such as the correction period is shortened), the hierarchical adjustment optimizes the core refractive adjustment threshold value, and the auxiliary rest interval is only fine-tuned, so that the threshold value parameters considering the individual and the population can be obtained for the first training of a new user, and the limitations of the traditional threshold value static or single-dimensional adjustment are broken through. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 FIG. 1 is a structure diagram of a function module of a multi-spectrum dynamic visual function training system according to an embodiment of the present application. DETAILED DESCRIPTION

[0039] The concept and technical effects of the present application will be described in detail below with reference to the embodiments, so as to fully understand the purposes, features and effects of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, but not all the embodiments. Based on the embodiments of the present application, other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application:

[0040] As shown in FIG. 1, a multi-spectrum dynamic visual function training system includes: Figure 1

[0041] An image acquisition module is arranged at a corresponding position of the eye, and is used for acquiring image information near the eyeball of the user in real time. The image information includes an eyeball movement trajectory, a pupil size and morphology, a corneal edge contour, and an eye white elasticity state.

[0042] ​The intelligent analysis module combines the image information with the image information collection time, analyzes the pupil light reaction speed according to the eyeball movement trajectory and the pupil size and shape, analyzes the corneal irregular region distribution according to the corneal edge contour and the eye white elasticity state, integrates the pupil light reaction speed and the corneal irregular region distribution into the eyeball characteristic parameter, and stores the eyeball characteristic parameter in association with the corresponding image information and the image information collection time; the personal information and the vision diagnosis result of the user are obtained, and the individual eye model is established by combining the eyeball characteristic parameter, the personal information and the vision diagnosis result;

[0043] The control module generates a personalized training scheme according to the individual eye model, the personalized training scheme contains a training picture, and each training parameter in the training scheme is updated according to the image information;

[0044] The multi-spectrum light source module is used for emitting light waves in a visible light wavelength range according to each training parameter, wherein green light of a preset green light wavelength is used to excite cone cells, and red light of a preset red light wavelength is used to stimulate choroid blood circulation; the intensity and the switching frequency of the light waves are dynamically adjusted according to the eyeball characteristic parameter;

[0045] The training execution module includes a contrast adjustment unit and a cyclic stimulation unit; the contrast adjustment unit obtains the corneal irregular region distribution, and adjusts the contrast of the training picture according to the corneal irregular region distribution to simulate the visual environment required for myopia prevention and control; the cyclic stimulation unit adjusts the red light irradiation wavelength of the multi-spectrum light source module through the eyeball characteristic parameter to promote the water circulation and the blood circulation of the eye and thicken the choroid; and the training execution module sends the adjusted training parameter to the multi-spectrum light source module;

[0046] The eye white elasticity state refers to the deformation ability and recovery ability of the sclera (i.e. “eye white”, the tough white connective tissue on the outer layer of the eyeball) under stress or natural movement (such as blinking, eyeball rotation). Under normal conditions, a healthy sclera has a certain elasticity and will slightly deform under the pressure of the eyelid when blinking, and quickly recover to be flat without wrinkles after blinking; if the elasticity of the eye white decreases (such as long-term eye fatigue, pre-stage of dry eye, sclera aging, etc.), it will show slow deformation recovery after blinking, local temporary wrinkles, or decreased adhesion of the sclera and conjunctiva during eyeball rotation (appearance of fine gaps).

[0047] The image acquisition module acquires dynamic images of the eye white at a frame rate of 15-20 frames per second by a high-resolution miniature camera (≥ 2 million pixels) combined with a near-infrared light supplement device, which is installed 5-10 cm away from the eye. When the image acquisition module detects that the eyelid blocks the iris by no less than a preset proportion (i.e., when blinking, the preset proportion is generally 80%, which is set by the administrator according to the user's personal information), the images of the eye white before, during and after blinking are recorded. When the eye rotation (pupil displacement and eyelid horizontal and vertical ratio) is greater than a preset rotation amplitude (generally 15%, which is set by the administrator according to the user's personal information), the stretched state of the eye white during rotation is collected (i.e., the morphological changes of the sclera due to the pulling effect of eye movement). After collection, the images are denoised by Gaussian filtering, the eye white area is separated by threshold segmentation, and the deformation recovery time (the length of time for the eye white to recover after blinking) and the deformation amplitude (area change rate) are quantified by edge detection. Finally, these parameters are associated with the collection time and other eye information and transmitted to the intelligent analysis module.

[0048] Among them, each training parameter in the training scheme includes two categories: multi-spectral light source parameters and training scene and intensity parameters. Multi-spectral light source parameters directly affect the light stimulation effect and are executed by the multi-spectral light source module, including red light wavelength (such as 640 nm and 655 nm, used to stimulate the choroid), green light wavelength (such as 530-570 nm, used to excite cone cells), light wave intensity (such as red light intensity 50-230 cd / m², adjusted according to the differences in eye axis development stage), and light wave switching frequency (such as green light switching frequency 4-5 Hz, which can optimize the activation efficiency of cone cells). Training scene and intensity parameters are regulated by the training execution module, including training duration (such as 12 minutes and 15 minutes, which can be adjusted according to user feedback), picture contrast (such as 20%-40%, which is adjusted according to the distribution gradient of irregular areas of the cornea), red light irradiation duration (such as 2-3 minutes, which can be extended due to reduced elasticity of the eye white), and pulse frequency (such as red light pulse frequency 2 Hz, which is dynamically modified according to eye feature parameters).

