Dynamic illumination prediction method and device based on refractive state and program product

Abstract

The invention relates to the field of intelligent medical treatment, in particular to a dynamic illumination prediction method and device based on a refraction state and a program product. Comprising the steps of obtaining eye refraction data of a to-be-tested person; carrying out clustering and feature fusion on the refractive data to obtain fusion features; outputting the fusion features to a trained prediction model for prediction to obtain a brightness prediction value and a color temperature prediction value; and the prediction model carries out weight parameter initialization by compressing an abnormal value of refractive data and compensating distribution deviation of refractive clustering. According to the method, personalized expression of different myopia degrees on illumination perception differences can be achieved, brightness and color temperature parameters are dynamically predicted, actual lighting equipment is driven, dynamic balance of visual comfort and environmental adaptability is achieved, and the method has good actual application value.

Description

Technical Field

[0001] This application relates to the field of intelligent healthcare, specifically to a dynamic illumination prediction method, device, program product, and computer-readable storage medium based on refractive state. Background Technology

[0002] As modern people become increasingly reliant on electronic devices and artificial lighting in their daily lives, visual health problems, especially the number of people with refractive errors, are growing rapidly. Visual fatigue and discomfort caused by lighting have become common issues. In particular, highly myopic users are significantly more sensitive to changes in lighting than the general population. Excessive brightness or inappropriate color temperature can easily cause glare, dry eyes, or even visual fatigue. However, most existing intelligent lighting systems adjust lighting based on environmental sensors or fixed scene settings, failing to fully consider the differences in refractive states of individual users. Summary of the Invention

[0003] To address the above problems, this invention provides a dynamic illumination prediction method based on refractive state, specifically including: Obtain the refractive data of the subject's eye; The refractive data are clustered and feature fusion is performed to obtain fused features; The fused features are input into the trained prediction model to obtain the brightness prediction value and the color temperature prediction value; The prediction model initializes its weight parameters by compressing outliers in the refractive data and compensating for distribution biases in refractive clustering.

[0004] Optionally, outliers in the compressed refractive data are compressed using a skewness-aware scaling matrix; the distribution bias of the compensated refractive clustering is compensated by calculating a bias correction term using refractive clustering information.

[0005] Optionally, the prediction model learns the characteristic relationship between refractive sensitivity and brightness through an adaptive activation function of refractive perception; the adaptive activation function of refractive perception is obtained by dynamically adjusting the steepness and threshold position of the activation curve by the average spherical power.

[0006] Optionally, the clustering is based on the average spherical power in the refractive data to obtain grouped sample clusters, and a projection matrix is ​​constructed to remove redundant features from the grouped sample clusters to obtain feature clusters; the feature fusion is based on the density of the feature clusters to calculate the fusion weight, and feature clusters with different fusion weights are fused to obtain fused features; The projection matrix is ​​constructed by performing singular value decomposition on the feature vectors of samples within the cluster to obtain a left singular matrix and a singular value matrix, and then calculating the projection matrix of the current cluster based on the left singular matrix and the singular value matrix.

[0007] The purpose of this invention is to provide a dynamic lighting adjustment method based on refractive state adaptation, comprising a light source controller and a lighting source, including: The luminance prediction value and color temperature prediction value are obtained based on the above dynamic illumination prediction method based on refractive state. The predicted brightness and color temperature values ​​are input to the light source controller; The brightness and color temperature of the lighting source are dynamically adjusted by the light source controller.

[0008] The purpose of this invention is to provide a computer program product that includes a computer program or instructions, wherein the computer program or instructions are executed by a processor to implement the above-described dynamic illumination prediction method based on refractive state, or to implement the above-described dynamic illumination adjustment method based on refractive state adaptation.

[0009] The purpose of this invention is to provide a computer device comprising a memory, a processor, and a computer program or instructions stored in the memory, wherein the computer program or instructions are executed by the processor to implement the above-described dynamic illumination prediction method based on refractive state, or to implement the above-described dynamic illumination adjustment method based on refractive state adaptation.

[0010] The purpose of this invention is to provide a computer-readable storage medium having a computer program or instructions stored thereon, wherein the computer program or instructions are executed by a processor to implement the above-described dynamic illumination prediction method based on refractive state, or to implement the above-described dynamic illumination adjustment method based on refractive state adaptation.

[0011] Advantages of this invention: 1. Existing feature extraction methods often employ PCA or static dimensionality reduction strategies, neglecting the refractive modulation dependencies between features, resulting in the loss of key interactive features such as "myopic users are more sensitive to brightness." Therefore, this application constructs a clustering projection matrix and fusion weights based on refractive state grouping, reducing dimensionality while retaining key discriminative features, and capturing the nonlinear interactions related to refractive states.

[0012] 2. Regarding data processing methods, traditional normalization methods assume that features are processed at the same scale, failing to consider the nonlinear photosensitive modulation of refractive states, resulting in weak modeling ability for individual visual differences. Therefore, this application employs a refractive state influence function to dynamically calculate the normalization scaling factor based on the user's average spherical power, thereby adjusting the feature scale and addressing the personalized expression of differences in light perception caused by myopia degree.

[0013] 3. For prediction models, conventional activation functions, initialization methods, and loss functions are unstable for extreme refractive samples such as high myopia, have difficulty converging during training, and exhibit poor generalization ability, failing to accurately serve heavy users. Therefore, this application adopts a prediction model structure that includes an adaptive activation function, an adaptive initialization strategy, a refractive weighted loss function, and a modulation path, embedding refractive state modulation capabilities throughout the entire process from input to output, thereby improving the prediction accuracy and generalization ability for high myopia users.

[0014] 4. Traditional lighting adjustment relies primarily on preset scenes, lacking real-time perception and feedback adjustment of users' visual physiological parameters, and cannot optimize lighting conditions in real time based on refractive status. The model in this application adopts an end-to-end training method, which enables end-to-end closed-loop lighting adjustment after training. It collects user refractive and environmental information in real time, dynamically outputs brightness and color temperature parameters, drives actual lighting equipment, and achieves a dynamic balance between visual comfort and environmental adaptability. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A schematic diagram of the dynamic illumination prediction method based on refractive state provided in an embodiment of the present invention; Figure 2 A schematic diagram of a dynamic illumination prediction system based on refractive state provided in an embodiment of the present invention; Figure 3 A schematic diagram of a computer device provided in an embodiment of the present invention; Figure 4 A comparison diagram of the normalization effect under different refractive states provided in an embodiment of the present invention; Figure 5 The brightness prediction error distribution for different refractive states provided in the embodiments of the present invention; Figure 6 The color temperature prediction error distribution for different refractive states provided in the embodiments of the present invention; Figure 7 The effect of refractive state on light sensitivity is provided in the embodiments of the present invention; Figure 8 The influence of refractive state on preferred brightness provided in the embodiments of the present invention; Figure 9 The influence of refractive state on the response time to light changes under different technologies provided in the embodiments of the present invention. Detailed Implementation

[0017] To enable those skilled in the art to better understand the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0018] In some of the processes described in the specification, claims, and accompanying drawings of this invention, multiple operations appearing in a specific order are included. However, it should be clearly understood that these operations may not be executed in the order they appear herein, or may be executed in parallel. The operation numbers, such as S101, S102, etc., are merely used to distinguish different operations and do not represent any execution order. Furthermore, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should be noted that the descriptions such as "first," "second," etc., in this document are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to different types.

