A myopia prevention method based on neural network model

Abstract

The present application relates to the technical field of artificial intelligence, and discloses a myopia prevention method based on a neural network model, including: obtaining the eye behavior data of a target person in real time; the eye behavior data includes behavior type, motion data and ambient light data; retrieving the target person's fundus image label classification result and pre-associating the matching correlation coefficient from a multi-label classification module; based on the eye behavior data and the correlation coefficient, using a preset myopia risk assessment method to assess the myopia risk of the target person, and obtaining a corresponding myopia risk value; after comparing the myopia risk value with a specified risk threshold, issuing a corresponding reminder signal; by comprehensively considering the differences in motion data generated by the target person in different states and the different requirements for ambient light data, as well as the label classification results that can reflect the target person's eye conditions, a more reasonable and accurate myopia risk warning reminder can be provided in different eye scenarios according to the actual eye conditions.

Description

Technical Field

[0001] The present application relates to the technical field of artificial intelligence, and in particular to a myopia prevention method based on a neural network model. Background Art

[0002] Myopia in adolescents is a high-frequency problem facing the current society, and its occurrence is related to a variety of factors. Specifically, if both or one of the parents is myopic, the risk of their children developing myopia will be relatively high, but genetic factors only play a certain role in the occurrence and development of myopia, and are not the only determining factor; if the eyes are used at close range for a long time, such as continuously reading, playing with mobile phones, using computers, etc., and reading and writing postures are incorrect, the risk of myopia will be greatly increased.

[0003] In addition, if children lack enough time for outdoor activities, their eyes cannot be fully exposed to natural light, which is not conducive to the normal development of the eyeball. In addition, too dark or too bright indoor lighting environments may cause damage to the eyes. When the light is insufficient, the eyes need to work harder to see objects clearly, which can easily lead to fatigue. Strong light can irritate the eyes. Lack of food rich in nutrients such as vitamins and minerals may also affect the normal development and function of the eyes. However, the most important and easily overlooked point is that most myopic and potential myopic patients rarely undergo regular vision tests. Many people believe that as long as their vision has not significantly decreased, there is no need for vision tests.

[0004] However, vision testing is not as simple as checking the eye chart. It also includes computer optometry, intraocular pressure testing, slit lamp examination, fundus examination and other aspects. These examinations can comprehensively evaluate the health of the eyes and promptly detect and prevent potential eye diseases. For example, eye diseases such as cataracts and glaucoma may have no obvious symptoms in the early stages, but vision tests can help with early detection and timely intervention, which can reduce the difficulty for patients to see objects clearly and reduce eye fatigue, thereby preventively reducing the probability of myopia. Therefore, the problem of myopia in adolescents needs to be viewed from all aspects, and preventive measures need to be integrated into all aspects of life as much as possible, and external forces should be used reasonably to achieve the formation of good eye habits. Summary of the invention

[0005] In order to provide more reasonable and accurate myopia risk warning reminders in different eye usage scenarios according to the actual eye conditions of the target person, the present application provides a myopia prevention method based on a neural network model.

[0006] The present application provides a myopia prevention method based on a neural network model, which adopts the following technical solutions:

[0007] A myopia prevention method based on a neural network model, comprising:

[0008] Acquire the eye behavior data of the target person in real time; the eye behavior data includes behavior type and action data and ambient light data;

[0009] Retrieving the correlation coefficient of the pre-correlated matching of the target person's fundus image label classification result from the multi-label classification module; the multi-label classification module includes a multi-label classification model and a conversion unit obtained by training based on a neural network model;

[0010] Based on the eye behavior data and the correlation coefficient, a preset myopia risk assessment method is used to assess the myopia risk of the target person to obtain a corresponding myopia risk value;

[0011] After comparing the myopia risk value with a specified risk threshold, a corresponding reminder signal is issued;

[0012] Wherein, the preset myopia risk assessment method determines the myopia risk value according to the behavior type, the action data and the ambient light data;

[0013] The behavior types include game status and reading status; the action data include the frequency of eye rubbing, the duration of a single eye rubbing action, the time period of a single blink, the blinking frequency value, the duration of eye closure per unit time, and the distance between the eyes and the visual object; the ambient light data includes the light color temperature, light brightness and light uniformity related to the behavior type.