[0049] Among them, the intelligent analysis module is further configured to obtain an individual eye model before training as an initial model, obtain an individual eye model at the end of training as a second model, obtain an interval time length between the initial model and the second model as a training duration, and construct a training improvement model based on the training parameters, the initial model, and the second model during the training process. The intelligent analysis module is further configured to obtain an individual eye model before the next training after the end of training as a third model, obtain a time length between the second model and the third model as a decay time length, and construct a decay model based on the second model, the third model, and the decay time length.

[0050] The intelligent analysis module sets a training parameter threshold value according to user personal information and vision diagnosis results, wherein the threshold value is adjusted differently according to the eye axis development stage and the range of astigmatism of users of different ages.

[0051] The intelligent analysis module identifies the initial model, the second model, the third model and the corresponding collection time according to the collection time of the image information and the eyeball feature parameter, inputs the initial model and the corresponding training parameter into the training improvement model, outputs different training durations and the corresponding second model as the training prediction result under the constraint of the training parameter threshold, and then inputs the second model into the attenuation model to output different attenuation durations and the corresponding third model as the attenuation result of the training prediction result. The training prediction result and the attenuation result of the training prediction result are integrated into the training prediction result and sent to the user terminal. The training duration fed back by the user terminal is received as the current training duration, and the control module adjusts the personalized training scheme according to the current training duration.

[0052] In one specific embodiment of the present scheme, when the intelligent analysis module acquires the initial model, the user's eye use state (such as the duration of near-distance eye use and the ambient light intensity) within a preset time (such as 1 hour) before training is recorded synchronously. When the training improvement model is constructed, the historical initial model, the second model and the corresponding training parameter (such as the irradiation duration of green light with a wavelength of 530-570 nm and the red light intensity level) of more than 3 times of training are used as training samples by using a deep learning algorithm, so that the error of the training prediction result output by the model is controlled within a preset range (generally 5%, which is determined by the administrator according to the expected accuracy of the training prediction result).

[0053] When the attenuation model is constructed, the user's daily eye use intensity coefficient is introduced, and the attenuation duration is weighted with the coefficient to make the deviation of the third model prediction value from the actual collection value ≤0.01 mm (in the dimension of choroidal thickness). The eye use intensity coefficient is a core parameter for quantifying the influence of “daily near-distance eye use time” on the eye fatigue degree and the training effect attenuation speed. The longer the near-distance eye use time is, the higher the sustained tension of the ciliary muscle is, the lower the choroidal repair efficiency is, and the faster the attenuation speed of the vision improvement effect after training is. Therefore, the eye use intensity coefficient increases with the increase of the eye use time. In the present specific embodiment, the adjusted attenuation duration = original attenuation duration × (1 / eye use intensity coefficient). The greater the eye use intensity coefficient is, the heavier the eye fatigue is, and the faster the attenuation speed is. Therefore, the adjusted attenuation duration is shorter, which ensures that the model prediction value is close to the actual collection value. In the present specific embodiment, the users mainly include the elderly leisure population, the primary and secondary school population, the working population and the high-intensity eye use population. The differences of the eye use intensity coefficient and the choroidal thickness prediction of the four kinds of users are shown in Table 1 as follows:

[0054] Population type Daily near vision time (hours) Eye use intensity coefficient Original attenuation duration (hours, ideal state) Weighted calculation attenuation duration (hours) Choroidal thickness prediction deviation (mm) Old leisure population ≤2 (mostly distance vision) 1 72 72×(1 / 1.0)=72 ≤0.005 Primary and secondary school population 3-4 (schoolwork + after-school homework) 1.2 72 72×(1 / 1.2)=60 ≤0.008 Working population 5-6 (computer office + commuting eye use) 1.5 72 72×(1 / 1.5)=48 ≤0.010 High-intensity eye use population ≥8 (such as postgraduate examination / design practitioners) 1.8 72 72×(1 / 1.8)=40 ≤0.010

[0055] Table 1

[0056] The data in Table 1 is based on the table data support principle and is developed around the correlation of "eye use scenario-eye use intensity coefficient-decay duration". The eye use intensity coefficient is based on the "eye use duration-ciliary muscle fatigue degree" experiment, and the coefficient is set to 1.0 (baseline, no additional decay burden) for the elderly group ≤2 hours of near-distance eye use; 1.2 for students for 4 hours of eye use; 1.5 for office workers for 6 hours of eye use; and 1.8 for high-intensity eye use for 8 hours, following the rule that the intensity coefficient increases by 0.3-0.4 for every additional 2 hours of eye use, which is consistent with the clinical conclusion that the longer the eye use, the faster the choroid decay. The weighted calculation takes 72 hours as the original decay duration (universal value), the greater the eye use intensity coefficient, the heavier the eye fatigue, and the shorter the decay duration (e.g., 48 hours for office workers). The eye use intensity coefficient of each type of user is fitted based on 200 data, ensuring that the deviation between the predicted choroid thickness and the actual collected value is ≤0.01 mm (the accuracy requirement set by the administrator).