[0019] Figure 1 A schematic diagram of the dynamic illumination prediction method based on refractive state provided in this embodiment of the invention specifically includes: S1: Obtain the refractive data of the subject's eye; In one embodiment, the ocular refractive data is obtained by calculating the ocular refractive data through a refractive state influence function, which is obtained by calculating the dynamic relationship between the initial ocular refractive data and environmental data.

[0020] In one embodiment, the process of calculating ocular refractive data is as follows: Obtain initial refractive data and environmental data of the subject's eyes; Preprocessing is performed on the initial ocular refractive data and environmental data to obtain preprocessed data; The preprocessed data is used to calculate ocular refractive data through the refractive state influence function; wherein the refractive state influence function is obtained by calculating the dynamic relationship between the initial ocular refractive data and environmental data.

[0021] In one embodiment, the dynamic relationship of the refractive state influence function is obtained based on the nonlinear response of an individual's refractive state to light sensitivity; The formula for calculating the effect function of refractive state is:

[0022] in, Adjusting hyperparameters for the first refractive state controls the magnitude of the scaling factor. To adjust the hyperparameters for the second light state, the curvature of the scaling factor is controlled. This represents the average spherical power, which is the calculation dimension for a single sample.

[0023] In one embodiment, the initial ocular refractive data includes one or more of the following: binocular spherical power, cylindrical power, axis, and pupil diameter.

[0024] Optionally, initial ocular refractive data may also include axial length, corneal curvature, and tear film quality index.

[0025] In one embodiment, the environmental data includes one or more of the following: ambient brightness, ambient color temperature, and sensor X / Y axis readings.

[0026] In one specific embodiment, user refractive state parameters and ambient light characteristics are collected through multi-source heterogeneous devices: Use an optometry device to obtain the user's core refractive parameters, such as binocular spherical power, cylindrical power, and axis. Pupil diameter was recorded using a pupillometer. Combined with an ambient light sensor, it captures ambient brightness, ambient color temperature, and sensor X / Y axis readings in real time. Biometric devices were used to obtain axial length, corneal curvature, and tear film quality index.

[0027] In one embodiment, the preprocessing includes one or more of the following: normalization, data cleaning, and data anonymization; Optionally, the initial ocular refractive data and environmental data are normalized to obtain normalized data; the normalized data are used to calculate ocular refractive data through a refractive state influence function.

[0028] In one embodiment, the dynamic relationship of the refractive state influence function is obtained based on the nonlinear response of an individual's refractive state to light sensitivity.

[0029] In one embodiment, the process of generating the refractive state influence function is as follows: Obtain the initial ocular refractive dataset, environmental dataset, and labels, including labels for individuals with different refractive states; The refractive state influence function is generated by dynamically calculating the relationship between the initial ocular refractive data set and the environmental data set.

[0030] Specifically, individuals with different refractive states include those with mild myopia, those with severe myopia, those with moderate myopia, and those with normal refractive status.

[0031] In one specific embodiment, refractive features such as spherical power and cylindrical power have problems with inconsistent units and mismatched numerical ranges with environmental features such as ambient brightness. Highly myopic users are more sensitive to changes in ambient light. Global normalization methods such as Min-Max normalization cannot effectively handle the multi-scale characteristics and dynamic range differences between refractive state data and environmental data. This can easily lead to the normalization process ignoring the nonlinear influence of refractive state on light sensitivity, making it difficult to adapt to individual differences.

[0032] This invention calculates the mean and standard deviation of each feature dimension based on the entire training dataset, performs basic normalization on the original feature values, then uses a refractive state influence function, which takes the average spherical power as input, to calculate a scaling factor. Finally, the result after basic normalization is multiplied by this scaling factor to obtain the normalized feature values ​​adapted to the refractive state. This allows the normalization process to dynamically adjust the scaling ratio according to the user's refractive state, expressed as:

[0033] In the formula, Indicates the first The normalized value of the dimensional feature is the calculated dimension of a single sample, representing the scaled data after adaptation to the refractive state; This represents the average spherical power, expressed in diopters. It is a calculation dimension for a single sample, and the calculation method is as follows: ,in This refers to the spherical lens power of the right eye. The left eye spherical power represents the effect of the user's refractive state on light sensitivity. For feature dimension indexing, identify the location of the feature in the sample, such as, Corresponding age characteristics; Indicates the first The original values ​​of the dimensional features reflect the original attributes of the sample, such as spherical diopter and ambient brightness. Indicates the first of all samples The mean of the dimensional features is calculated based on the entire training dataset; Indicates the first of all samples The standard deviation of the 3D features is calculated based on the entire training dataset; The function representing the influence of refractive state is calculated as follows: ; It is the hyperbolic tangent function; Adjusting hyperparameters for the first refractive state controls the magnitude of the scaling factor, such as... ; Adjust the hyperparameters for the second light state to control the curvature of the scaling factor, such as... ; This represents a small offset used to prevent... To prevent division by zero errors or numerical instability when the value is close to zero, and to ensure robustness of normalization calculations, the default value is [value missing]. .

[0034] In one specific embodiment, an example of the refractive data and environmental data attributes and original values ​​of a sample is as follows: User age (35 years), gender (1 for male, no unit, One-hot encoding format), right eye spherical power (-4.5 diopters), right eye cylindrical power (-1.2 diopters), right eye axis (90 diopters), left eye spherical power (-4.0 diopters), left eye cylindrical power (-1.0 diopters), left eye axis (85 diopters), pupil diameter (3.8 millimeters), ambient luminance (300 lux), ambient color temperature (4000 Kelvin), axial length (24.2 millimeters), corneal curvature (43.5 diopters), tear film quality index (85, percentage value).

[0035] Furthermore, based on the attribute values ​​of this sample, the average spherical power is calculated according to the formula. ,in This refers to the spherical lens power of the right eye. Substitute the left eye spherical lens power into the right eye spherical lens power. Refractive error, left eye spherical power is Refractive power, then (Diopter), i.e., the average spherical power is Refractive power.

[0036] In one specific embodiment, conventional normalization ignores refractive state modulation, and the refractive state influence function dynamically adapts to individual differences. It is designed based on the nonlinear characteristics of refractive state sensitivity to illumination. The term provides a smooth S-shaped curve through the hyperbolic tangent function. Control the scaling factor. Controlling curvature, when Time indicates high myopia, Magnifying feature values ​​improves low-light region discrimination. hour, Reduce interference.

[0037] In addition, output value Characterizes the scaling factor and adjusts it dynamically. The scaling ratio is adjusted to magnify or compress the range of characteristic values ​​to accommodate the nonlinear response of users with different refractive states to light sensitivity. For example, users with high myopia require greater scaling to enhance low-light discrimination.

