An intelligent display control method and system for liquid crystal panel image recognition driving

Abstract

The application discloses an intelligent display control method and system for liquid crystal panel image recognition driving, and relates to the technical field of liquid crystal panels. The method comprises the following steps: collecting user image data of a liquid crystal panel through a camera and preprocessing the user image data to obtain a user image of the liquid crystal panel; inputting the user image of the liquid crystal panel into an improved VGG-Face model to output a user age prediction value; collecting environment data of the liquid crystal panel, and based on the environment data of the liquid crystal panel and the user age prediction value, performing display adjustment on the liquid crystal panel through a fuzzy logic control method to obtain a first display adjustment result; collecting user feedback data, training a gradient boosting tree model by taking the user feedback data as a training set to obtain a trained gradient boosting tree model, inputting the first display adjustment result into the trained gradient boosting tree model, and outputting a second display adjustment result, so that intelligent display control of the liquid crystal panel is realized.

Description

Technical Field

[0001] This invention relates to the field of liquid crystal panel technology, specifically to an intelligent display control method and system for driving image recognition in liquid crystal panels. Background Technology

[0002] With its advantages of clear image quality, low power consumption, and wide adaptability, LCD panels have become a core component of various display terminals. As users' demands for visual experience and visual health continue to rise, adaptive display adjustment technology for LCD panels has become an important direction for industry development. Ambient light intensity is a key environmental factor affecting the visibility of LCD panels and user viewing comfort. Adjusting display parameters based on ambient light perception allows the panel to adapt to different lighting environments, which is an important technical means to achieve basic intelligent control of LCD panels and has become one of the mainstream application directions for LCD panel display adjustment.

[0003] Traditional LCD panels typically adjust their display based on ambient light intensity. An ambient light sensor integrated into the LCD panel collects real-time ambient light data around the panel. The collected raw light data undergoes preprocessing operations such as filtering, noise reduction, and range normalization. Then, according to a pre-set mapping rule between light intensity and display parameters, the processed light intensity value is matched to the corresponding display adjustment parameters such as blue light intensity and panel brightness. These adjustment parameters are directly transmitted to the LCD panel's driving circuit to complete the adaptive adjustment of the panel's display state.

[0004] However, this method of adjusting the display based solely on light intensity only designs the adjustment logic around a single environmental parameter. It cannot combine the user's personalized physiological characteristics to formulate differentiated display adjustment strategies. As a result, the uniform adjustment scheme under the same lighting environment cannot adapt to the visual physiological needs of users of different age groups. Ultimately, the display adjustment of the LCD panel lacks personalized adaptability and it is difficult to improve the overall viewing comfort and visual health protection effect from the perspective of user physiological characteristics. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent display control method and system for LCD panel image recognition driving, thereby resolving the problems existing in the background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an intelligent display control method and system for LCD panel image recognition driving, comprising the following steps: Step S1: Acquire user image data of the LCD panel through a camera and perform preprocessing to obtain the user image of the LCD panel; Step S2: Input the user image of the liquid crystal panel into the improved VGG-Face model and output the predicted user age value; Step S3: Collect LCD panel environmental data. Based on the LCD panel environmental data and the user's age prediction value, adjust the display of the LCD panel using fuzzy logic control to obtain the first display adjustment result. Step S4: Collect user feedback data, use the user feedback data as a training set to train the gradient boosting tree model, obtain a trained gradient boosting tree model, input the first display adjustment result into the trained gradient boosting tree model, and output the second display adjustment result, thereby realizing intelligent display control of the LCD panel.

[0007] Preferably, the step of acquiring user image data from the LCD panel via a camera and performing preprocessing to obtain the user image of the LCD panel includes the following specific steps: The camera captures the original image of the user's face at a preset sampling frequency; the user's facial region is detected and located in the original image; the image of the user's facial region is standardized and enhanced to finally obtain the user image on the LCD panel.

[0008] Preferably, the step of inputting the user image from the liquid crystal panel into the improved VGG-Face model and outputting the user's age prediction value includes the following steps: Construct an improved VGG-Face model; input the user image of the liquid crystal panel into the improved VGG-Face model: ; Where age is the predicted age of the user. The output scalar value is the sigmoid value of the fully connected layer, and Sigmoid() is the Sigmoid activation function. S is a flattened one-dimensional vector, and S is a preset scaling factor.