[0014] By adopting the above technical scheme, the present invention comprehensively considers the differences in motion data and the different demands for ambient light data generated by the target person in two different states, namely, the gaming state or the reading state, and retrieves the target person's fundus image label classification results from the multi-label classification module. By using the fundus image label classification results that can reflect the target person's eye conditions, a myopia risk assessment method is comprehensively formulated, which can provide more reasonable and accurate myopia risk warning reminders in different eye usage scenarios according to the actual eye conditions.

[0015] Optionally, the method of evaluating the myopia risk of the target person according to a preset myopia risk evaluation method to obtain a corresponding myopia risk value includes:

[0016]

[0017] in, is the myopia risk value, A lighting health score is calculated based on the ambient light data; is the correlation coefficient, provided by the conversion unit; is the frequency of eye rubbing per unit time, is the average duration of eye rubbing action per unit time, is the average length of a single blink period, is the blink frequency per unit time, is the duration of eye closure per unit time, For the The normal blinking reference frequency value of the target person under the behavior type, the unit time is a preset value; is the distance between the eye and the visual object, For the The standard distance value between the target person's eyes and the visual object under the behavior type, is the change in distance between the eye and the visual object per unit time, is the weight coefficient.

[0018] Optionally, the process of calculating the lighting health score obtained according to the ambient light data includes:

[0019]

[0020] in, For the The color temperature value of the ambient light of the target person under the behavior type, For the The recommended color temperature value for the above behavior type is: For the The brightness value of the ambient light of the target person under the behavior type, For the Recommended brightness values ​​for the above behavior types. is the unevenness value, and are the corresponding weighting coefficients respectively.

[0021] Optionally, the method for obtaining the unevenness value includes:

[0022] Obtaining a visual area image of the target person;

[0023] A central point is randomly selected in the central area of ​​the visual area image to perform brightness detection;

[0024] Taking the central point as the center, the visual area image is divided into m partitions, and a detection point is randomly selected at the non-edge of each partition to perform brightness detection;

[0025] Connect adjacent detection points in a clockwise direction to form a detection area;

[0026]

[0027] in, and Respectively and The brightness value of the detection point corresponding to each partition, is the brightness value of the center point, is the average brightness value corresponding to m detection points.

[0028] Optionally, the training acquisition method of the multi-label classification module includes:

[0029] Acquire a fundus image set, and add a detection frame to each fundus image data in the fundus image set;

[0030] After labeling the detection frame content of the fundus image data, the data is divided into a training set and a validation set according to a preset ratio;

[0031] Constructing a neural network model based on the YOLO algorithm, integrating the principle of long short-term memory network, inputting the training set into the neural network model for training to obtain a multi-label classification model;

[0032] The multi-label classification model is modified by inputting the validation set to verify the obtained results.

[0033] Optionally, the method for training the neural network model includes:

[0034] For each training sample in the training set, the training sample is divided into S×S regions, and B detection boxes are predicted for each region to achieve A types of target classification;

[0035] Get the total loss function of the neural network model , and in the total loss function The training of the neural network model is completed when minimizing;

[0036]

[0037] in, , , Preset weighting coefficients for the corresponding items respectively, is the category loss, is the detection box loss of the target prediction, is the confidence loss for the target prediction.

[0038] Optionally, the class loss The methods for obtaining include:

[0039]

[0040] in, To predict the classification results, is the sequence number of the region, is the sequence number of the detection frame, Indicates that the predicted classification result corresponds to the object falling in the area The probability of is the target predicted value, is the predicted confidence value after the sigmoid function.

[0041] Optionally, the detection box loss The methods for obtaining include:

[0042]

[0043] Among them, x and y are the coordinate values ​​of the upper left corner of the detection frame, w and h are the width and height data of the detection frame, respectively. Indicates that the predicted classification result corresponds to the object falling in the area Middle detection box Probability within.

[0044] Optionally, the confidence loss The methods for obtaining include:

[0045]

[0046] in, and Respectively indicate whether the corresponding target has a marked value, is the weight coefficient, Indicates that the object corresponding to the predicted classification result does not fall into the area Middle detection box Probability within.

