Self-adaptive visual training method based on eye characteristics

Abstract

The invention relates to a self-adaptive visual training method based on eye features, and belongs to the technical field of visual health and artificial intelligence. In order to solve the problems that traditional visual training equipment is heavy in structure, single in training mode and lack of personalized regulation and control and concentration state monitoring, the visual training equipment obtains eye images of a user in real time through an image acquisition module, adopts visual intelligent analysis, extracts double features of eye contours and pupil positions, calculates the visual concentration degree, and improves the visual training efficiency. And a concentration state is judged by combining a dynamic self-adaptive threshold value. When the concentration degree is insufficient, the control module triggers intervention mechanisms such as prompt or training pause and the like, and the remote interaction module uploads data to the cloud platform to generate a personalized training scheme. Intelligent and personalized regulation and control of the training process are achieved, the training effect and compliance are improved, the system is compact in structure and suitable for various scenes, and an efficient solution is provided for vision health.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of vision health and artificial intelligence, and relates to an adaptive vision training method based on eye features. BACKGROUND

[0002] With the widespread popularity of electronic products and the significant increase in close-range eye behavior, the incidence of myopia in adolescents continues to rise, and vision health problems are increasingly concerned. Traditional vision training methods mainly rely on training equipment based on ciliary muscle accommodation, such as fog mirrors or visual tracking instruments, but these devices have many shortcomings in practical application. First, existing devices are usually bulky, inconvenient to carry and use, and difficult to adapt to the training needs of families or diverse scenarios. Second, the training method is relatively single, lacking diverse visual stimulation and interactive design, making it difficult to maintain the training interest and compliance of users. In addition, existing devices generally lack intelligent judgment mechanisms and cannot dynamically adjust training content according to the real-time state of users, nor do they have effective monitoring and feedback on the user's concentration state. This makes the training process difficult to achieve individualization and unable to optimize training effectiveness according to individual differences.

[0003] In recent years, artificial intelligence visual analysis technology has begun to be applied in the field of eye feature recognition and analysis, which can extract features such as pupil position, eye movement trajectory, and blink frequency by processing eye images. However, the application of existing AI visual analysis technology in vision training is still limited, and most schemes only stay at the basic level of eye feature detection, failing to deeply combine with dynamic regulation of the training process. In addition, existing technologies lack quantitative evaluation mechanisms for user concentration states and do not fully utilize cloud data analysis to achieve remote optimization and generation of individualized training programs. Therefore, there is an urgent need for a vision training method that can monitor user eye features in real time, dynamically evaluate concentration states, and optimize training programs in combination with cloud data, to improve training effectiveness and meet the needs of different users. SUMMARY

[0004] In view of the above, the purpose of the present application is to provide an adaptive vision training method based on eye features, which monitors the user's gaze behavior and eye movement features in real time, dynamically judges their concentration state, and triggers appropriate interventions when distraction or fatigue occurs, to improve the effectiveness of vision training.

[0005] To achieve the above purpose, the present application provides the following technical solutions:

[0006] An adaptive vision training method based on eye features, comprising the following steps:

[0007] S1: Real-time acquisition of eye images of the user through an image acquisition module;

[0008] S2: Processing of the eye images through a visual intelligence analysis module, extraction of eye features and calculation of visual concentration, wherein the eye features include eye contour features, pupil center position, eye movement trajectory, gaze duration, blink frequency and visual line deviation;

[0009] S3: Comparison of the visual concentration with a dynamic adaptive threshold to determine the user's concentration state;

[0010] S4: When the visual concentration is lower than the dynamic adaptive threshold, triggering an intervention mechanism through a control module, including issuing a prompt, pausing training or adjusting training tasks;

[0011] S5: Uploading the visual concentration and training data to a cloud platform through a remote interaction module to generate and issue a personalized training plan.

[0012] Further, the visual intelligence analysis module in S2 uses a double-feature fusion model for eye feature extraction, including:

[0013] S21: Extracting geometric feature points of the eye corner and eyelid margin through eye contour detection, and constructing an eye structure constraint region;

[0014] S22: Generating a pupil candidate point heat map through pupil detection;

[0015] S23: Spatially constraining and cropping the pupil candidate point heat map according to the eye structure constraint region to obtain a cropped heat map;

[0016] S24: Concatenating the cropped heat map with the feature vector of the eye structure constraint region and inputting it into a regression module to output the pupil center coordinates.