[0038] In one specific embodiment, to verify the feature discrimination improvement effect of the proposed adaptive normalization method on user groups with different refractive states, the feature discrimination index, a key indicator for measuring the quality of feature spatial distribution, is used. This index directly affects the learning ability and prediction accuracy of the subsequent model. Experiments compared the performance of the conventional normalization method (global normalization) and the adaptive normalization method of this invention on three user groups: high myopia, moderate myopia, and mild myopia. A higher feature discrimination index indicates that the feature distribution in space is more conducive to the model distinguishing different patterns, such as... Figure 4 As shown, in the high myopia user group, the feature discrimination of the method of the present invention is significantly higher than that of the conventional method, and the data distribution is more concentrated (as reflected by the width of the box plot and violin plot). In the moderate myopia group, the method of the present invention also maintains a significant advantage, with a smaller range of data fluctuation. In the mild myopia group, the gap between the two methods narrows, but the method of the present invention still maintains a slight advantage. The experimental results show that the present invention, by adopting the refractive state influence function, can dynamically adjust the feature scaling ratio according to the user's refractive state, improve the feature representation of high myopia users, and solve the problem that conventional normalization ignores the nonlinear influence of refractive state on light sensitivity. The vertical axis "feature discrimination index" in the figure is a dimensionless index with no unit, ranging from 0 to 1. The closer it is to 1, the better the discrimination.

[0039] S2: The refractive data is clustered and feature fusion is performed to obtain fused features; In one embodiment, the clustering is based on the average spherical power in the refractive data to obtain grouped sample clusters, and a projection matrix is ​​constructed to remove redundant features from the grouped sample clusters to obtain feature clusters; the feature fusion is based on the density of the feature clusters to calculate the fusion weight, and feature clusters with different fusion weights are fused to obtain fused features. In one embodiment, the projection matrix is ​​constructed by performing singular value decomposition on the feature vectors of samples within the cluster to obtain a left singular matrix and a singular value matrix, and then calculating the projection matrix of the current cluster based on the left singular matrix and the singular value matrix.

[0040] Optionally, the fusion weight calculation involves using the average squared distance from the samples within a cluster to the centroid of the cluster as the inverse of the cluster density, and calculating the fusion weight of each cluster based on the inverse of the density of all clusters.

[0041] S3: The fused features are input into the trained prediction model to obtain the predicted brightness and color temperature values; the prediction model is initialized with weight parameters by compressing outliers in the refractive data and compensating for the distribution bias of refractive clusters.

[0042] In one embodiment, outliers in the compressed refractive data are compressed using a skewness-aware scaling matrix; the distribution bias of the compensated refractive clusters is compensated by calculating a bias correction term using refractive cluster information.

[0043] In one embodiment, the skewness-aware scaling matrix calculates the skewness values ​​of each dimension of the fused feature vector, and calculates the scaling coefficients corresponding to each dimension based on the skewness values ​​and scaling adjustment hyperparameters. Then, the scaling coefficients are used to construct the skewness-aware scaling matrix.

[0044] In one embodiment, the calculation process of the bias correction term is as follows: Calculate the global mean vector of the fused feature vectors of all training samples; Calculate the difference between the centroid of each cluster and the global mean, and then sum the differences using the fusion weights of the clusters. Multiplying the weighted summation result by the bias learning rate hyperparameter yields the bias correction term.

[0045] In one embodiment, the weight parameter initialization process is as follows: calculate the mean of the fused feature vectors of all samples, multiply the mean vector by the skewness-aware scaling matrix, and add it to the bias correction term to obtain the initial weight matrix of the first layer of the neural network.

[0046] In one embodiment, the weight initialization only compresses outliers in the refractive data and compensates for distribution biases in refractive clustering in the first layer of the prediction model, while random weight initialization is performed on the other layers of the prediction model.

[0047] In one embodiment, the prediction model learns the characteristic relationship between refractive sensitivity and brightness through an adaptive activation function of refractive perception; the adaptive activation function of refractive perception is obtained by dynamically adjusting the steepness and threshold position of the activation curve by the mean spherical power.

[0048] In one embodiment, the prediction model includes an input layer, a feature extraction module, and a prediction module. The fused features are sequentially passed through the input layer, the feature extraction module, and the prediction module. The prediction module includes a dual-branch structure. The first branch performs a linear mapping of the deep feature vector through a weight vector and a global bias term to obtain a basic prediction value. The second branch generates a refractive dynamic modulation factor for brightness prediction and a refractive dynamic modulation factor for color temperature prediction based on the deep feature vector and the average spherical power. The brightness prediction value and the color temperature prediction value are obtained based on the basic prediction value and the refractive dynamic modulation factor.

[0049] In one embodiment, the method further includes: acquiring behavioral data; inputting the behavioral data and fusion features into a trained first prediction model to predict brightness and color temperature values.

[0050] In one embodiment, the training process of the first prediction model is as follows: Obtain the subject's ocular refractive error dataset, environmental dataset, behavioral dataset, and labels; The ocular refractive dataset, environmental dataset, behavioral dataset, and labels are input into the prediction model for training. The predicted values ​​are compared with the true values, the loss is calculated, and the prediction model is iterated until the loss function remains unchanged, thus obtaining the trained prediction model.

[0051] In one embodiment, behavioral data includes user-defined preferred brightness and color temperature levels, as well as historical average brightness and color temperature data automatically calculated by the system.

[0052] In one embodiment, the prediction model has the same network architecture as the first prediction model, including but not limited to the same parameter initialization and loss function processing method, the difference being the different input data, thus forming a prediction model that is the same as the first prediction model.

[0053] In one specific embodiment, user refractive state parameters, ambient light characteristics, and behavioral context data are collected using multi-source heterogeneous devices. Specific collection methods include: The system uses an optometer to obtain the user's core refractive parameters, such as spherical power, cylindrical power, and axis; a pupillometer to record the pupil diameter; an ambient light sensor to capture ambient brightness, ambient color temperature, and sensor X / Y axis readings in real time; device logs to extract the user's age, gender, usage time, activity type, screen brightness, and screen color temperature; and biometric devices to obtain axial length, corneal curvature, and tear film quality index.

[0054] At the same time, it records the user's preferred brightness and color temperature levels, as well as the historical average brightness and color temperature automatically calculated by the system.

[0055] Furthermore, the data is labeled, focusing on the target parameters of the dynamic lighting mode. Professional vision engineers label the target brightness value and target color temperature value of each sample based on user subjective comfort feedback and environmental adaptation requirements, forming supervised learning label pairs. The label categories strictly distinguish between two continuous numerical labels: brightness and color temperature.

[0056] In one specific embodiment, the attributes and raw values ​​of a sample's behavioral data are as follows: Current time (14.5, time-encoded format, indicating time is 14:30), usage duration (120, in minutes), activity type (2, reading, no unit, one-hot encoded format), screen brightness (200, in candela per square meter), screen color temperature (5000, in Kelvin), ambient light sensor X-axis reading (0.8, normalized value, no unit), ambient light sensor Y-axis reading (0.6, normalized value, no unit), user preferred brightness level (70, percentage value, in %), user preferred color temperature level (5500, in Kelvin), historical average brightness (180, in candela per square meter), historical average color temperature (4500, in Kelvin), axial length (24.2, in millimeters), corneal curvature (43.5, in diopters), and tear film quality index (85, percentage value, in %).