[0057] When setting the training parameter threshold, the eye axis development stage threshold for users aged 6-12 years is relaxed by 15% compared to adults (e.g., the upper limit of red light intensity is increased from 200 lux to 230 lux), and the contrast adjustment threshold gradient for users with astigmatism > 200 degrees is reduced by 20% (e.g., from 5% gradient to 4%). Users with astigmatism > 200 degrees have high corneal irregularity, steep contrast sensitivity curves, and low tolerance to contrast changes (prone to ghosting and visual fatigue). Reducing the gradient can reduce the amplitude of single adjustment and match the sensitive characteristics of their visual system; dynamic feedback adjustment adapts to individual tolerance differences and avoids overstimulation. In this specific embodiment, among the 200 high astigmatism users of the same type, the user discomfort rate (increased blinking frequency, eye white wrinkles) reached 45% at a 5% gradient, and decreased to 12% after a 20% compression, while ΔC could still be maintained at ≥7.5% (close to the effect of 8% for the universal population, with no significant decay); at the same time, from the parameter coordination, the 20% compression ratio complements the 15% relaxation of the red light intensity threshold for users with astigmatism > 200 degrees.

[0058] The image acquisition module first acquires user corneal image data, the intelligent analysis module extracts the astigmatism degree parameter, and after determining > 200 degrees, the default 5% (which can be set by the administrator) contrast adjustment gradient is compressed by 20% (which can be set by the administrator according to the situation), and the personalized gradient value is generated through the formula "5% × (1-20%) = 4%". The control module dynamically adjusts the picture contrast in 4% steps, while real-time collecting the eye white stretching state and blinking frequency, if frequent blinking and other fatigue signals are detected, the gradient is temporarily reduced to 3%; if there is no abnormality, the 4% gradient is maintained, and finally the corneal irregularity improvement rate ≥8% is verified. Through fine adjustment to ensure training effectiveness, and through fatigue relief to improve long-term compliance, the precise adaptation of pathological characteristics and technical parameters is achieved.

[0059] Based on historical data of similar users, the algorithm fits the association between the duration and the improvement of the cornea. At 10 minutes, the irregular area of the cornea is continuously stimulated by multi-spectrum green light (530 nm) with 5% contrast adjustment, which makes the smoothness of the cornea increase by 8% and the irregularity decrease by 8%; at 15 minutes, the stimulation time is extended, the synergistic effect of green light and red light is enhanced, and the improvement rate is superimposed to 12%.

[0060] When the user terminal receives the training prediction result, the association curve between the training duration and the expected effect is displayed synchronously (such as 10 minutes of training corresponding to an 8% decrease in corneal irregularity, and 15 minutes corresponding to a 12% decrease); after the user selects the current training duration, the control module links the multi-spectrum light source module, and if the duration is shortened, the green light peak wavelength is automatically increased (such as from 530 nm to 550 nm) to enhance the activation efficiency of cone cells and offset the impact of insufficient duration.

[0061] Specifically, the intelligent analysis module is also used for multi-dimensional classification according to user personal information and vision diagnosis results, statistical analysis of the attenuation model and the training improvement model corresponding to each category, extraction of universal user features in similar users according to the distribution of user numbers in different categories, marking of high-quality features with better training recovery effect than universal user features and weak features with worse recovery effect than universal user features, extraction of personal information and vision features for users who use the multi-spectrum dynamic visual function training system for the first time, and matching with universal user features, high-quality features, and weak features to locate the user's category. Based on the positioning result, the attenuation model and the training improvement model of the corresponding category are used as the initial template and are bound with the user's personal information, and the reinforcement parameters corresponding to the high-quality features or the compensation parameters corresponding to the weak features are synchronously integrated.

[0062] When extracting universal user features, the intelligent analysis module includes the association between the training duration and the attenuation duration in the statistical dimension; when marking user features, it combines the eye feature parameters in the image information; for users who use it for the first time, if there is a difference between the eye feature parameters and the universal features after feature matching, adjust the attenuation duration in the training prediction result and link the training execution module to adjust the training parameters.

[0063] More specifically, for the same type of users (in this embodiment, taking the primary and secondary school student group as an example), 6 preset precision categories are divided according to the user personal information and the vision diagnosis results, and the sample size of each precision category is not less than 500 (experimental data shows that when the sample size is greater than or equal to 300, the K-means clustering error is less than 3%). Among them, the personal information is divided into age and gender, and the age is divided into 7-12 years old, 13-18 years old, and more than 18 years old. The vision diagnosis results are divided into axial length, astigmatism degree, and corneal curvature; the axial length is divided into <22mm, 22-26mm, and >26mm; the astigmatism degree is divided into 0-100 degrees, 101-200 degrees, and >200 degrees; and the corneal curvature is divided into 40-45D and >45D.

[0064] The improved K-means clustering algorithm is used for universal feature extraction, and the training time (t), the decay time (td), the corneal irregularity (C), and the pupil response speed (V) are used as inputs. The target function clearly reflects the weight distribution of the four parameters, and is specifically shown in the following formula (1):

[0065] (1),

[0066] Wherein, n is the sample number in a single preset precision category, and n≥500; is the training time of the ith sample, is the corresponding decay time, is the corneal irregularity of the ith sample, is the pupil response speed of the ith sample; , , , The cluster center mean of the four types of parameters is respectively. The value of J is minimized through iteration. When extracting the core universal features, the average training time threshold (such as the 7-12-year-old astigmatism 101-200-degree group minutes); the average decay time threshold (the same group hours); the average corneal irregularity and the improvement rate (the same group / second); and the average pupil response speed (the same group mm / ms).