[0017] Further, S2 also includes further calculation of the eye movement trajectory, gaze duration, blink frequency and visual line deviation based on the pupil center coordinates, and calculation of the visual concentration quantitative value through the following visual concentration quantification function:

[0018] F = w1·T + w2·S - w3·B - w4·D

[0019] Where F is the visual concentration quantitative value, T is the continuous gaze duration, S is the gaze center stability score, B is the blink frequency, D is the visual line drift probability, and w1, w2, w3, w4 are preset empirical weights.

[0020] Further, the gaze center stability score is calculated by the following formula:

[0021]

[0022] wherein, σ x and σ y are the standard deviations of the two-dimensional coordinate distribution of the gaze center within the target region, D max is the maximum allowable offset distance constant of the region.

[0023] Further, the dynamic adaptive threshold is calculated by the following formula:

[0024] T threshold = μ + k·σ

[0025] wherein, μ is the sliding mean of the visual concentration in the recent period of time, σ is the fluctuation degree of the visual concentration, and k is an adjustment factor based on the training stage, user historical data, and rest period.

[0026] Further, the intervention mechanism in S4 includes suspending the training and starting the rest countdown when the visual concentration is continuously lower than the dynamic adaptive threshold for more than a preset duration, and entering the protection mode and suspending the training task if the cumulative interruption times exceed a preset threshold.

[0027] Further, the personalized training scheme in S5 includes the parameter settings of the brightness rhythm, audio synchronization mode, training duration, and intervention frequency, and the cloud platform generates the personalized training scheme by curve fitting and behavior statistical analysis on the uploaded visual concentration.

[0028] An adaptive vision training system based on eye features, comprising:

[0029] an image acquisition module for acquiring eye images of a user in real time;

[0030] a visual intelligent analysis module connected with the image acquisition module, for processing the eye images, extracting eye features, and calculating visual concentration;

[0031] a control module connected with the visual intelligent analysis module, for comparing the visual concentration with a dynamic adaptive threshold, judging the concentration state, and triggering an intervention mechanism;

[0032] a training execution module connected with the control module, for executing a visual training task according to the instructions of the control module;

[0033] a remote interaction module connected with the control module, for uploading the visual concentration and training data to a cloud platform and receiving a personalized training scheme.

[0034] Further, the visual intelligence analysis module comprises a double-feature fusion model, the double-feature fusion model comprising an eye contour detection, a pupil detection and a fusion module, wherein the eye contour detection is used to extract an eye structure constraint region, the pupil detection is used to generate a pupil candidate point heat map, and the fusion module is used to crop the heat map according to the eye structure constraint region and output a pupil center coordinate.

[0035] Further, the training execution module supports fixed target fixation training, target movement tracking training and audio rhythm stimulation training, and the parameters of the training task are dynamically configured by the remote interaction module according to the individualized training scheme of the cloud platform.

[0036] The present application has the advantages that the present application provides an eye feature-based adaptive vision training method and system, which can significantly improve the pertinence and effect of vision training by monitoring the fixation behavior and eye movement features of the user in real time, dynamically evaluating the concentration state and triggering the corresponding intervention mechanism. The vision training system of the present application realizes individualized and intelligent control, accurately extracts eye features and calculates visual concentration through the visual intelligence analysis module, judges the attention state of the user in combination with the dynamic adaptive threshold, and thus adjusts the training content or suspends the training in time, effectively avoids fatigue training and improves training compliance. The system adopts modular design, has compact structure and high integration, is suitable for various use scenarios such as families and institutions, and enhances the convenience and universality of use. In addition, through the collaborative work of the remote interaction module and the cloud platform, the system can continuously optimize the individualized training scheme based on big data analysis, further improving the training effect and user experience. The present application ensures the safety and stability of the training process through a closed-loop regulation mechanism, and provides an efficient and individualized vision training solution for different users.