[0057] The data entry is labeled as dynamic lighting mode parameters, specifically: target brightness L_"target" = 250 (candela per square meter), target color temperature T_"target" = 4800 (Kelvin).

[0058] In one specific embodiment, dynamic feature fusion is performed based on refractive clustering: There is a strong interaction effect between the pupil diameter of myopic users and the ambient brightness. Conventional principal component analysis methods are prone to losing such key discriminative information when processing normalized data, and cannot eliminate the redundancy of highly correlated features such as spherical power and cylindrical power.

[0059] This invention achieves dynamic grouping and weighted fusion through a feature fusion layer guided by refractive state. First, samples are clustered based on average spherical power. Then, a projection matrix is ​​constructed for each cluster to eliminate feature redundancy. Finally, features are fused by combining cluster density weights. This reduces feature dimensionality while retaining key discriminative information, enhancing the ability to capture nonlinear interaction effects related to refractive state. The specific steps are as follows: 1) Refractive state clustering and cluster assignment; Using the K-means algorithm, clustering is performed based on the average spherical power of all training samples, which are divided into 3 clusters by default. The Euclidean distance from each sample to the centroid of each cluster is calculated, and the sample is assigned to the nearest cluster. The refractive state group to which each sample belongs is determined, as follows:

[0060] In the formula, The number of clusters is represented by the average spherical power calculated using the K-means algorithm. Clustering is determined, with a default value of ; Indicates the first The cluster index of each sample identifies the cluster to which the sample belongs; The centroid of the c-th cluster, also known as the cluster center, is calculated through iterative optimization using the K-means algorithm. This includes initializing random centers, assigning samples to the nearest cluster, updating the centers to the mean of the average spherical diopter within the cluster, until convergence. For clustered indexes; For the first The average spherical power of each sample; It is an L2 norm; This represents the cluster index that minimizes the Euclidean distance. .

[0061] The setting of C=3 is an empirical default value, corresponding to common refractive state groups, namely mild, moderate and high myopia. The K-means algorithm is applied to the average spherical power of all training samples to automatically determine the cluster center.

[0062] 2) Cluster projection matrix and weight calculation: For each cluster, singular value decomposition is performed using the normalized eigenvectors of the samples within the cluster. The resulting left singular vector matrix and singular value matrix are used to calculate the projection matrix of the cluster. Simultaneously, the mean squared distance from the samples within the cluster to the cluster centroid is calculated as the inverse of the cluster density. Based on the inverse densities of all clusters, the fusion weights for each cluster are calculated, resulting in the cluster projection matrix used for dimensionality reduction and redundancy elimination, as well as the fusion weights reflecting the importance of the clusters, expressed as:

[0063]

[0064] In the formula, This represents the projection matrix of the c-th cluster, used to compress the feature dimension while preserving the main variations within the cluster; This represents the left singular vector matrix obtained by singular value decomposition of the normalized feature matrix of the sample in the c-th cluster; Let represent the singular value matrix of the c-th cluster. It is a diagonal matrix whose diagonal elements are singular values. Represents the singular value matrix Take each diagonal element The power of the power is the reciprocal square root; This represents the fusion weight of the c-th cluster; the larger the value, the greater the contribution of that cluster to the fusion result. This represents the weight adjustment factor, which controls the sensitivity of the density to the weights, such as... ; This represents the reciprocal of the sample density of the c-th cluster, which is the average squared distance from the sample to the cluster center. It is calculated as follows: ; This represents the reciprocal of the sample density of the k-th cluster; k is the cluster index that distinguishes it from c. This represents the number of samples in the c-th cluster; Indicates the first The normalized feature vector of each sample, where all feature values ​​are normalized values; For sample index; This represents the natural exponential function.

[0065] The term is used to calculate the first... The projection matrix of the cluster is used to compress the feature dimension, achieving dimensionality reduction while preserving the main variations within the cluster. It is the left singular vector of the singular value decomposition. Scaling singular values ​​balances dimensional contributions.

[0066] 3) Dynamic feature fusion: For each sample, its normalized feature vector is transformed using the projection matrix of its cluster. Then, the transformation results of all clusters are weighted and summed according to their respective fusion weights to generate a fusion feature vector. This process fuses the feature information of each cluster, focusing on preserving key variations and nonlinear interaction effects within the refractive state group to which the sample belongs, thereby reducing the feature dimensionality. This is represented as:

[0067] In the formula, Indicates the first The fusion feature vector of each sample is the computational dimension of a single sample, representing a compact representation of the adaptive refractive state.

[0068] The fusion feature vector uses the cluster to which the sample belongs. and Weighted summation, combining intra-cluster variation preservation and density weighting, adapts to the modulation of feature interactions by refractive states, such as the nonlinear coupling between ambient brightness and pupil diameter for myopic users, thereby improving the generated fused feature vector. While retaining key discriminative information, the feature dimension is reduced.

[0069] It should also be noted that the first Sample based on Select the projection matrix of the cluster. and fusion weight This enables feature fusion that adapts to refractive states.

[0070] In one embodiment, conventional principal component analysis neglects the modulation of feature associations by refractive state. For example, the pupil diameter of myopic users is more sensitive to ambient brightness. This invention captures local nonlinearities through clustering and grouping, and utilizes projection matrices. Eliminating redundancy, such as the correlation redundancy between spherical and cylindrical lenses, while also incorporating density weighting. To address the problem of data imbalance, higher weights are assigned to dense clusters of samples, thereby enhancing the expressive power of fusion features on highly correlated features such as spherical and cylindrical power, and reducing the generalization error for sparse clusters such as extreme refractive errors.

[0071] In one specific embodiment, a deep neural network (prediction model or first prediction model) is constructed for predicting dynamic lighting parameters: A fully connected deep neural network is constructed as the core prediction architecture, which specifically includes an input layer, a feature extraction module, and a dynamic lighting parameter prediction module.

[0072] The input layer dimension of the neural network is consistent with the dimension of the fused feature vector, and it receives the feature representation after adaptive processing of the refractive state.

[0073] The feature extraction module is a hidden layer structure of the neural network, implemented using a three-layer fully connected structure. Each layer employs batch normalization to stabilize the training process, specifically including: The first hidden layer contains 128 neurons, which realize the primary abstraction of environment-refractive interaction features through the refractive perception adaptive activation function; The second hidden layer is reduced to 64 neurons, enhancing the ability to transform nonlinear features; The third hidden layer is further compressed to 32 neurons to extract high-level discriminative patterns.

[0074] The dynamic lighting parameter prediction module serves as the output layer and adopts a dual-branch structure. The first branch outputs the luminance prediction value through a linear fully connected layer, while the second branch independently outputs the color temperature prediction value.