[0047] In summary, the present application includes at least one of the following beneficial technical effects:

[0048] 1. The present invention comprehensively considers the differences in motion data and the different demands for ambient light data generated by the target person in two different states, namely, the gaming state or the reading state, and retrieves the target person's fundus image label classification results from the multi-label classification module. By using the fundus image label classification results that can reflect the target person's eye conditions, a myopia risk assessment method is comprehensively formulated, which can provide more reasonable and accurate myopia risk warning reminders in different eye usage scenarios according to the actual eye conditions.

[0049] 2. In the specific calculation and acquisition process of myopia risk value, the present invention comprehensively considers the distance between the target person and the visual object and the distance change, as well as the calculation of blinking and eye rubbing related data, and can evaluate the myopia risk more reasonably and accurately. For example, when the target person unconsciously and slowly reduces the distance value between the eye and the visual object, the calculation method can amplify the impact of the distance change between the target person and the visual object on the myopia risk value to a certain extent.

[0050] 3. The present invention calculates the unevenness of the light by comparing the difference between the brightness of the center point and the surrounding brightness and the difference between the brightness of adjacent points. It has a small amount of calculation and comprehensive dimensional considerations, can adapt to high-frequency analysis and calculation, and is practical and reliable. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 is a diagram of a myopia prevention method based on a neural network model in the present invention;

[0052] Figure 2 This is an example of fundus image data with a detection frame added in the present invention. DETAILED DESCRIPTION

[0053] Embodiments of the present application are described in detail below, examples of which are shown in the accompanying drawings.

[0054] In the description of this specification, the description with reference to the terms "certain embodiments", "one embodiment", "some embodiments", "illustrative embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials, or characteristics described in conjunction with the embodiments or examples are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0055] The present application embodiment discloses a myopia prevention method based on a neural network model, such as Figure 1 As shown, including:

[0056] Acquire the eye behavior data of the target person in real time; the eye behavior data includes behavior type and action data and ambient light data;

[0057] Retrieving the correlation coefficient of the pre-correlated matching of the target person's fundus image label classification result from the multi-label classification module; the multi-label classification module includes a multi-label classification model and a conversion unit obtained by training based on a neural network model;

[0058] Based on the eye behavior data and the correlation coefficient, a preset myopia risk assessment method is used to assess the myopia risk of the target person to obtain a corresponding myopia risk value;

[0059] After comparing the myopia risk value with a specified risk threshold, a corresponding reminder signal is issued;

[0060] Wherein, the preset myopia risk assessment method determines the myopia risk value according to the behavior type, the action data and the ambient light data;

[0061] The behavior types include game status and reading status; the action data include the frequency of eye rubbing, the duration of a single eye rubbing action, the time period of a single blink, the blinking frequency value, the duration of eye closure per unit time, and the distance between the eyes and the visual object; the ambient light data includes the light color temperature, light brightness and light uniformity related to the behavior type.

[0062] When gaming, for example, using a desktop computer and playing games through a monitor, if a traditional desk lamp is placed directly next to the computer monitor, the light may shine directly into the eyeballs, affecting the visual experience and increasing eye fatigue, exacerbating the risk of myopia; when reading, natural light is needed first, and then choose lights with warmer color temperatures, such as incandescent lamps or LED eye protection lamps, which emit soft and uniform light and cause less irritation to the eyes.

[0063] Therefore, when gaming, try to use lights that may produce reflections and maintain a brightness balance between the screen and the surrounding environment, such as screen hanging lights, which can provide sufficient lighting while avoiding direct light. The color temperature can be selected to be a neutral color temperature of 4000K-5000K, which is suitable for users who use computers for a long time, including gamers; when reading, in the absence of natural light, give priority to warm white light or warm yellow light with a color temperature between 3500K and 4000K. The color temperature light in this range is soft and not dazzling, which can simulate the natural light environment and help reduce eye fatigue.

[0064] In this way, taking into account the differences in motion data and the different demands for ambient light data generated by the target person in the two different states of gaming or reading, and retrieving the target person's fundus image label classification results from the multi-label classification module, the fundus image label classification results can reflect the target person's eye conditions. Through the above content, a myopia risk assessment method is comprehensively formulated, which can provide more reasonable and accurate myopia risk warning reminders in different eye usage scenarios according to the actual eye conditions.