[0037] Other advantages, objects, and features of the present application will be apparent to those skilled in the art from the following specification, in certain respects, are deemed obvious by virtue of the teachings provided herein and in others respects are pointed out in the following detailed description. BRIEF DESCRIPTION OF DRAWINGS

[0038] In order to make the objectives, technical solutions and advantages of the present application clearer, the preferred detailed description of the present application will be given below in combination with the drawings, in which:

[0039] Figure 1 a double-feature fusion model for accurate identification of eye features;

[0040] Figure 2 a visual attention detection and intervention model architecture based on multi-dimensional eye movement analysis;

[0041] Figure 3 A flow chart of a processing procedure for a visual concentration quantification model based on eye movement characteristics;

[0042] Figure 4 A flow chart of a visual concentration state detection and training interruption decision-making procedure based on a sliding window. DETAILED DESCRIPTION

[0043] The advantages and effects of the present application can be easily understood by those skilled in the art from the description of the specific embodiments of the present application. The present application can also be implemented or applied in other different specific embodiments, and various modifications or changes can be made to the details in the description based on different views and applications without departing from the spirit of the present application. It should be noted that the drawings provided in the following embodiments only illustrate the basic concept of the present application in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0044] The drawings are only used for illustrative description, and the representation is only a schematic diagram, not a physical diagram, and should not be understood as a limitation of the present application; in order to better illustrate the embodiments of the present application, some components in the drawings may be omitted, enlarged or reduced, and do not represent the actual size of the product; it is understandable for those skilled in the art that some well-known structures and their descriptions in the drawings may be omitted.

[0045] The same or similar reference numerals in the drawings of the embodiments of the present application correspond to the same or similar components; in the description of the present application, it should be understood that if the terms "upper", "lower", "left", "right", "front", "back" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, therefore the terms describing the positional relationship in the drawings are only used for illustrative description, and should not be understood as a limitation of the present application, for those skilled in the art, the specific meaning of the above terms can be understood according to the specific situation.

[0046] System structure and module configuration: the present application provides a visual training system, including a control module, an image acquisition module, a visual intelligent analysis module, a training execution module and a remote interaction module. Among them, the control module monitors the visual concentration in the training process in real time, and when the quantitative value continuously falls below the adaptive threshold, the training state intervention or interruption is carried out, and the training adaptability parameter is generated; the training execution module specifically implements the visual training task according to the instruction of the control module; the image acquisition module acquires the eye image of the user in real time, providing input for subsequent analysis; the remote interaction module uploads the visual concentration and other data in the training process to the cloud platform, which is used for user portrait construction and generation of personalized training suggestions.

[0047] Vision intelligence analysis module structure and task: The vision intelligence analysis module uses a double feature fusion model to process the collected eye images in real time on the edge device. The model includes two parts that work together, the first part is eye contour detection, which extracts key geometric feature points such as eye corners and eyelids, constructs eye structure constraint regions, and forms spatial boundary basis for pupil positioning. The second part is pupil detection, which detects the pupil to generate a pupil candidate point heat map. The output results of the above two parts are integrated through a fusion module, and the fusion process is divided into the following two steps:

[0048] S1: Spatial constraint cropping: The system crops the heat map generated by the pupil detection part according to the structure constraint region constructed by the eye contour detection part, and only retains the high response region within the eye structure region, suppressing the false response of the non-pupil region in the heat map.

[0049] S2: Feature fusion and coordinate regression: The cropped saliency heat map and eye structure feature vector are spliced and input into the subsequent regression module, and the final pupil center coordinates are output through convolution + fully connected structure.

[0050] On the basis of pupil center positioning, the vision intelligence analysis module further completes the following subtasks, including: eye movement trajectory extraction, gaze duration statistics, blink frequency estimation, and visual line offset detection. By fusing the analysis results of each subtask, the module calculates the user's visual concentration in real time, and synchronizes the quantitative value with the dynamic training task to guide the adjustment of the training process.

[0051] Vision attention quantification and adaptive threshold mechanism: The system defines a visual concentration quantification function based on multi-dimensional eye movement features to quantify the user's attention concentration. The system can dynamically calculate an adaptive threshold based on the historical quantification value sequence and combined with the user's individual state and training time. When the real-time quantification value is continuously below the threshold, the control module will trigger feedback mechanisms such as prompts or pauses in training to guide the user to restore attention and generate corresponding training adjustment parameters.

[0052] Cloud platform remote training optimization method: A cloud platform is constructed to remotely manage and optimize training data. Specifically, it includes:

[0053] S1: The cloud platform receives data such as visual concentration, training duration, and abnormal interruption records uploaded by the terminal;

[0054] S2: Based on visual concentration curve fitting and behavior statistical analysis, identify user training tolerance and attention sensitive period;

[0055] S3: Generate and issue personalized training programs, including brightness rhythm, audio synchronization mode, training duration, and intervention frequency parameter settings.