[0075] The network as a whole processes data in a forward propagation manner, and the fused feature vectors are sequentially processed by linear transformation and adaptive activation at each layer. Finally, the output layer generates the predicted values ​​of brightness and color temperature.

[0076] Adaptive weight initialization in deep neural networks: Conventional random initialization methods such as Xavier initialization are sensitive to outliers such as skewed distributions and extreme spherical diopters in refractive data, which can easily lead to instability and slow convergence during neural network training.

[0077] This invention employs an adaptive weight initialization mechanism. It compresses the influence of outliers through a skewness-aware scaling matrix, calculates a bias correction term to compensate for distribution bias by combining refractive clustering information, and integrates feature statistics and the correction term to generate initial weights. This enhances the model's robustness to the distribution characteristics of refractive data and accelerates training convergence. The specific steps are as follows: 1) Calculate the adaptive scaling matrix: The skewness statistics of each dimension of the fused feature vector are calculated. Based on these skewness values ​​and the scaling adjustment hyperparameter, the scaling coefficients corresponding to each dimension are calculated. Then, these scaling coefficients form an adaptive scaling matrix to compress the influence of outliers with significant skewed distribution characteristics and balance the initial scale of each dimension, expressed as:

[0078] In the formula, This represents the adaptive scaling matrix, used to adjust the initial scale of the weights; This represents a diagonal matrix generating function; Indexed by feature dimensions, ; To integrate the total dimension of features; For the first The scaling factor of the dimension is calculated as follows: Accordingly, The scaling factor for the first dimension. Let be the scaling factor for the D-th dimension; For the first The skewness statistic of the dimensional feature, with spherical power as the main factor, is because there are few extreme positive values ​​such as hyperopia but the amplitude is large. Therefore, the refractive data is often right-skewed, i.e. positively skewed, resulting in a right-tailed distribution. To adjust the scaling hyperparameters and control the compression intensity of outliers, such as... .

[0079] 2) Calculate the bias correction term: Calculate the global mean vector of the fused feature vectors of all training samples. Then, calculate the difference between the centroid of each cluster and the global mean, and sum these differences using the cluster fusion weights. Finally, multiply the weighted sum by the bias learning rate hyperparameter to obtain the bias correction term, which compensates for the initialization bias caused by the uneven distribution of refractive state data, expressed as:

[0080] In the formula, This represents a bias correction term used to offset clustering bias. The global sample mean vector is calculated as follows: ; The total number of training samples; The bias learning rate hyperparameter controls the correction strength, such as... .

[0081] 3) Generate the initial weight matrix: The mean of the fused feature vectors of all training samples is calculated. This mean vector is then multiplied by an adaptive scaling matrix (skewness-aware scaling matrix) and added to a bias correction term to obtain the initial weight matrix for the first layer of the neural network. This matrix integrates the overall information of the fused features, the scaling adjustment for feature skewness, and the bias compensation for data imbalance, providing robust initial parameters for the first layer of the neural network, expressed as:

[0082] In the formula, This represents the initial weight matrix of the first layer of the neural network; It should be noted that, in order to handle the skewed distribution of refractive data, this initialization calculation method is only performed for the first layer, while the weight matrices of other layers are initialized using the standard Xavier method.

[0083] It should also be noted that the initialization method of the weight matrix in the first layer of the neural network integrates the feature mean and the adaptive scaling matrix. and bias correction term In particular, it enhances the predictive stability for anomalous samples such as those with high refractive errors.

[0084] In one specific embodiment, a feature extraction module based on an adaptive activation function is constructed: Users with high myopia require stronger nonlinear expression capabilities in low-light environments. Conventional ReLU activation functions cannot dynamically respond to changes in refractive state, which can easily lead to difficulties in capturing the complex interaction patterns between illumination and refractive state in deep features.

[0085] This invention employs a refractive sensing adaptive activation function, dynamically adjusting the nonlinear response characteristics using the mean spherical power. By modulating the steepness and threshold position of the activation curve, it enhances the neural network's ability to express refractive state-sensitive features and optimizes the modeling effect of the illumination-refractive interaction mode. The specific steps are as follows: 1) Perform linear transformation calculations: Using the weight matrix and bias vector of the current layer, a linear transformation is performed on the output feature vector of the previous layer to obtain the input value of the activation function, expressed as:

[0086] In the formula, This represents the input value of the activation function and characterizes the result of the linear transformation; Let be the weight matrix of the l-th layer of the neural network, and be the trainable parameters; For neural network layer indexing; For the lth neural network The output feature vector of layer 1; Let be the bias vector of the l-th layer of the neural network, and be the trainable parameters.

[0087] It should be noted that the input feature vector of the first layer of the neural network is , The fused feature vector of a sample represents the computational dimension of a single sample.

[0088] 2) Applying the refractive sensing activation function: The linear transformation result is input into an adaptive activation function, which dynamically adjusts its nonlinear response characteristics based on the user's average spherical power to calculate the output feature vector of the current layer. This allows the neural network to flexibly enhance or suppress the nonlinear expression of features according to the user's refractive state, better capturing the complex interaction patterns between illumination and refractive state, as expressed below:

[0089]

[0090] In the formula, This represents the output feature vector of the l-th layer of the neural network, capturing the deep patterns of the interaction between refractive state and environment; An adaptive activation function; As a nonlinear steepness control parameter, it determines the slope of the function curve, such as... ; The threshold offset parameter controls the threshold position of the adaptive activation function, adjusting the starting point of the nonlinear response to adapt to different refractive states. Highly myopic users require a lower activation threshold, such as... .

[0091] It should be noted that the adaptive activation function, combined with the mean spherical power, dynamically modulates the nonlinear steepness control parameters. A low average spherical power indicates high myopia. The coefficient increases, enhancing the nonlinear expression capability under low light and improving the ability of deep features to capture the light-refractive interaction.

[0092] Furthermore, the data is processed through forward propagation using a fully connected neural network, followed by activation via a layer-by-layer linear transformation activation function until the last layer of the fully connected neural network. The output feature vector of the last layer of the fully connected neural network is defined as... .

[0093] In one specific embodiment, a dynamic illumination parameter prediction module based on refractive perception is constructed: Although the deep features after feature extraction contain environmental-refractive interaction information, direct linear mapping cannot adapt to the heterogeneous influence of refractive state on brightness and color temperature. Highly myopic users prefer low brightness to alleviate glare sensitivity, and conventional fully connected regression methods ignore the nonlinear modulation effect of refractive state in the output stage.