[0067] The universal characteristics of this category are obtained by K-means clustering (AC0=8% / time, μ_V=0.2mm / ms, document universal characteristics example); then the AC (corneal irregularity improvement rate) and V (pupil reaction speed) of 500 samples are sorted, and the "top 20%" users (corresponding to AC average higher than AC0+15% and V average higher than μ_V+10%) are defined as high-quality users; the "bottom 15%" users (corresponding to AC average lower than AC0-10% or V average lower than μ_V-10%) are defined as weak users. High-quality users are users with and in the same category, and the training parameters are red light wavelength 640nm (to increase the penetration depth) and green light switching frequency increased by 20%; weak users are users with or in the same category, and the compensation parameters are picture contrast reduced by 5% and red light duration extended by 15%.

[0068] When matching the initial user, the feature matching degree is calculated using the following formula (2), which directly substitutes the difference between the four parameter measured values and the category average:

[0069] +0.3 (2),

[0070] Formula (2) contains five parameters, respectively corresponding to training time (t), decay time (td), corneal irregularity (C), pupil reaction speed (V) and axial length (A), and the weights are 0.1, 0.1, 0.25, 0.25 and 0.3 respectively. This weight distribution is based on the influence degree of the parameters on the vision training. Axial length (A) is the core indicator for the prevention and control of myopia in adolescents (directly related to the eye development stage), so it has the greatest influence on the adaptability of the training program (such as red light penetration depth and green light stimulation intensity), and therefore is given the highest weight (0.3); corneal irregularity (C) and pupil reaction speed (V) directly reflect the integrity of the eye structure and the activity of the visual cells (key features of visual function development in primary and secondary school students), so they are given the second highest weight (0.25 each); training time (t) and decay time (td) are time dimension parameters, which can be compensated by dynamic adjustment (such as increasing the green light wavelength when shortening the time), so they are given the lowest weight (0.1 each).

[0071] Taking corneal irregularity (C) as an example, represents the relative difference between the user's measured value and the category average. If the user's C value is equal to the category average , then this item is 1 (i.e. complete matching); if the user's C value is equal to the maximum value or the minimum value , then this item is 0 (i.e. complete mismatch). The calculation logic of other parameters such as axial length (A), pupil response speed (V) is the same, and the parameter difference is converted into positive matching degree by "1 minus relative difference".

[0072] The total matching degree S is obtained by multiplying the five single matching degrees by the corresponding weights and summing them up. When S≥0.7 (the threshold of different S is tested on the training set, which is 0.6, 0.65, 0.7, 0.75, 0.8 respectively; it is found that when S is 0.7, the matching rate is >85%, and the proportion of the deviation value D of the training effect after matching is <0.2 is the highest), it is determined that the user matches the category, and the system automatically binds the attenuation model and the training improvement model of the category as the initial template. This threshold setting takes into account the regularity of primary and secondary school students (such as the rapid development characteristics of 7-12 year-old axial length) and individual differences (such as different corneal curvatures at the same age), which avoids mismatching caused by small parameter differences (such as a difference of 0.5mm in axial length), and ensures that the initial template is basically consistent with the actual vision status of the user.

[0073] In the statistical dimension of the correlation relationship, the correlation between the training duration and the attenuation duration is calculated as t and td in the five parameters, as shown in the following formula (3):

[0074] (3),

[0075] When , it is marked as a high attenuation sensitive category (such as "13-18 year-old + astigmatism > 200 degrees + axial length > 26mm" of primary and secondary school students), and t and td in this group show a strong negative correlation (such as a 10% increase in training duration shortens the attenuation duration by 12%).

[0076] The Pearson correlation coefficient r is used to quantify the linear relationship between the training duration (t) and the attenuation duration (td), and its value range is ; n represents the sample size 500, and i represents the index of the sample (from 1 to n in turn). When , it indicates that they are strongly negatively correlated, that is, when the training duration increases, the attenuation duration (the duration of the training effect) will be significantly shortened. For example, in the 13-18 year-old, astigmatism > 200 degrees and axial length > 26mm group, the coefficient can be as low as -0.65, which means that for every 10% increase in training duration, the attenuation duration will be shortened by 12%. This statistical dimension breaks through the limitation of traditional single parameter (such as only looking at the training duration or only looking at the attenuation duration), and by mining the correlation rules in the time dimension, it can predict the persistence of the training effect in advance, and provide data support for the adjustment of programs for high-risk groups such as adolescents in the rapid development period.

[0077] For first-time users, if the user's axial length , and the corneal irregularity ​ , (For the cluster mean of the matching categories), then the decay coefficient is used by the user. (Limited to 1.3-1.8), adjusted decay time During the linkage adjustment, the red light wavelength is 655nm, the pulse frequency is reduced from 2Hz to 1.5Hz, and the screen contrast gradient is reduced from 5% to 4%.

[0078] Initial corneal irregularity in first-time users ( ) and pupillary reaction speed ( When a value deviates from the universal characteristics of its category, the formula for calculating the attenuation coefficient K will simultaneously incorporate the differences in two parameters: / Among them, users Reflecting the degree of corneal irregularity exceeding the standard, the user Reflecting the degree of lag in pupillary response speed, both factors work together to cause K to exhibit a dynamic change where the more severe the corneal irregularity, the slower the response speed, and the larger the coefficient (as in the case of user). , ,user , =0.15、 Calculated At the same time, the red light wavelength is adjusted to 655nm (to enhance the choroidal repair ability), the pulse frequency is reduced to 1.5Hz (to reduce stimulation), and the image contrast gradient is reduced to 4% (to reduce the burden on the cornea), forming a multi-parameter collaborative compensation mechanism.