[0009] Preferably, the construction of the improved VGG-Face model includes the following steps: The improved VGG-Face model includes: an input convolutional layer, four feature extraction stages, and a fully connected layer; The input convolutional layer contains 64 convolutional kernels of size 3×3, with a stride of 1 and padding of 1; The four feature extraction stages are as follows: the first feature extraction stage includes two Fire modules and a pooling layer; the second feature extraction stage includes four Fire modules, residual connections, and a pooling layer; the third feature extraction stage includes eight Fire modules, residual connections, and a pooling layer; and the fourth feature extraction stage includes four Fire modules, residual connections, and a pooling layer. The fully connected layer contains one output neuron.

[0010] Preferably, the collection of LCD panel environmental data includes the following steps: LCD panel environmental data includes ambient light intensity measured by an ambient light sensor. Local time value And the geographic location factor L.

[0011] Preferably, the step of adjusting the display of the liquid crystal panel based on the liquid crystal panel environmental data and the user's age prediction value using fuzzy logic control to obtain a first display adjustment result includes the following steps: Construct a membership function between LCD panel environmental data and predicted user age values; Substitute the LCD panel environmental data and the user age prediction value into the corresponding membership function to obtain the membership vector of the LCD panel environmental data and the membership vector of the user age prediction value. The aggregate membership function value is calculated based on the membership vector of the LCD panel environmental data, the membership vector of the user age prediction value, and the preset fuzzy rule base. The first display adjustment result is calculated based on the aggregate membership function value.

[0012] Preferably, the calculation of the first display adjustment result based on the aggregate membership function value includes the following specific steps: The centroid method is used to convert the aggregate membership function values ​​into the first display adjustment result. The calculation formula is as follows: ; in, The first displayed adjustment result is shown, where u is the blue light adjustment variable. This represents the aggregate membership function value.

[0013] Preferably, the step of collecting user feedback data and using the user feedback data as a training set to train the gradient boosting tree model to obtain a trained gradient boosting tree model specifically involves: Collect user feedback data, wherein the user feedback data is the user's feedback value f, f= - , Indicates the target adjustment value; The feedback values, along with the corresponding LCD panel environmental data and the predicted user age, form a training set. The forward stepwise algorithm is used to train the gradient boosting tree model using the training set, resulting in a trained gradient boosting tree model.

[0014] Preferably, the step of inputting the first display adjustment result into the trained gradient boosting tree model and outputting the second display adjustment result specifically involves: The initial adjustment result is then input into the trained gradient boosting tree model: ; in, This represents the feedback prediction value. () represents a gradient boosting tree model. This represents the input feature vector; The second display adjustment result is obtained by adding the first display adjustment result to the feedback prediction value: ; in, The second display shows the adjustment results.

[0015] An intelligent display control system driven by image recognition of a liquid crystal panel includes a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement the steps of the above method.

[0016] This invention provides an intelligent display control method driven by image recognition of a liquid crystal panel, involving machine learning and deep learning technologies, which has the following beneficial effects: (1) The VGG-Face model is used to predict the age of LCD panel users. The deep convolutional neural network designed for facial features has the ability to represent and extract facial features, and can accurately capture key facial features related to age, providing a reliable model foundation for age prediction. Relying on the face-specific feature extraction advantage of the VGG-Face model, the accuracy of user age prediction can be effectively improved, providing accurate user attribute data support for subsequent LCD panel display adjustment based on user age, allowing display adjustment to revolve around the core personalized indicator of user age, and ensuring the personalized design basis of the display adjustment scheme.

[0017] (2) Using the improved VGG-Face model, while retaining the original model's powerful facial feature extraction capabilities, the number of model parameters was significantly reduced by embedding the Fire module and introducing residual connections. This greatly reduced the model's computational overhead and memory usage, adapting to the hardware computing power and storage resources of LCD panel terminal devices and meeting the device's local real-time inference requirements. The introduction of residual connections effectively solved the gradient vanishing problem in the training process of deep neural networks, improved the model's training effect and convergence, and further optimized the accuracy and stability of age prediction. The Fire module, through channel compression and parallel convolution design, reduced the amount of computation while retaining rich feature information, ensuring that the accuracy of age prediction was not affected.