[0094] This invention employs a dual-path structure to predict target brightness and color temperature. The main path linearly maps the deep feature vector through a weight vector and a global bias term to obtain the basic predicted value. The modulation path, based on the deep feature vector and the average spherical power, calculates the refractive dynamic modulation factor for brightness prediction and the refractive dynamic modulation factor for color temperature prediction through refractive influence function scaling, feature projection, and interaction terms. The final predicted value is obtained by adding the product of the corresponding modulation factor and its fusion coefficient to the basic predicted value. An offset dynamically modulated by the refractive state is superimposed on the basic prediction, prioritizing the preference of highly myopic users for brightness reduction and color temperature adjustment, capturing the interaction effect between refractive state and deep features, expressed as:

[0095] In the formula, Indicates the first The predicted brightness value for each sample, in candela per square meter; Indicates the first The predicted color temperature value for each sample, in Kelvin; The weight vector representing brightness prediction is a trainable parameter; for transpose; The global bias term representing brightness prediction is a trainable parameter; The weight vector representing color temperature prediction is a trainable parameter. for transpose; The global bias term representing color temperature prediction is a trainable parameter; The fusion coefficient represents the luminance modulation path and controls... The intensity of contribution, such as, ; Represents the blending coefficient of the color temperature modulation path, controlling The intensity of contribution, such as, ; Indicates the first The refractive dynamic modulation factor predicted for the brightness of each sample, based on and The calculation showed that the brightness of the enhanced high myopia sample was reduced, and the calculation method is expressed as follows: ; The refractive dynamic modulation factor representing color temperature prediction is based on and The calculation shows that the predicted color temperature is dynamically adjusted to respond to the nonlinear modulation effect of refractive state on color temperature perception. The calculation method is expressed as follows: ; The brightness refractive effect function is expressed as follows: ; This represents the brightness modulation amplitude hyperparameter, controlling the output range of the function, such as... ; This represents the hyperparameter of brightness modulation curvature, controlling the sensitivity of the function, such as... ; The color temperature refractive effect function is represented by the following calculation method: ; Indicates the color temperature modulation amplitude over-parameter (e.g.) ), which are trainable parameters; Indicates the color temperature modulation curvature hyperparameter (e.g.) ), which are trainable parameters; Represents the brightness modulation weight vector. yes The transpose of the vector is used for projection. The generation basis offset, which extracts the feature subspace related to brightness modulation, is a trainable parameter; This represents the color temperature modulation weight vector. yes The transpose of the vector is used for projection. The generation basis offset, which extracts the feature subspace related to color temperature modulation, is a trainable parameter. Represents the brightness interaction weight vector. yes The transpose of the deep feature vector The interaction weight vector extracted for color temperature prediction is a trainable parameter. This represents the color temperature interaction weight vector. yes The transpose of the deep feature vector The interaction weight vector extracted for brightness prediction is a trainable parameter. This represents the luminance interaction gain coefficient, which controls the interaction terms in the luminance prediction path. The gain coefficient, such as, ; This represents the color temperature interaction gain coefficient, which controls the interaction terms in the color temperature prediction path. The gain coefficient, such as, .

[0096] It should be noted that, and Characterize the main path calculation and provide basic predictions. and Characterization of modulation path calculation, based on Dynamically scaling the modulation amplitude, for example, when hour, This amplifies the brightness adjustment for users with high myopia. Furthermore, in the dual-path structure, the main path ensures global consistency, while the modulation path incorporates local refractive dependence and bias correction terms. Synergistically improve robustness to extreme refractive samples.

[0097] It should also be noted that, As deep features, they contain high-dimensional environment-refractive interaction information, such as brightness, color temperature, and pupil diameter, but not all dimensions are equally modulated by refractive states. Item and The two projection operations are performed using a trainable weight matrix. Mapping to a low-dimensional subspace preserves only features strongly correlated with refractive state, such as the ambient brightness sensitivity of highly myopic users, thereby reducing the modulation factor. and This enables efficient and targeted refractive adaptation, avoids noise interference, and improves prediction robustness.

[0098] It should also be noted that, Item and Item implementation and The interaction enhances nonlinearity, such as in high myopia. Increasing the negative value leads to negative saturation of the output, reducing its efficiency. At the same time, combined and accomplish and explicit coupling, thereby capturing the fused feature vector The nonlinear interaction characteristics emphasized in the text, such as the interaction effect between ambient brightness and pupil diameter in patients with high myopia.

[0099] It should also be noted that, Item and The term represents the weighted fusion modulation term, prioritizing optimization of highly myopic samples, such as... It exhibits higher brightness sensitivity, through and Dynamically adjust forecast values ​​and respond directly. , and based on the influence function of refractive state The normalized scaling operation and the adaptive activation operation based on the adaptive activation function form an end-to-end refractive adaptation chain.

[0100] In one embodiment, the prediction model dynamically adjusts the loss weights of brightness and color temperature prediction errors using a refractive state-weighted loss function and a weighting factor calculated using the average spherical power.

[0101] In one specific embodiment, highly myopic users are more sensitive to brightness errors, but conventional loss functions cannot prioritize optimizing such key parameters, and conventional mean squared error loss functions do not consider the heterogeneous impact of refractive state on prediction errors.

[0102] This invention employs a refractive state-weighted loss function, dynamically adjusting the loss weights for brightness and color temperature prediction errors through weighting factors calculated based on the average spherical power. This allows the model optimization process to focus on improving the accuracy of brightness prediction for highly myopic users. The specific steps are as follows: 1) Calculate the refractive weighting factor: Based on the user's average spherical power, the refractive weighting factor is calculated using a mapping function in the form of a sigmoid activation function, expressed as:

[0103] In the formula, Indicates the first The refractive weighting factor for each sample is used to adjust the weight of the luminance error term; As a slope control parameter, it adjusts the sensitivity to changes in weights, such as... .

[0104] It should be noted that the refractive weighting factor The calculation method is similar to the Sigmoid activation function, which can map the mean spherical power to... To ensure the high myopia sample Approaching 1 improves the accuracy of brightness prediction for highly myopic users.

[0105] 2) Calculate the refractive weighted loss value: The squares of the brightness prediction error and color temperature prediction error are calculated. These errors are then weighted using a refractive weighting factor. The weighted error terms for all samples are then averaged to obtain the final loss value. This allows the loss function to automatically increase the accuracy requirement for brightness prediction based on the user's refractive state during the optimization process, especially prioritizing brightness prediction for highly myopic users. This is expressed as:

[0106] In the formula, Represents the loss function; For the first The predicted brightness value for each sample; For the first The target brightness value for each sample; For the first The predicted color temperature value for each sample; For the first The target color temperature value for each sample.

[0107] It should be noted that during the calculation of the refractive weighting factor, the refractive weighting factor... This helps the loss function prioritize optimizing brightness prediction. This indicates high myopia; users with high myopia are more sensitive to brightness errors, such as in the settings. Under the premise that, when hour, ,when hour, to improve The weights of the items are prioritized for optimizing brightness prediction.

[0108] It should also be noted that, The term represents the color temperature error weighting term, when When the value increases, the brightness weight is higher. To reduce the loss function, the weights must remain constant at 1 to ensure optimal balance.

[0109] In one specific embodiment, parameter updates are performed based on a refractive sensing gradient optimization algorithm: The standard Adam optimizer ignores the modulating effect of refractive state on parameter updates, which can easily lead to slow convergence speed and easy getting trapped in local optima for samples with high refractive errors.