[0079] More specifically, the verification index integrates four parameters to show the improvement degree as shown in the following formula (4):

[0080] (4),

[0081] in, The expected improvement rate is calculated by the training improvement model based on the user's initial eye model and training parameters (such as light wavelength and training duration). For example, it is predicted that the corneal irregularity of a certain astigmatic user will be reduced by 8% after training. This represents the difference between the corneal irregularity measured by the image acquisition module after training and before training (e.g., an actual reduction of 7.8%).

[0082] Weighting ( ,Right now 30%, Similarly, ,Right now (Accounting for 20%) reflects the principle of prioritizing core visual functions. Time (as in the case) This indicates that the adjusted training effect deviates little from the expected result, and no further adjustment is needed; if If so, repeated adjustment is triggered, up to a maximum of 5 times. This mechanism ensures the stability of the training effect through quantitative indicators, and avoids parameter disorder caused by excessive adjustment through the number limit, realizing the closed-loop management of effect verification and dynamic correction.

[0083] The intelligent analysis module is further configured to compare the attenuation model, the training improvement model and the positioning result of the current user with the universal user features of the corresponding category, generate current state information containing a training effect deviation value and send it to the user terminal; obtain the individual eye model and the training parameter calculation state change trend of the user at a preset period; combine a preset emotional support information library including training bottleneck period encouragement and effect improvement congratulations, match the accompanying guide information from the emotional support information library according to the state change trend and the training effect deviation value, and send it to the user terminal synchronously.

[0084] Specifically, the training effect deviation value is calculated by a weighted comprehensive algorithm, as shown in the following formula (5):

[0085] (5),

[0086] wherein, is the actual corneal irregularity improvement rate of the user, is the universal value; the user is the actual attenuation duration, is the universal threshold value; the user is the actual pupil reaction speed, is the universal mean value. The specific values of the weights 0.4, 0.3 and 0.3 are obtained by training effect data of a large number of similar primary and secondary school students (such as groups of different ages and vision problem categories). is the effect matching, is the slight deviation, is the significant deviation.

[0087] Formula (5) integrates the corneal irregularity improvement rate (R), the attenuation duration (T) and the pupil reaction speed (P) by weighting. ​​The relative deviation of the same universal feature is realized. First, avoid one-sidedness of single index evaluation, give weight distribution (AC accounts for 40%, td and V each accounts for 30%) to highlight the core improvement index, while considering the effect persistence and eye dynamic function, more comprehensive reflect the training "quality" and "quantity". Second, through standardization classification (i.e. D<0.1 is effect matching, 0.1≤D<0.2 is mild deviation, D≥0.2 is significant deviation), clear positioning label is provided for users and system, and professional data is visualized. Third, link the front and rear technical links, reuse the universal feature as the benchmark, and its result directly provides quantitative basis for parameter adjustment (such as triggering attenuation coefficient optimization when significant deviation) and emotion guidance (such as pushing explanatory speech when mild deviation), realizing the closed-loop cooperation of group rule, individual deviation and dynamic response.

[0088] The state change trend is calculated by 3-period moving average method, as shown in the following formula (6):

[0089] (6),

[0090] Wherein is the current period deviation value, , is the value of the previous two periods. is the deviation expansion, is the trend stable, is the deviation reduction.

[0091] The trend algorithm accurately captures the long-term change direction of the user state through the average change rate of the 3-period deviation value (D). First, filter short-term fluctuation interference, eliminate accidental factors (such as user fatigue) by moving average, and get closer to the true trend (such as Trend=-0.03 reveals that the deviation is continuously reduced). Second, realize the pre-judgment intervention, start the response strategy in advance through the trend symbol (i.e. Trend>0 is deviation expansion, <0 is reduction, ≈0 is stable), expand the training parameters when expanding, strengthen the encouragement and guidance when reducing, and maintain the scheme when stable, so that the scheme is upgraded from passive response to active guidance. In addition, it can also enhance the user's training resilience, especially for children or weak vision groups, use trend data (such as a 3% reduction in deviation per week) instead of absolute results to provide process evidence for emotional guidance, alleviate the anxiety caused by short-term effect fluctuation, improve the possibility of long-term adherence, and realize the deep integration of technical parameters and psychological support.

[0092] The deviation value algorithm (D) and the trend algorithm (Trend) form a golden combination of static positioning and dynamic tracking: D tells you where the user is now (i.e., the gap with the group), and Trend tells you where the user is going (i.e., the direction of change). This combination avoids both the shortsightedness of looking only at the current state (e.g., slight deviation but deteriorating trend needs to be vigilant) and the ambiguity of looking only at the trend (e.g., good trend but still significant deviation needs to be adjusted). For example, a 15-year-old high myopia user with D = 0.18 (i.e., slight deviation) and Trend = -0.02 (i.e., deviation reduction), the program will combine the two to judge that the current level is slightly lower than the group level, but is improving steadily, so it matches the positive progress combined with the guidance of maintaining the program, which is both professional and rigorous and fits the user's psychology, achieving a deep integration of technical parameters and humanistic care.

[0093] According to the state change trend and the training effect deviation value, the accompanying guidance information is matched from the emotional support information library, as shown below:

[0094] When the effect matches and the trend is stable, the accompanying guidance information can be that the current training effect is consistent with the average level of the myopic group of the same age, and the corneal irregularity improvement rate is stable at 8%, which can continue to maintain the existing rhythm.

[0095] When the deviation is slight and the deviation is reduced, the accompanying guidance information can be that the decay duration is 5% shorter than the group average, but the reduction in the shortening amplitude is reduced by 3% compared with the previous week, the repair stability is improving, and it is recommended to practice once a week to consolidate the effect.