[0056] Through the method, the system can continuously optimize the training scheme according to the big data analysis result, and improve the training effect.

[0057] The visual attention quantitative function model based on eye movement characteristics: in order to quantitatively evaluate the attention state, the system designs a visual attention quantitative function in the form of weighted linear combination. Let the current continuous fixation time be T continuous (second), the fixation center stability score be S focus , the blink frequency be F blank (times / minute), and the visual line drift probability be P drift , and the experience weights be w1, w2, w3 and w4, then the visual attention can be expressed as:

[0058] Score=w1T continuous +w2S focus -w3F blank -w4P drift

[0059] Among them, the longer continuous fixation time and the higher stability correspond to the higher quantitative value, and the higher blink frequency and the visual line drift probability reduce the quantitative value. The weight parameter can be set and adjusted according to the actual training feedback experience. The quantitative value is updated at a fixed frequency, and is compared with an adaptive threshold to judge the visual attention of the user in real time.

[0060] I. Overall structure design of the system

[0061] The overall architecture of the system includes a control module, an image acquisition module, a visual intelligent analysis module, a training execution module and a remote interaction module. Each module is integrated in the main control system through an internal data bus or logical connection and runs on an edge processing platform (such as an embedded AI chip or a microcontroller with a neural network acceleration unit) to form a closed-loop control system.

[0062] In the system initialization stage, the control module first completes the hardware self-checking, including the camera state, the main control chip temperature, the power state and the like, and confirms that the device running environment is stable before entering the training process. The image acquisition module starts to continuously acquire eye image frame streams and transmits them to the visual intelligent analysis module; the control module synchronously activates the visual stimulation function of the training execution module.

[0063] II. Implementation method of the image acquisition module

[0064] The image acquisition module is composed of a camera assembly, an image preprocessing unit and a data buffer module. The camera assembly is fixedly installed in the user's forward direction and faces the eye area for continuously capturing eye images. The raw image data collected is first transmitted to the image preprocessing unit, which optimizes the image quality through processing procedures such as gray scale conversion, noise suppression and image enhancement, to improve the feature extraction accuracy and stability of the subsequent visual intelligent analysis module. The image data after preprocessing is stored in a ring buffer as an input data source for visual analysis tasks, ensuring the continuity and efficiency of the data flow.

[0065] III. Visual intelligent analysis module

[0066] The visual intelligent analysis module deploys a double-feature fusion model and runs in an edge-side local inference mode, which is suitable for embedded platforms such as ESP32 that support neural network acceleration. The task of this module is to extract key attention features from the eye images collected by the camera during the training process and calculate the visual concentration degree accordingly. The whole processing procedure is as follows:

[0067] 1. Pupil positioning and gaze center estimation

[0068] The image acquisition module provides continuous eye image frames, and the visual intelligent analysis module processes each image based on the double-feature fusion model to complete the accurate estimation of the pupil center position. The double-feature model includes eye contour detection and pupil detection.

[0069] Among them, the eye contour extraction is based on the lightweight face key point detection framework MediaPipe Face Mesh, which extracts multiple sets of eye structure key point coordinates including outer corner points, eyebrow arch center points, upper and lower eyelid boundary points, etc. The system constructs an eye geometric structure region based on this key point set to form a spatial constraint boundary where the pupil position may appear, which is used as a guide for subsequent fusion.

[0070] The pupil detection uses MobileNetV3 as the backbone network structure, which mainly includes three modules.

[0071] 1) Feature extraction network:

[0072] Perform layer-by-layer convolution operation on the preprocessed eye image to extract multi-scale visual features including texture shape, edge gradient, local brightness contrast, etc.

[0073] 2) Attention mechanism module:

[0074] Embed a lightweight attention structure (such as SE module or CBAM module) in the intermediate feature map layer to enhance the model's response ability to the pupil candidate area and suppress the interference of image background and eye outer edge;

[0075] 3) Heat map generation module:

[0076] The pupil saliency heat map is generated at the tail of the model by convolution and up-sampling structure, which represents the probability distribution value of each pixel point in the image as the pupil center.

[0077] The outputs of the above two feature extraction parts will be jointly transmitted to the geometric constraint fusion module. The system first extracts the key points such as the eye corner, eyebrow arch and eyelid margin according to the eye contour detection, and delimits an eye structure region in the image as the limited region where the pupil may appear. Then, the system processes the heat map generated by the pupil detection by taking this region as a constraint condition: only the response value within this region is retained, and the interference signal outside the region is suppressed, so as to improve the accuracy of the pupil position determination. On this basis, the system can also encode the key point coordinate data into a feature vector, and splice it with the feature results in the saliency heat map as the fusion input, and then make a comprehensive judgment through the subsequent neural network layer to output the final pupil center position.