[0065] Optionally, the method of evaluating the myopia risk of the target person according to a preset myopia risk evaluation method to obtain a corresponding myopia risk value includes:

[0066]

[0067] in, is the myopia risk value, A lighting health score is calculated based on the ambient light data; is the correlation coefficient, provided by the conversion unit; is the frequency of eye rubbing per unit time, The average duration of eye rubbing in a unit of time. The more times you rub your eyes and the longer the duration, the greater the probability of eye fatigue. is the average length of a single blink cycle. A single blink cycle is the process of eyelid closing 10% - closing the eyes - eyelid opening 90%. is the blink frequency per unit time, is the duration of eye closure per unit time, For the The normal blinking reference frequency value of the target person under the behavior type, the unit time is a preset value, which is set to 3s in this embodiment; is the distance between the eye and the visual object, For the The standard distance value between the target person's eyes and the visual object under the behavior type, is the change in distance between the eye and the visual object per unit time, is the weight coefficient.

[0068] The correlation coefficient provided by the conversion unit , pre-configured according to the fundus image label classification results; the fundus image label classification results may include: retinal atrophy, retinal hole, optic nerve abnormality, macular hemorrhage, retinal vein occlusion, retinal artery occlusion and lens opacity, and the correlation coefficient can be adjusted according to the severity of the above classifications Pre-configuration.

[0069] In this embodiment, multiple cameras can be set up in the target person's seating area to capture their face and sides, and the obtained video data can be uploaded to the cloud server. The facial features of the target person can be recognized through a large model to obtain motion data. The acquisition of this motion data belongs to the existing technology and will not be elaborated here.

[0070] in, The distance between the target person and the visual object and the distance change are combined. When the target person unconsciously and slowly reduces the distance between the eye and the visual object, this calculation method can amplify the impact of the change in the distance between the target person and the visual object on the myopia risk value to a certain extent; when hour, Can be set smaller, when hour, Can be set larger.

[0071] Optionally, the process of calculating the lighting health score obtained according to the ambient light data includes:

[0072]

[0073] in, For the The color temperature value of the ambient light of the target person under the behavior type, For the The recommended color temperature value for the above behavior type is: For the The brightness value of the ambient light of the target person under the behavior type, For the Recommended brightness values ​​for the above behavior types. is the unevenness value, and are the corresponding weighting coefficients respectively.

[0074] In this embodiment, the recommended color temperature value in the gaming state can be selected as 4500K, and the recommended color temperature value in the reading state can be selected as 3750K; when the target person reads or writes under light with uneven brightness, the line of sight will continue to move at a high frequency within the lighting range. If the brightness within the illumination range is uneven, especially if there is a large gap between the brightness of the central area and the surrounding area, the pupil will also frequently contract and expand during the movement of the eyes. Over time, this frequent pupil adjustment will cause visual fatigue, increase the adjustment load of the eyes, and keep the eyes in a state of tension for a long time. This continuous state of tension may cause eye muscle fatigue, which in turn aggravates the development of myopia. Therefore, how to accurately evaluate the unevenness is a problem that needs to be solved.

[0075] Optionally, in this embodiment, the method for obtaining the non-uniformity value includes:

[0076] Obtain a visual area image of the target person; considering that the desktop background may be dark, the visual area image may be a data content excluding the reading and writing areas;

[0077] A central point is randomly selected in the central area of ​​the visual area image to perform brightness detection;

[0078] Taking the central point as the center, the visual area image is divided into m partitions, and a detection point is randomly selected at the non-edge of each partition to perform brightness detection;

[0079] Connect adjacent detection points in a clockwise direction to form a detection area;

[0080]

[0081] in, and Respectively and The brightness value of the detection point corresponding to each partition, is the brightness value of the center point, is the average brightness value corresponding to m detection points; Represents permutations and combinations. The above formula compares the difference between the brightness of the center point and the surrounding brightness, as well as the difference between the brightness of adjacent points. The calculation amount is small but the dimensions are comprehensive. It can adapt to high-frequency analysis and calculation, and is practical and reliable. It is worth noting that in order not to affect the normal operation of the above formula, the above formula is calculated using dimensionless operation.