[0018] (3) The gradient boosting tree model is used to perform personalized calibration on the first display adjustment result obtained by fuzzy logic control. This model is suitable for the characteristics of small-scale and structured user feedback data in the LCD panel display control scenario. It is not easy to overfit when training with a small amount of data. It can automatically learn the complex nonlinear relationship and interaction effect between environmental data such as ambient light intensity, user age, geographical location, time and user personalized display preferences. No manual design of feature combination is required, and accurate personalized correction of general adjustment results is achieved. The gradient boosting tree model has high training and inference efficiency and supports local incremental learning on LCD panel terminal devices. All data processing and model training are completed locally. Combined with the first-in-first-out buffer design, while protecting the privacy of user feedback data, it can realize dynamic updating and adaptive learning of the model based on continuous user operation feedback. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of 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.

[0020] Figure 1 This is a flowchart illustrating the steps of an intelligent display control method for driving image recognition on a liquid crystal panel, as proposed in this invention. Figure 2 This is a framework diagram of the improved VGG-Face model in the intelligent display control method driven by image recognition of a liquid crystal panel proposed in this invention. Figure 3 This is a step hierarchy diagram of obtaining the first display adjustment result in an intelligent display control method driven by image recognition of a liquid crystal panel proposed in this invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see Figures 1-3 The present invention provides a technical solution: an intelligent display control method for driving image recognition of a liquid crystal panel.

[0023] Step S1: Acquire user image data of the LCD panel through a camera and perform preprocessing to obtain the user image of the LCD panel; Step S1 begins with capturing image data using a camera module integrated above or to the side of the LCD panel. When the LCD panel is in operation, the camera module operates intermittently at a preset low sampling frequency (e.g., 1 Hz), with its optical lens facing the user's viewing area 0.5 to 2 meters in front of the panel to capture raw RGB images that may include the user's face.

[0024] Subsequently, real-time preprocessing is performed on the captured images. The first step of preprocessing is the detection and localization of the user's facial region. A lightweight convolutional neural network face detection algorithm (such as SSD or YOLO algorithm) is used to analyze the image, identify and define the location coordinates of valid face regions. The valid face region is the region with a confidence level higher than a preset threshold (e.g., 0.7). When a single valid face is detected, the image of that face region is cropped as the basis for subsequent processing. If multiple faces are detected, they are selected according to a preset priority rule: faces whose bounding box center point falls within a certain range in the center of the image (e.g., a circular area with the image center as the center and a radius of 20% of the image's shorter side length) are prioritized; if no face is detected in the central region, the face with the largest area is selected as the primary target for cropping. The second step is to standardize and enhance the cropped user facial region image. First, geometric normalization is performed, using proportional scaling combined with boundary padding to scale the face image to a fixed pixel size. Subsequently, a contrast-limited adaptive histogram equalization method is applied to compensate for uneven illumination, and a Gaussian filter is used for smoothing to suppress random noise. Finally, pixel value normalization is performed, and the processed image pixel values ​​are linearly transformed and scaled to the range [0,1].

[0025] After the above detection, cropping, standardization and enhancement operations, a user image of the LCD panel with uniform size, illumination correction and noise suppression is finally output. In this process, all image data is processed in the device's local memory. The original image frame is discarded after the face region is cropped. The LCD panel user image generated after preprocessing is transmitted to the subsequently improved VGG-Face model and the age prediction calculation is completed. It will also be immediately cleared from memory, without storing or transmitting any user facial image data, thus ensuring user privacy and security.

[0026] Step S2: Input the user image of the liquid crystal panel into the improved VGG-Face model and output the predicted user age value; Step S2 takes the preprocessed LCD panel user image output from step S1 as input, performs forward inference through a deep convolutional neural network based on the improved VGG-Face architecture, and finally outputs a predicted user age value representing the user's age.

[0027] The improved VGG-Face model comprises an input convolutional layer, a feature extraction stage, and a fully connected layer. This VGG-Face model embeds multiple Fire modules and introduces residual connections within the traditional VGG-Face framework, significantly reducing the number of parameters while maintaining representational power.