[0078] The pupil center points located in each frame are mapped to the coordinate system of the display content to calculate the gaze direction of the user. If the two pupils point to the same direction and intersect at the center region of the interesting target, the gaze direction is defined as the "gaze center". The judgment basis of the gaze center is:

[0079] In the current frame, the deviation of the intersection point of the two pupil direction vectors from the center of the training image is less than the threshold Δθ;

[0080] Stable falling within the target region for a plurality of consecutive frames (e.g. >5 frames).

[0081] Only when the number of frames satisfying the deviation less than the threshold Δθ consecutively exceeds 5 frames, the system determines it as a valid gaze event, and adds the duration of the consecutive frames to T continuous .

[0082] 2. Blinking behavior recognition and frequency estimation

[0083] Based on the eye lid key points extracted by the double feature fusion model, the system dynamically calculates the geometric features such as the eye aspect ratio (EAR) representing the eye lid opening and closing state. When the EAR is below the set threshold in consecutive frames and quickly recovers, the system determines it as a blinking behavior. Within a set time window (e.g. 30 seconds or 60 seconds), the number of such events is counted to obtain the blinking frequency F blank (unit: times / minute).

[0084] 3. Gaze deviation detection

[0085] The system detects the degree and duration of the gaze center offset from the display target region according to the gaze center located by the double feature fusion model, and recognizes the gaze drift behavior. The judgment logic is as follows:

[0086] If the gaze center deviates from the target region of the training image by more than a set threshold;

[0087] And the deviation frame accounts for more than 30% of the total number of analysis frames;

[0088] The system determines that there is a line of sight drift behavior in the current training section, and calculates the deviation probability P drift , ranging from [0, 1].

[0089] 4. Gaze stability calculation

[0090] During the time period of valid gaze, the system records the two-dimensional coordinate distribution of the gaze center within the target region. By calculating the standard deviation (σ x ,σ y ) of these points and performing normalization processing, the system constructs the gaze stability score S focus , defined as follows:

[0091] S focus =1-(σ x 2 +σ y 2 ) / ρ 2

[0092] Where ρ is the maximum allowed deviation distance constant of the region. S focus The closer to 1, the more concentrated the gaze.

[0093] The system finally outputs the attention quantitative indicator Score in the following linear weighted form:

[0094] Score=w1T continuous +w2S focus -w3F blank -w4P drift , where:

[0095] T continuous : continuous effective gaze duration;

[0096] S focus : gaze stability score;

[0097] F blank : blink frequency (times / minute);

[0098] P drift : line of sight drift probability;

[0099] w1-w4: are preset empirical weights.

[0100] The score is updated every 5 seconds and provided to the control module to participate in training state intervention, visual attention detection, and behavior feedback generation.

[0101] IV. Scheduling and feedback mechanism of control module

[0102] The control module in the present application undertakes the real-time scheduling and behavior feedback tasks of the system running state, and the core responsibilities include: collecting and analyzing the visual attention quantification value sequence, dynamically generating threshold reference standard, judging whether the concentration state is abnormal, and making decisions such as prompting, pausing, resuming or protective termination based on the abnormality degree.

[0103] Firstly, the system receives the visual concentration degree output by the visual intelligent analysis module at a fixed time step, and maintains a sliding quantification value window internally. Based on the quantification value sequence in the last 5 minutes, the control module calculates the sliding mean value as the basic threshold reference value μ hist On this basis, the system introduces a time-dependent adjustment factor δ(t) to construct a dynamic threshold function:

[0104] θ(t) = μ hist + δ(t)

[0105]

[0106] Among them, δ(t) is composed of the following three factors: first, the visual concentration degree model parameters α and β corresponding to the current training stage (early, middle or late); second, the quantification value fluctuation degree σ u of the user's past training data, which reflects the stability of the concentration degree; the third is the last rest period τ, which reflects the current potential fatigue accumulation state. This dynamic threshold can more accurately reflect the current training tolerance of the individual, effectively avoiding false triggering and false intervention.