[0110] This invention utilizes a refractive sensing dynamic learning rate mechanism to amplify the parameter update step size proportional to the average spherical power, thereby accelerating the convergence of relevant parameters for high refractive error samples and improving model training efficiency. The specific steps are as follows: 1) Calculate the refractive perception learning rate: Based on the absolute value of the average spherical power of users, the base learning rate is scaled up proportionally to obtain the dynamic learning rate for refractive perception. This allows the model to use a larger parameter update step size for user samples with more severe refractive errors during the optimization process, expressed as:

[0111] In the formula, The dynamic learning rate represents refractive perception and is used to adjust the parameter update step size. The dynamic learning rate of basic refractive perception controls the overall optimization speed, such as... ; The refractive accommodation coefficient determines the strength of the effect of refractive state on the learning rate, such as... ; It is the absolute value of the average spherical power.

[0112] It should be noted that in the settings Under the premise that, when hour, Increase by 12% because the parameters of high refractive error samples require a larger update step size to accelerate convergence; continue adjusting. Can make hour, The increase is even higher.

[0113] 2) Perform refractive sensing parameter updates: An improved Adam optimization rule is adopted, and the bias-corrected first-order moment and second-order moment estimates are calculated using the Adam optimizer. Then, the standard Adam update rule is scaled using the dynamic learning rate of refractive perception to update the parameters of the neural network, thereby accelerating the convergence speed of relevant parameters for high refractive error samples. This is expressed as:

[0114] In the formula, express The parameter values ​​for the next iteration specifically refer to the trainable parameters of the neural network, including all trainable parameters of the neural network. express The parameter values ​​for the next iteration; is the first-order moment estimation vector of the Adam optimizer, representing the mean gradient after bias correction; Let be the second-order moment estimation vector of the Adam optimizer, representing the gradient variance after bias correction. Indicates to Take the square root of each element; For numerical stability constants, to prevent division by zero errors, such as, ; Represents the identity matrix. Implement construction and Vectors of the same dimension.

[0115] In one specific embodiment, iterative training and early stopping strategy: The model training employs a mini-batch gradient descent strategy, combined with a refractive sensing gradient optimization algorithm to dynamically update network parameters. Each iteration consists of three steps, specifically: Forward propagation calculates the predicted brightness and color temperature for the current batch; Calculate the batch loss value using the loss function; Backpropagation is used to calculate the gradient and update the refractive sensing parameters.

[0116] The performance of the validation set is continuously monitored during the training process. After each training cycle, the validation set loss is calculated and the optimal loss value is recorded. If the verification loss does not decrease for 10 consecutive periods, that is, the current loss is higher than the historical best value, the early stopping mechanism is triggered, and the model parameters before that period are saved as the final model. At the same time, set a maximum training cycle limit, such as 2000 cycles, to prevent infinite iteration.

[0117] Early stopping conditions and maximum period limits work together to prevent overfitting and ensure that the model terminates training when the performance on the validation set peaks.

[0118] This invention provides a dynamic lighting adjustment method based on refractive state adaptation, comprising: a light source controller and a lighting source; The luminance prediction value and color temperature prediction value are obtained based on the above dynamic illumination prediction method based on refractive state. The predicted brightness and color temperature values ​​are input to the light source controller; The brightness and color temperature of the lighting source are dynamically adjusted by the light source controller.

[0119] In one embodiment, the light source controller controls the lighting source through instructions, computer programs, or signals, including one or more of the following: wireless Bluetooth control, wired control, and local area network control.

[0120] In one specific embodiment, dynamic lighting parameter prediction and lighting adjustment: 1) Real-time data preprocessing: Collect new user raw data in the required format and perform adaptive normalization scaling based on the dynamic data processing method of eye refractive state; 2) Feature fusion: Based on the pre-stored cluster centroid and projection matrix, calculate the index of the refractive cluster to which the user belongs, and generate a fused feature vector by applying the corresponding projection matrix and fusion weight; 3) Neural network prediction: The fused features are input into the trained deep neural network, and the output brightness prediction value and color temperature prediction value are output through forward propagation.

[0121] Furthermore, lighting adjustment is performed based on the prediction results. Specifically, the predicted parameters are sent to the lighting controller to drive the LED light source to dynamically adjust the output luminous flux and color temperature. The system continuously monitors changes in ambient light and user feedback. When environmental sensor data or user activity type is updated, the above process is re-triggered to achieve closed-loop adaptive adjustment.

[0122] In one embodiment, a comparative analysis of prediction accuracy under different refractive states is conducted to evaluate the predictive performance of the invention's technology in user groups with different refractive states. By comparing the prediction error distribution of five types of users—high myopia, moderate myopia, mild myopia, emmetropia, and hyperopia—when using the invention's technology, the advantages of the invention in refractive state adaptive processing are verified, particularly its optimization effect on high myopia users. The experiment compares the error distribution of the five types of users (high myopia: average spherical power < -6 diopters; moderate myopia: -6 to -3 diopters; mild myopia: -3 to 0 diopters; emmetropia: 0 to +1 diopters; hyperopia: > +1 diopters) on two key indicators: luminance prediction and color temperature prediction. The unit of luminance prediction error is candela per square meter (measuring light intensity), and the unit of color temperature prediction error is Kelvin (measuring the color characteristics of a light source). The experiment uses a violin plot to illustrate the distribution characteristics of the prediction error, while also displaying the overall distribution range, density, and key statistics of the data. For example, in the luminance prediction error distribution plot... Figure 5 As shown, the horizontal axis represents the five refractive state user groups, and the vertical axis represents the luminance prediction error value (unit: candela per square meter). Experimental data shows that the prediction error distribution is most concentrated (narrowest violin shape) for highly myopic users, with the lowest median error. As the degree of myopia decreases, the error distribution range gradually widens (widening violin shape), and the median error gradually increases. The error distribution characteristics of hyperopic users are similar to those of mildly myopic users, both exhibiting positive skewness (longer upper tail), indicating the existence of a small number of relatively large prediction errors. In the color temperature prediction error distribution map, as shown... Figure 6As shown, the horizontal axis represents the five user groups with different refractive states, and the vertical axis represents the color temperature prediction error (in Kelvin). The distribution trend is similar to that of brightness prediction. The error distribution is most concentrated and has the lowest median for highly myopic users, while the error distribution is widest and has the highest median for emmetropic users. The error distributions for moderate and mild myopia users fall between these two extremes. All groups exhibit a positively skewed error distribution. Experimental results show that highly myopic users have the smallest and most stable errors in brightness and color temperature prediction, verifying the targeted optimization effects of key technologies such as the refractive state influence function, adaptive activation function, and dynamic feature fusion for highly myopic users. The significantly narrower violin shape and lower median in the figure visually demonstrate this advantage. Furthermore, the error distribution shows a regular gradient change with refractive state, indicating that the technology can dynamically adjust model parameters and processing strategies according to the user's refractive state. The significant performance difference from highly myopic to moderate myopia demonstrates the technology's sensitivity to subtle changes in refractive state.