[0082] In this embodiment of the present invention, the training acquisition method of the multi-label classification module includes:

[0083] Acquire a fundus image set, and add a detection frame to each fundus image data in the fundus image set;

[0084] After labeling the detection frame content of the fundus image data, the data is divided into a training set and a validation set according to a preset ratio;

[0085] Constructing a neural network model based on the YOLO algorithm, integrating the principle of long short-term memory network, inputting the training set into the neural network model for training to obtain a multi-label classification model;

[0086] The multi-label classification model is modified by inputting the validation set to verify the obtained results.

[0087] In this embodiment, if Figure 2As shown in the figure, the detection frames of the fundus image data can be multiple, and the low feature area of ​​the eye can be divided by the detection frame, and then the feature points with abnormalities in the detection frame can be further divided into detection frames; the neural network model mainly includes the Backbone part, the Neck part and the Head part. The Backbone part is responsible for extracting useful features from the input image. It is usually a convolutional neural network (CNN) trained in large-scale image classification tasks, such as ImageNet. With the iteration of versions, the Backbone network structure is also continuously optimized to improve the feature extraction capability and computational efficiency; the Neck part, as the intermediate component connecting the Backbone and the Head, aggregates and refines the features extracted by the backbone network, and usually focuses on strengthening the spatial and semantic information of different scales. The Head part is the last component of the object detector, which is responsible for making predictions based on the features provided by the Backbone and Neck parts. It is usually composed of one or more sub-networks for specific tasks, and performs tasks such as classification and positioning.

[0088] Specifically, the method for training the neural network model includes:

[0089] For each training sample in the training set, the training sample is divided into S×S regions, and B detection boxes are predicted for each region to achieve A types of target classification;

[0090] Get the total loss function of the neural network model , and in the total loss function The training of the neural network model is completed when minimizing;

[0091]

[0092] in, , , Preset weighting coefficients for the corresponding items respectively, is the category loss, is the detection box loss of the target prediction, is the confidence loss of the target prediction, when the total loss function When it reaches the minimum, the neural network model training is completed.

[0093] In addition, considering the influence of the accuracy of the neural network model on the calculation accuracy of the myopia risk value, the correlation coefficient The calculation method can be limited to ,in, is the real table lookup correlation coefficient.

[0094] Optionally, the class loss The methods for obtaining include:

[0095]

[0096] in, To predict the classification results, is the sequence number of the region, is the sequence number of the detection frame, Indicates that the predicted classification result corresponds to the object falling in the area The probability of falling in can generally be configured as 1 or 0, 1 means falling in, and 0 means not falling in. is the target predicted value, is the predicted confidence value after the sigmoid function.

[0097] Optionally, the detection box loss The methods for obtaining include:

[0098]

[0099] Among them, x and y are the coordinate values ​​of the upper left corner of the detection frame, w and h are the width and height data of the detection frame, respectively. Indicates that the predicted classification result corresponds to the object falling in the area Middle detection box The probability of falling within can generally be configured as 1 or 0, 1 means falling within, and 0 means not falling within.

[0100] Optionally, the confidence loss The methods for obtaining include:

[0101]

[0102] in, and Respectively indicate whether the corresponding target has a marked value, is the weight coefficient, Indicates that the object corresponding to the predicted classification result does not fall into the area Middle detection box Probability within.

[0103] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limitations on the present application. Ordinary technicians in the field can change, modify, replace and modify the above embodiments within the scope of the present application.