[0107] When the quantification value is lower than the dynamic threshold θ(t) in multiple consecutive time slices, the system will be considered as lack of attention and automatically enter the control response process. The response strategy is executed in stages: when the quantification value is first detected to decrease, the system plays a voice prompt through the audio module to guide the user to recover; if the quantification value is continuously lower than the threshold value for more than 30 seconds, the system automatically pauses the training, plays a rest prompt, and starts a fixed time of rest countdown; after the countdown is over, the system re judges whether the current visual concentration degree has recovered, if it is recovered above the threshold, the training process is restarted, and the training interruption state in this segment is recorded in the training log.

[0108] To prevent fatigue accumulation caused by overtraining, the system counts the number of interruptions in a longer period (e.g., 1 hour). If the number of attention interruptions exceeds the preset threshold (e.g., 3 times) in this period, the control module will automatically enter the protection mode, suspend the training task, enter the standby state, and prompt the user for manual intervention (e.g., decide whether to restart training by parents or administrators). This mechanism helps to ensure user training safety and prevent adverse experiences caused by fatigue training or excessive intervention.

[0109] Through the above scheduling and feedback strategy, the control module realizes closed-loop adaptive regulation of the training process, which can not only sensitively perceive the fluctuations in user visual concentration, but also intervene in the training state as needed to ensure the continuity, stability, and personalized adaptability of system operation.

[0110] Five, training execution module

[0111] This module exercises ciliary muscle regulation and focusing ability through various training tasks, typical tasks include:

[0112] 1. Fixed target fixation training: the user fixes his gaze on a stationary light-emitting target, and the LED brightness changes with the rhythm;

[0113] 2. Target movement tracking training: display the pattern slowly moving, train the user's eye movement ability;

[0114] 3. Audio rhythm stimulation training: play rhythmical sound with visual rhythm, stimulate concentration.

[0115] The rhythm parameters of all training tasks can be dynamically configured by the remote interaction module, realizing multi-mode switching such as "daily training", "visual attention detection", "recovery training", etc.

[0116] Six, remote interaction module and cloud platform data optimization process

[0117] The remote interaction module is mainly used to realize data interaction and strategy synchronization between the training device and the cloud platform, and to build a remote intelligent closed loop of the system. This module uploads terminal training data to the cloud platform through the MQTT protocol, and the data content includes: training start and end time period, visual concentration sequence, training interruption number, system internal adaptive parameters (such as θ(t), μ hist ), as well as user identification information, training period and environmental parameters, etc.

[0118] After receiving the data, the cloud platform first performs sliding curve fitting on the visual attention time series, extracts the attention change trend and identifies potential visual attention decay segments. Secondly, the system performs statistics on the fluctuation amplitude of visual attention, calculates the attention stability of the user during training, and identifies whether the individual training state is continuously low or shows obvious fluctuations. Finally, the platform generates user portraits such as "high-frequency distraction type", "fluctuation improvement type", "stable low type", etc. using behavior clustering algorithm, so as to implement more targeted training optimization strategies.

[0119] Based on the above analysis results, the platform can intelligently generate personalized recommendation schemes, including recommended training time period, light frequency parameters adapted to the user's attention rhythm, rhythm mode of audio prompts, and recommended single task duration, etc. After being reviewed and confirmed by the system administrator, the platform will issue the optimized parameters to the user terminal device, realizing the dynamic closed-loop regulation and control of remote-local collaborative training.

[0120] Seven, edge optimization strategy of visual attention quantification function module

[0121] In order to ensure that the quantification mechanism runs stably and efficiently on the edge device, the present application optimizes the implementation of the visual attention quantification function for edge computing. The specific measures are as follows:

[0122] 1. All input variables (such as T continuous , F blank , S focus , P drift ) are normalized before quantization to unify the dimension and reduce the complexity of floating point calculation;

[0123] 2. The EAR threshold judgment in blink detection uses a look-up table (Look-Up Table) instead of function calculation to reduce repeated logical judgment;

[0124] 3. The batch pre-computation mechanism is used for various statistical values in the sliding window to uniformly process within the quantization update period, avoiding frequent invocation of expensive functions.

[0125] Through the above optimization methods, the quantification algorithm can run in low-power mode on the resource-constrained microcontroller platform, and ensure real-time and accuracy, providing stable visual attention input for the control module.

[0126] Finally, it should be pointed out that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the present application, and all should be covered in the scope of the claims of the present application.