[0123] In one specific embodiment, this invention analyzes the nonlinear influence of refractive state on light perception, exploring the complex nonlinear relationship between refractive state (quantified by mean spherical power) and light perception characteristics, and verifying the superiority of this invention in capturing these relationships by comparing it with traditional methods. Three key indicators are examined: light sensitivity factor, preferred brightness value, and response time to changes in light intensity. The experiment is divided into three parts: the relationship between refractive state and light sensitivity factor, the relationship between refractive state and preferred brightness value, and a comparison between traditional methods and the technology of this invention on the influence of refractive state on response time, such as... Figure 7 , Figure 8 , Figure 9As shown in the figure, the horizontal axis of all charts represents the average spherical power (diopters), and the vertical axes represent the light sensitivity factor (dimensionless relative value), preferred brightness value (candela per square meter), and response time (milliseconds), respectively. In the light sensitivity factor analysis, the scatter plot shows the relationship between refractive power and sensitivity for each sample. The color intensity indicates sample density (darker areas show denser samples). The red dashed trend line exhibits a distinct S-shaped curve. When refractive power < -4 diopters, the sensitivity factor increases significantly; when refractive power > 0 diopters, the sensitivity factor tends to stabilize. In the preferred brightness value analysis, the scatter distribution shows a negative correlation between refractive power and preferred brightness. The purple dashed trend line exhibits an inverse S-shaped characteristic. Users with high myopia (refractive power < -6 diopters) have significantly lower preferred brightness values ​​than other groups. The sample density shows that most users' refractive powers are concentrated in the range of -8 to +2 diopters. In the response time comparison, the blue scatter plots and trend line represent the traditional method, while the orange scatter plots and trend line represent the technology of this invention. Both technologies show a shortened response time when the refractive power decreases (myopia worsens). Experimental results show that the S-shaped trend line perfectly verifies the design rationality of the refractive state influence function (based on the hyperbolic tangent function) proposed in the technical document, proving that the technology of this invention can accurately model the complex nonlinear relationship between refractive state and light perception. Furthermore, the technology of this invention comprehensively outperforms the traditional method in terms of response time, especially in the high myopia region (refractive power < -6), verifying the effectiveness of the refractive perception gradient optimization algorithm and adaptive activation function, indicating that the technology of this invention can dynamically adjust the learning rate and nonlinear response characteristics according to the user's refractive state.

[0124] The present invention also discloses a computer program product or system, including a computer program that, when executed by a processor, implements the above-described dynamic illumination prediction method steps based on refractive state.

[0125] Figure 2 A schematic diagram of a dynamic illumination prediction system based on refractive state provided in this embodiment of the invention specifically includes: Acquisition module: Acquires the refractive data of the subject's eye; Feature module: The refractive data is clustered and feature fusion is performed to obtain fused features; Prediction module: Inputs the fused features into the trained prediction model to obtain the brightness prediction value and the color temperature prediction value; The prediction model initializes its weight parameters by compressing outliers in the refractive data and compensating for distribution biases in refractive clustering.

[0126] This invention provides a dynamic illumination adjustment system based on refractive state adaptation, comprising: Light source controller, lighting source; Prediction module: Based on the above dynamic illumination prediction method based on refractive state, the predicted values ​​of luminance and color temperature are obtained; Transmission module: Inputs the predicted brightness and color temperature values ​​to the light source controller; Control module: Dynamically adjusts the brightness and color temperature of the lighting source through the light source controller.

[0127] Figure 3 An embodiment of the present invention provides a schematic diagram of a computer device, specifically including: The system includes a memory and a processor; the memory is used to store program instructions; the processor is used to invoke the program instructions, when the program instructions are executed to perform any of the above-described dynamic illumination prediction methods based on refractive state, or to execute the above-described dynamic illumination adjustment methods based on refractive state adaptation.

[0128] The present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, performs any of the above-described dynamic illumination prediction methods based on refractive state, or performs the above-described dynamic illumination adjustment method based on refractive state adaptation.

[0129] The verification results of this verification embodiment show that assigning inherent weights to indications can improve the performance of this method compared to the default settings. Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. In the several embodiments provided in this application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for example, the division of units is merely a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling or direct coupling or communication connection may be through some interfaces, indirect coupling or communication connection of devices or units, and may be electrical, mechanical, or other forms. The units described as separate components may or may not be physically separated; the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of this embodiment. Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units. Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. This program can be stored in a computer-readable storage medium, which may include: read-only memory (ROM), random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0130] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0131] The computer device provided by the present invention has been described in detail above. For those skilled in the art, there will be changes in the specific implementation and application scope based on the ideas of the embodiments of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A dynamic illumination prediction method based on refractive state, characterized in that, include: Obtain the refractive data of the subject's eye; The refractive data are clustered and feature fusion is performed to obtain fused features; The fused features are input into the trained prediction model to obtain the brightness prediction value and the color temperature prediction value; The prediction model initializes its weight parameters by compressing outliers in the refractive data and compensating for distribution biases in refractive clustering.

2. The dynamic illumination prediction method based on refractive state according to claim 1, characterized in that, Outliers in the compressed refractive data are compressed using a skewness-aware scaling matrix; the distribution bias of the compensated refractive clusters is compensated by calculating a bias correction term using refractive cluster information.

3. The dynamic illumination prediction method based on refractive state according to claim 1, characterized in that, The prediction model learns the characteristic relationship between refractive sensitivity and brightness through an adaptive activation function of refractive perception; the adaptive activation function of refractive perception is obtained by dynamically adjusting the steepness of the activation curve and the threshold position by the average spherical power.

4. The dynamic illumination prediction method based on refractive state according to claim 1, characterized in that, The clustering is based on the average spherical power in the refractive data to obtain grouped sample clusters, and a projection matrix is ​​constructed to remove redundant features from the grouped sample clusters to obtain feature clusters; The feature fusion is based on the density of feature clusters to calculate fusion weights, and feature clusters with different fusion weights are fused to obtain fused features; The projection matrix is ​​constructed by performing singular value decomposition on the feature vectors of samples within the cluster to obtain a left singular matrix and a singular value matrix, and then calculating the projection matrix of the current cluster based on the left singular matrix and the singular value matrix.

5. A dynamic illumination adjustment method based on refractive state adaptation, characterized in that, Light source controller and lighting source, including: The dynamic illumination prediction method based on refractive state according to any one of claims 1-4 obtains luminance prediction values ​​and color temperature prediction values; The predicted brightness and color temperature values ​​are input to the light source controller; The brightness and color temperature of the lighting source are dynamically adjusted by the light source controller.

6. A computer program product comprising a computer program or instructions, characterized in that, The computer program or instructions are executed by the processor to implement the dynamic illumination prediction method based on refractive state as described in any one of claims 1-4, or to implement the dynamic illumination adjustment method based on refractive state adaptation as described in claim 5.

7. A computer device comprising a memory, a processor, and a computer program or instructions stored in the memory, characterized in that, The computer program or instructions are executed by the processor to implement the dynamic illumination prediction method based on refractive state as described in any one of claims 1-4, or to implement the dynamic illumination adjustment method based on refractive state adaptation as described in claim 5.

8. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, The computer program or instructions are executed by the processor to implement the dynamic illumination prediction method based on refractive state as described in any one of claims 1-4, or to implement the dynamic illumination adjustment method based on refractive state adaptation as described in claim 5.

Images