[0096] When the deviation is significant and the deviation is enlarged, the accompanying guidance information can be that the pupil reaction speed is 22% slower than the group average, and the gap is enlarged by 5% compared with the previous two weeks, which is consistent with the volatility of 12-year-old children's visual development. Next week, the green light frequency will be adjusted to 5Hz (originally 4Hz) to help improve reaction sensitivity.

[0097] More specifically, the deviation value calculation reuses the universal features (i.e. , ) and the decay duration correlation law, and the trend analysis data feeds back parameter adjustment (such as triggering secondary optimization of training parameters when the deviation is enlarged), forming a closed loop of feature extraction, deviation analysis, parameter adjustment, and emotional guidance.

[0098] Among them, after the intelligent analysis module generates the training effect deviation value and the state change trend, if the training effect deviation value is in the significant deviation interval and the trend is deviation enlargement for two consecutive preset periods, the decay coefficient in the decay model is automatically called to optimize the decay model, and the basic parameters of the decay model are dynamically adjusted based on the current user's eye feature parameter changes; At the same time, the light stimulation parameter weight in the training improvement model is optimized synchronously, so that the adaptability of the decay model and the training improvement model is iterated in real time with the state change trend.

[0099] The intelligent analysis module compares the group trend change formed by the same type of users with the state change trend corresponding to the individual user when acquiring the current state information by period, and when the deviation rate between the state change trend corresponding to the individual user and the group trend exceeds the preset deviation threshold, the training parameters of the individual user are supplemented to the same type of universal features as a sample, and the correlation of the attenuation model is optimized in reverse, realizing the closed loop of individual optimization, group feature iteration and model universality improvement.

[0100] Specifically, the intelligent analysis module performs model optimization checking every 14 days (i.e., a preset period), calculates the training effect deviation value D, and when D≥0.2 (i.e., a significant deviation interval) and the state change trend Trend>0.03 (i.e., a trend expansion) for two consecutive times, the attenuation coefficient in the attenuation model is automatically called to optimize the attenuation model. When optimizing the attenuation model, the attenuation coefficient K is dynamically corrected based on the user's eyeball feature parameter change (corneal irregularity fluctuation ΔV=the difference between the current V and the V in the first period), and the correction formula is shown in the following formula (7):

[0101] (7),

[0102] Koptis the optimized attenuation coefficient, Koptis the optimized attenuation coefficient. Koptis the optimized attenuation coefficient.

[0103] The basic parameters of the attenuation model (such as the weight of the baseline attenuation duration Td0) are also adjusted (from 0.3 to 0.3+0.02×Trend).

[0104] The gradient descent method is used to update the light stimulation parameter weight (red light wavelength λ weight w1, green light frequency f weight w2), and the loss function is a weighted combination of multi-dimensional deviations, and the loss function is shown in the following formula (8):

[0105] (8),

[0106] The learning rate η=0.01, and the iteration is stopped until Loss<0.05, ensuring that the light stimulation parameters and the attenuation model are cooperatively adapted.

[0107] wherein, Doptis the training effect deviation value predicted by the attenuation model, Doptis the measured deviation value (deviation value algorithm); Doptis the corneal irregularity improvement rate predicted by the attenuation model, Doptis the actual improvement rate of the user; Doptis the pupil reaction speed predicted by the model, The user's actual speed is measured. The weight distribution (0.5, 0.3, 0.2) reflects the priority of the parameters: the deviation value (D) directly reflects the overall adaptability, with the highest weight; the corneal irregularity improvement rate (ΔC) is the core functional indicator, with the second highest weight.

[0108] This loss function can integrate multi-dimensional errors (not a single indicator) to avoid model optimization bias towards a certain parameter while ignoring the overall effect; at the same time, as a navigator for optimization algorithms such as gradient descent, it guides the adjustment of parameters in the direction of loss reduction by calculating the partial derivative of the loss function with respect to the light stimulus parameter weight (such as the red light wavelength weight w1) (such as the iterative process of learning rate η=0.01); Furthermore, it forms a linkage with the deviation value and the trend algorithm: when the loss function value falls below the threshold (such as Loss<0.05), it indicates that the model optimization has met the standard, and iteration can be stopped to ensure the stability of parameter adjustment.

[0109] Write the updated , Td0 weight, w1, w2) to the user's local model (decay model) and synchronize it to the initial template candidate library of similar users in the cloud.

[0110] The intelligent analysis module calculates the group and individual trends (i.e., the state change trend of the individual user) every 14 days. The group trend is the average of the Trend of similar users in the past 3 periods; the individual trend is the Trend of the current user in the past 3 periods, and the deviation rate is calculated as shown in the following formula (9):

[0111] (9),

[0112] When >15% (preset deviation threshold), optimization is started. Effective training parameters (such as red light wavelength , single training duration ) of the individual are selected, which need to meet the condition that ΔC is improved by ≥8% after application. For the selected effective parameters, a dynamic weighting method is used to integrate them into the universal characteristics of the similar group. Taking the red light wavelength as an example, the updated formula is shown in the following formula (10):

[0113] (10),

[0114] is the original universal red light wavelength of the group (such as 640nm), is the effective red light wavelength of the individual user (such as 655nm), and D is the current training effect deviation value of the individual (reuse the deviation value algorithm, range 0-1).