Claims

1. An adaptive vision training method based on eye features, characterized in that: The method comprises the following steps: S1: Real-time acquisition of eye images of the user by an image acquisition module; S2: Processing of the eye images by a visual intelligence analysis module, extraction of eye features, and calculation of visual concentration, wherein the eye features include eye contour features, pupil center position, eye movement trajectory, gaze duration, blink frequency, and visual line deviation; S3: Comparison of the visual concentration with a dynamic adaptive threshold to determine the user's concentration state; S4: Triggering of an intervention mechanism by a control module when the visual concentration is lower than the dynamic adaptive threshold, including issuing a prompt, pausing the training, or adjusting the training task; S5: Uploading of the visual concentration quantification value and training data to a cloud platform by a remote interaction module to generate and issue a personalized training plan.

2. The eye feature based adaptive vision training method of claim 1, wherein: The visual intelligence analysis module in S2 extracts double features of eye contour and pupil position, including: S21: Extraction of geometric feature points of the eye corner and eyelid margin by eye contour detection to construct an eye structure constraint region; S22: Generation of a pupil candidate point heat map by pupil detection; S23: Spatial constraint cropping of the pupil candidate point heat map according to the eye structure constraint region to obtain a cropped heat map; S24: Concatenation of the cropped heat map and the feature vector of the eye structure constraint region, input into a regression module, and output of the pupil center coordinates.

3. The adaptive vision training method based on eye features according to claim 2, characterized in that: S2 further includes further calculation of the eye movement trajectory, gaze duration, blink frequency, and visual line deviation based on the pupil center coordinates, and calculation of the visual concentration by the following visual concentration quantification function: F=w1·T+w2·S-w3·B-w4·D Wherein, F is the visual concentration, T is the continuous gaze duration, S is the gaze center stability score, B is the blink frequency, D is the visual line drift probability, and w1, w2, w3, w4 are preset empirical weights.

4. The adaptive vision training method based on eye features according to claim 3, characterized in that: The gaze center stability score is calculated by the following formula: where σ x and σ y are the standard deviations of the two-dimensional coordinate distribution of the gaze center within the target region, and D max is a region maximum allowed offset distance constant.

5. The eye feature based adaptive vision training method as claimed in claim 1, wherein: The dynamic adaptive threshold is calculated by the following formula: T threshold = μ + k · σ Wherein, μ is the sliding mean of the visual concentration in the recent period of time, σ is the fluctuation degree of the visual concentration quantification value, and k is an adjustment factor based on the training stage, user historical data, and rest period.

6. The eye feature based adaptive vision training method of claim 1, wherein: The intervention mechanism in S4 includes pausing the training and starting the rest countdown when the visual concentration is continuously lower than the dynamic adaptive threshold for more than a preset duration, and entering a protection mode and pausing the training task if the cumulative interruption frequency exceeds a preset threshold.

7. The eye feature based adaptive vision training method of claim 1, wherein: The personalized training plan in S5 includes parameter settings of brightness rhythm, audio synchronization mode, training duration, and intervention frequency, and the cloud platform generates the personalized training plan by curve fitting and behavior statistical analysis of the uploaded visual concentration.

8. An adaptive vision training system based on eye features, characterized by: It comprises: An image acquisition module for real-time acquisition of eye images of the user; A visual intelligence analysis module connected with the image acquisition module for processing the eye images, extracting eye features, and calculating visual concentration; A control module connected with the visual intelligence analysis module for determining the concentration state and triggering the intervention mechanism according to the comparison of the visual concentration with a dynamic adaptive threshold. The training execution module is connected with the control module and is configured to execute a visual training task according to an instruction of the control module. The remote interaction module is connected with the control module and is configured to upload the visual concentration and training data to a cloud platform and receive a personalized training scheme.

9. The eye feature based adaptive vision training system of claim 8, wherein: The visual intelligent analysis module comprises a double-feature fusion model, the double-feature fusion model comprising an eye contour detection, a pupil detection and a fusion module, wherein the eye contour detection is configured to extract an eye structure constraint region, the pupil detection is configured to generate a pupil candidate point heat map, and the fusion module is configured to crop the heat map according to the eye structure constraint region and output a pupil center coordinate.

10. The eye feature based adaptive vision training system of claim 8, wherein: The training execution module supports fixed target fixation training, target moving tracking training and audio rhythm stimulation training, and parameters of the training task are dynamically configured by the remote interaction module according to the personalized training scheme of the cloud platform.

Images