[0028] The input convolutional layer uses 64 3×3 convolutional kernels to process the user image of the liquid crystal panel (224×224×3), with a stride of 1 and padding of 1 to ensure that the spatial dimensions remain unchanged. It is then connected to a ReLU activation function, ultimately outputting an initial feature map of size 224×224×64. This initial feature map from the input convolutional layer serves as the input to the Fire module stack in the subsequent feature extraction stage.

[0029] The improved VGG-Face model comprises four feature extraction stages. The first feature extraction stage consists of two Fire modules. The input to this stage is the initial feature map (224×224×64) output from the input convolutional layer. After processing by the two Fire modules, a 2×2 max pooling (stride 2) is directly performed, outputting a feature map of size 112×112×128.

[0030] The second feature extraction stage contains four Fire modules. The feature map after pooling in the first feature extraction stage is used as input and is called the second initial input feature map. The second initial input feature map passes through the four Fire modules in sequence to obtain the output feature map of the second Fire module. Then, a residual connection is introduced: the second initial input feature map and the output feature map of the second Fire module are added element by element. After processing, 2×2 max pooling (with a stride of 2) is performed to output a feature map with a size of 56×56×128.

[0031] The third feature extraction stage consists of 8 Fire modules. The feature map obtained from pooling in the second feature extraction stage is used as input, called the third initial input feature map. This third initial input feature map passes through the 8 Fire modules sequentially to obtain the output feature map of the third Fire module. A residual connection is then introduced: the third initial input feature map and the output feature map of the third Fire module are added element-wise. After processing, 2×2 max pooling (with a stride of 2) is performed, outputting a feature map of size 28×28×256.

[0032] The fourth feature extraction stage consists of four Fire modules. The feature map obtained from pooling in the third feature extraction stage is used as input, referred to as the fourth initial input feature map. This fourth initial input feature map passes through the four Fire modules sequentially, resulting in the output feature map of the fourth Fire module. A residual connection is then introduced: the fourth initial input feature map and the output feature map of the fourth Fire module are added element-wise. After processing, 2×2 max pooling (with a stride of 2) is performed, outputting a feature map of size 14×14×256.

[0033] It should be noted that, in the Fire module, each Fire module takes the feature map output from the previous feature extraction stage as input and performs the following operations sequentially: First, it passes through a Squeeze layer, which uses 16 1×1 convolutional kernels for convolution operations, followed by a ReLU activation function to compress the number of channels; then, the compressed feature map is fed into an Expand layer, which performs two convolution operations in parallel, one using 64 1×1 convolutional kernels and the other using 64 3×3 convolutional kernels (padded with 1s to maintain the spatial size), and both convolution operations are followed by a ReLU activation function; finally, the output feature maps of these two parallel convolutional paths are concatenated along the channel dimension, so that the final output channel number of each Fire module is 128.

[0034] The fully connected layer flattens the feature map (14×14×256) obtained from the pooling in the fourth feature extraction stage into a one-dimensional vector and feeds it into a fully connected layer containing an output neuron, followed by a sigmoid activation function, to finally obtain the user's age prediction value. ; ; Where age is the predicted age of the user. The output scalar value is the sigmoid value of the fully connected layer, and Sigmoid() is the Sigmoid activation function. For flattened one-dimensional vectors, Represents the weight matrix. The bias term is S, which is a preset scaling factor used to map the output in the (0,1) interval to the target age range (e.g., 0 to 100 years old).

[0035] It should be noted that the preset scaling factor S is a key model parameter used to linearly map the abstract probability values ​​(range 0, 1) output by the sigmoid activation function of the fully connected layer to an age range with practical physical meaning. The value of this factor is preset during the model design phase based on the target application scenario. For example, if the desired age prediction value output by the model is in the range of 0 to 100 years, then the scaling factor S is set to 100, ensuring that the final age prediction value... The model output is constrained to this range. The specific value of factor S determines the dimensions and upper limit of the model output value. Its setting must match the actual distribution range of age labels in the training dataset to ensure the rationality and interpretability of the model output.

[0036] The loss function of the improved VGG-Face model is: ; Where Loss is the value of the loss function. The number of samples in a training batch. For sample index, Indicates the first The true age of each training sample Indicates the first The predicted age of each training sample.