Claims

1. A myopia prevention method based on a neural network model, characterized in that: include: Obtain the target person’s eye behavior data in real time; The eye behavior data includes behavior type and motion data and ambient light data; Retrieving the correlation coefficient of the pre-correlated matching of the target person's fundus image label classification result from the multi-label classification module; the multi-label classification module includes a multi-label classification model and a conversion unit obtained by training based on a neural network model; Based on the eye behavior data and the correlation coefficient, a preset myopia risk assessment method is used to assess the myopia risk of the target person to obtain a corresponding myopia risk value; After comparing the myopia risk value with a specified risk threshold, a corresponding reminder signal is issued; Wherein, the preset myopia risk assessment method determines the myopia risk value according to the behavior type, the action data and the ambient light data; The behavior types include game status and reading status; the action data include eye rubbing action frequency, duration of a single eye rubbing action, single blinking time period, blinking frequency value, eye closing duration per unit time, and distance between the eye and the visual object; the ambient light data includes light color temperature, light brightness and light uniformity related to the behavior type; The method of using a preset myopia risk assessment method to assess the myopia risk of the target person and obtaining a corresponding myopia risk value includes: in, is the myopia risk value, A lighting health score is calculated based on the ambient light data; is the correlation coefficient, provided by the conversion unit; is the frequency of eye rubbing per unit time, is the average duration of eye rubbing action per unit time, is the average length of a single blink period, is the blink frequency per unit time, is the duration of eye closure per unit time, For the The normal blinking reference frequency value of the target person under the behavior type, the unit time is a preset value; is the distance between the eye and the object being viewed, For the The standard distance value between the target person's eyes and the visual object under the behavior type, is the change in distance between the eye and the visual object per unit time, is the weight coefficient.

2. The myopia prevention method based on the neural network model according to claim 1, characterized in that: The process of calculating and obtaining the lighting health score according to the ambient light data includes: in, For the The color temperature value of the ambient light of the target person under the behavior type, For the The recommended color temperature value for the above behavior type is: For the The brightness value of the ambient light of the target person under the behavior type, For the Recommended brightness values ​​for the above behavior types. is the unevenness value, and are the corresponding weighting coefficients respectively.

3. The myopia prevention method based on the neural network model according to claim 2, characterized in that: The method for obtaining the unevenness value comprises: Obtaining a visual area image of the target person; A central point is randomly selected in the central area of ​​the visual area image to perform brightness detection; Taking the central point as the center, the visual area image is divided into m partitions, and a detection point is randomly selected at the non-edge of each partition to perform brightness detection; Connect adjacent detection points in a clockwise direction to form a detection area; in, and Respectively and The brightness value of the detection point corresponding to each partition, is the brightness value of the center point, is the average brightness value corresponding to m detection points.

4. The myopia prevention method based on the neural network model according to claim 1, characterized in that: The training acquisition method of the multi-label classification module includes: Acquire a fundus image set, and add a detection frame to each fundus image data in the fundus image set; After labeling the detection frame content of the fundus image data, the data is divided into a training set and a validation set according to a preset ratio; Constructing a neural network model based on the YOLO algorithm, integrating the long short-term memory network principle, inputting the training set into the neural network model for training to obtain a multi-label classification model; The multi-label classification model is modified by inputting the validation set to verify the results.

5. The myopia prevention method based on the neural network model according to claim 4, characterized in that: The method for training the neural network model includes: For each training sample in the training set, the training sample is divided into S×S regions, and B detection boxes are predicted for each region to achieve A types of target classification; Get the total loss function of the neural network model , and in the total loss function The training of the neural network model is completed when minimizing; in, , , Preset weighting coefficients for the corresponding items respectively, is the category loss, is the detection box loss of the target prediction, is the confidence loss for the target prediction.

6. The myopia prevention method based on the neural network model according to claim 5, characterized in that: The category loss The methods for obtaining include: in, To predict the classification results, is the sequence number of the region, is the sequence number of the detection frame, Indicates that the predicted classification result corresponds to the object falling in the area The probability of is the target predicted value, is the predicted confidence value after the sigmoid function.

7. The myopia prevention method based on the neural network model according to claim 6, characterized in that: The detection box loss The methods for obtaining include: Among them, x and y are the coordinate values ​​of the upper left corner of the detection frame, w and h are the width and height data of the detection frame, respectively. Indicates that the predicted classification result corresponds to the object falling in the area Middle detection box Probability within.

8. The myopia prevention method based on the neural network model according to claim 7, characterized in that: The confidence loss The methods for obtaining include: in, and Respectively indicate whether the corresponding target has a marked value, is the weight coefficient, Indicates that the object corresponding to the predicted classification result does not fall into the area Middle detection box Probability within.

Images