[0115] Among them, 80% of the original group characteristic weight is retained (i.e., 0.8× ), to ensure the stability of the group characteristics and avoid drastic fluctuations caused by individual parameters;

[0116] The individual parameter weight is 20% of the base weight multiplied by a dynamic adjustment factor (1-0.5xD): the better the individual effect (D is smaller, such as D=0.1, the adjustment factor=0.95), the higher the actual weight (0.2x0.95=0.19), which strengthens the influence of high-quality individual parameters; when the individual effect is average (D=0.3, the adjustment factor=0.85), the weight is correspondingly reduced (0.2x0.85=0.17), avoiding excessive interference of low-efficiency parameters on the group characteristics.

[0117] Based on the new universal characteristics, the correlation coefficient r of the training duration and the decay duration is recalculated, and the factor in the correlation formula is adjusted (such as the original formula coefficient 0.02 is adjusted to 0.02+0.001x ), and the updated model is pushed to all user terminals of the same type. Parameter adjustment-emotion guidance closed loop.

[0118] Specifically, the intelligent analysis module dynamically adjusts the training parameter threshold based on the model change rate after updating the decay model or training improvement model parameters, and corrects the threshold by weighted integration of the decay coefficient change rate, so that the change rate correction threshold and the decay model or training improvement model iteration are coordinated, and single threshold adjustment is realized while adapting to the decay characteristics and stimulation effect.

[0119] The intelligent analysis module adjusts the training parameter threshold in layers after updating the group universal characteristics or optimizing the individual model, and the training parameter threshold includes the training duration threshold, the red light power threshold and the pulse frequency threshold.

[0120] More specifically, the variable definition As the decay model slope change rate (% / cycle), the calculation formula is shown in equation (11) as follows:

[0121] (11),

[0122] Define the variable As the training improvement model slope change rate (% / cycle), the calculation formula is shown in equation (12) as follows:

[0123] (12),

[0124] Define the threshold base value set , respectively corresponding to the training duration threshold (such as 15 minutes), the red light power threshold (such as 50 cd / m²), and the pulse frequency threshold (such as 2 Hz).

[0125] Define the correction weight , (wherein, based on 200 examples of the same user cross-validation, the balance attenuation characteristics and the stimulation effect weight are determined).

[0126] When the single threshold value is cooperatively corrected, the threshold value correction formula is shown in the following formula (13):

[0127] (13),

[0128] Forced restriction to prevent the threshold value fluctuation from exceeding the safe range and ensure the training stability.

[0129] The iteration trigger condition is that formula (13) is automatically called after each attenuation model or training improvement model parameter update; if or the threshold value is immediately corrected in real time and is synchronized to the user terminal and the cloud backup (for reference when a new user is matched).

[0130] When the hierarchical threshold value is adjusted, it is divided into group layer adjustment and individual layer adjustment. The group layer adjustment is that when the group slope in the universal characteristics of the same user is the update amplitude , the default threshold value of all users in this category is recalculated once according to formula (13), and is synchronized to the universal characteristic library as the reference of the initial threshold value of the new user. The individual layer adjustment is that if the individual slope of a user is changed by after the individual model is updated, the threshold value of the user is only corrected, and the group default threshold value remains unchanged.

[0131] The priority of the individual layer adjustment is higher than that of the group layer, so as to avoid that the group update covers the individual optimization result; meanwhile, if the threshold value deviation rate (the difference from the group default value) of the individual adjustment is it is automatically fed back to the universal characteristic iteration mechanism as a sample, so as to promote the group characteristic optimization.

[0132] The above only is the embodiment of the present application, and the well-known specific structure and characteristics in the scheme are not described too much. It should be pointed out that for those skilled in the art, without departing from the structure of the present application, a number of deformations and improvements can be made, which should also be regarded as the protection range of the present application, and these will not affect the effect and practicality of the present application. The protection range required by the present application should be subject to the content of its claims, and the specific implementation mode and the like in the description can be used to explain the content of the claims.

Claims

1. A multi-spectral dynamic visual function training system, characterized in that, The application relates to an intelligent myopia prevention and control system, which comprises the following parts: An image acquisition module arranged at a corresponding position of an eye part, which is used for acquiring image information near an eyeball of a user in real time, wherein the image information comprises an eyeball movement track, pupil size and shape, a corneal edge contour and an eye white elasticity state; An intelligent analysis module, which is combined with the image information and the acquisition time of the image information, analyzes pupil light reaction speed according to the eyeball movement track and the pupil size and shape, analyzes corneal irregular region distribution according to the corneal edge contour and the eye white elasticity state, integrates the pupil light reaction speed and the corneal irregular region distribution into eyeball characteristic parameters, and stores the eyeball characteristic parameters, corresponding image information and acquisition time of the image information in association; personal information and vision diagnosis results of the user are acquired, and an individual eye part model is established by combining the eyeball characteristic parameters, the personal information and the vision diagnosis results; A control module, which generates a personalized training scheme according to the individual eye part model, wherein the personalized training scheme comprises a training picture, and each training parameter in the training scheme is updated according to the image information; A multi-spectrum light source module, which is used for emitting light waves in a visible light wavelength range according to each training parameter, wherein green light with a preset green light wavelength is used for exciting cone cells, and red light with a preset red light wavelength is used for stimulating choroid blood circulation; the intensity and switching frequency of the light waves are dynamically adjusted according to the eyeball characteristic parameters; A training execution module, which comprises a contrast adjustment unit and a cyclic stimulation unit; the contrast adjustment unit acquires the corneal irregular region distribution, and adjusts the contrast of the training picture according to the corneal irregular region distribution, so as to simulate a visual environment required for myopia prevention and control; the cyclic stimulation unit adjusts the red light irradiation wavelength of the multi-spectrum light source module through the eyeball characteristic parameters, promotes eye water circulation and blood circulation, and thickens choroids; and the training execution module sends the adjusted training parameters to the multi-spectrum light source module.

2. The multi-spectral dynamic visual function training system of claim 1, wherein: The intelligent analysis module is also used for acquiring an individual eye part model before training as an initial model, acquiring an individual eye part model at the end of training as a second model, taking the interval time length between the initial model and the second model as a training time length, and constructing a training improvement model by combining the training parameters in the training process, the initial model and the second model; An individual eye part model before the next training after the end of training is acquired as a third model, a time length between the second model and the third model is acquired as a decay time length, and a decay model is constructed according to the second model, the third model and the decay time length; The intelligent analysis module sets a training parameter threshold value according to the personal information and the vision diagnosis results of the user, wherein the threshold value is adjusted differently according to the axial development stage and the astigmatism degree of users of different ages. The intelligent analysis module identifies the initial model, the second model, the third model and the corresponding collection time according to the collection time of the image information and the eyeball characteristic parameter, inputs the initial model and the corresponding training parameter into the training improvement model, outputs different training durations and the corresponding second model as the training prediction result under the constraint of the training parameter threshold, and then inputs the second model into the attenuation model to output different attenuation durations and the corresponding third model as the attenuation result of the training prediction result; the training prediction result and the attenuation result of the training prediction result are integrated into the training prediction result and sent to the user terminal, and the training duration fed back by the user terminal is received as the current training duration, and the control module adjusts the personalized training scheme according to the current training duration.

3. The multi-spectral dynamic visual function training system of claim 2, wherein: The intelligent analysis module is also used for multi-dimensional classification according to user personal information and vision diagnosis results, statistical analysis of the attenuation model and the training improvement model corresponding to each category, and extraction of the universal user characteristics in the same category according to the distribution of the number of users in different categories; The high-quality characteristics with better training recovery effect than the universal user characteristics and the weak characteristics with worse recovery effect than the universal user characteristics are marked out; For the user who uses the multi-spectrum dynamic visual function training system for the first time, the personal information and the vision characteristics are extracted and matched with the universal user characteristics, the high-quality characteristics and the weak characteristics to locate the category to which the user belongs, the attenuation model and the training improvement model of the corresponding category are taken as the initial template and bound with the user personal information based on the positioning result, and the reinforcement parameters corresponding to the high-quality characteristics or the compensation parameters corresponding to the weak characteristics are synchronously integrated.

4. The multi-spectral dynamic visual function training system of claim 3, wherein: When extracting the universal user characteristics, the intelligent analysis module takes the correlation between the training duration and the attenuation duration into the statistical dimension; when marking the user characteristics, the eyeball characteristic parameters in the image information are combined; for the user who uses the system for the first time, if there is a difference between the eyeball characteristic parameters and the universal characteristics after the characteristic matching, the attenuation duration in the training prediction result is adjusted, and the training execution module is adjusted to adjust the training parameter.

5. The multi-spectral dynamic visual function training system of claim 4, wherein: The intelligent analysis module is also used for comparing the attenuation model, the training improvement model and the positioning result of the current user with the universal user characteristics of the corresponding category, generating current state information containing a training effect deviation value and sending the current state information to the user terminal; The individual eye model and the training parameter calculation state change trend of the user are obtained at a preset period, the state change trend and the training effect deviation value are combined, the accompanying guide information is matched from the emotional support information library according to the state change trend and the training effect deviation value, and the accompanying guide information is synchronously sent to the user terminal.

6. The multi-spectral dynamic visual function training system of claim 5, wherein: After the intelligent analysis module generates the training effect deviation value and the state change trend, if the training effect deviation value is in a significant deviation interval and the trend is deviation expansion in two continuous preset periods, the attenuation coefficient in the attenuation model is automatically called to optimize the attenuation model, the basic parameters of the attenuation model are dynamically adjusted based on the change of the eyeball characteristic parameters of the current user, the light stimulation parameter weight in the training improvement model is simultaneously optimized, and the adaptability of the attenuation model and the training improvement model is iterated in real time with the state change trend.

7. The multi-spectral dynamic visual function training system of claim 5, wherein: The intelligent analysis module compares the group trend change formed by the same type of users with the state change trend corresponding to the individual user when obtaining the current state information by period, and when the deviation rate between the state change trend corresponding to the individual user and the group trend exceeds the preset deviation threshold, the training parameters of the individual user are supplemented to the same type of universal features as samples, and the correlation of the decay model is simultaneously optimized in the reverse direction, realizing the closed loop of individual optimization, group feature iteration and model universality improvement.

8. The multi-spectral dynamic visual function training system of claim 6 or 7, wherein: After updating the parameters of the decay model or the training improvement model, the intelligent analysis module dynamically adjusts the training parameter threshold value according to the model change rate, corrects the threshold value by weighted integration of the decay coefficient change rate, makes the change rate correction threshold value and the decay model or the training improvement model iteration cooperative, and realizes the adjustment of the single threshold value while adapting the decay characteristics and the stimulation effect.

9. The multi-spectral dynamic visual function training system of claim 8, wherein: After the group universal feature is updated or the individual model is optimized, the intelligent analysis module adjusts the training parameter threshold value in layers, and the training parameter threshold value includes a training duration threshold value, a red light power threshold value and a pulse frequency threshold value.

Images