[0037] Step S2 inputs the preprocessed LCD panel user image output from step S1 into a deep convolutional neural network based on an improved VGG-Face model for forward inference. This network significantly reduces the number of parameters while maintaining strong feature extraction capabilities by embedding a Fire module and introducing residual connections in the traditional VGG-Face model. Finally, the network maps the extracted high-dimensional features to a scalar through a fully connected layer, and after processing by a Sigmoid activation function and a preset scaling factor, outputs a predicted value representing the user's age.

[0038] It's important to note that the improved VGG-Face model is used to automatically estimate the user's age, rather than relying on manual input. This is primarily based on a comprehensive consideration of automation, privacy protection, and compatibility with terminal hardware. The improved VGG-Face model, by embedding a Fire module and residual connections, significantly reduces computational complexity and the number of parameters while maintaining high-precision age recognition capabilities. This allows it to run efficiently on the local processor integrated into the LCD panel, achieving seamless, real-time user status awareness. This approach avoids the cumbersome process of requiring users to manually input and update age information, ensuring a seamless experience for the intelligent adjustment system. Simultaneously, all image data is processed locally and immediately deleted, never storing or transmitting original facial information, eliminating the risk of personal biometric data leakage at the source—a more secure approach than manual input followed by age profile management. Furthermore, automatic recognition can dynamically adapt to the usage scenarios of multiple users of different ages, providing real-time and reliable physiological status input for subsequent personalized display adjustments based on fuzzy logic and gradient boosting trees.

[0039] Step S3: Collect LCD panel environmental data. Based on the LCD panel environmental data and the user's age prediction value, adjust the display of the LCD panel using fuzzy logic control to obtain the first display adjustment result. Step S3 collects environmental data of the LCD panel, including ambient illuminance measured by the ambient light sensor. (Unit: lux), local time value obtained from the system clock. (hours are represented by integers from 0 to 23), and the geographic location factor L.

[0040] It should be noted that the value of the geographic location factor L is based on an inductive analysis of ambient light conditions and visual health needs under different typical usage scenarios. Essentially, the geographic location factor L is an adjustment coefficient; its value is positively correlated with the recommended blue light filtering intensity for the corresponding scenario. That is, the higher the value, the more intense the blue light filtering should be applied in that scenario. The geographic location factor L for outdoor scenarios is set at 0.95 because outdoor environments typically have ample natural light and strong ambient light; appropriately reducing the blue light filtering intensity helps maintain screen visibility and color performance under strong light. The L value for office scenarios is set to a baseline of 1.0, as it represents a typical indoor working environment with stable, bright artificial lighting. The L value for home scenarios is set to the highest value of 1.1, considering that users in home environments typically engage in long, relaxing viewing activities, with a higher proportion of nighttime use; therefore, the strongest blue light filtering is required to minimize potential interference with visual comfort and circadian rhythms. These typical values ​​are preset empirical values ​​and can be calibrated in actual products based on display panel characteristics and the preferences of the target user group.

[0041] First, the environmental data of the LCD panel and the predicted user age (age) are fuzzified: For the ambient illuminance E (with a universe of discourse of [0, 1000] lux), a set of linguistic variables is defined as {dark, moderate, bright} and its corresponding triangular membership function. For the predicted user age (with a universe of discourse of [0, 100] years), a set of linguistic variables is defined as {childhood, youth, middle age, old age} and its corresponding membership function. For the local time T (with a universe of discourse of [0, 24) hours), a set of linguistic variables is defined as {night, daytime, evening} and its corresponding membership function. For the geographic location factor L, a set of linguistic variables is defined as {outdoor, office, home}. Since the geographic location factor L is discretely assigned, the values ​​of this linguistic variable (outdoor, office, home) are mutually exclusive and explicit categories. Its membership function is defined as: if the geographic location factor... ,but and If geographical location factor ,but and If geographical location factor ,but and ,in , , These are the membership function values ​​of the geographic location factor for outdoor, office, and home, respectively.

[0042] It should be noted that the domain of discourse for the predicted user age is defined as 0 to 100 years old. Based on this, four fuzzy subsets – “Childhood,” “Youth,” “Middle Age,” and “Old Age” – are quantified using triangular membership functions. Specifically, the center point of the membership function for the “Childhood” subset is set at approximately 12 years old, with its effective support range covering 0 to 25 years old; the center point for the “Youth” subset is set at approximately 30 years old, with its support range covering 18 to 45 years old; the center point for the “Middle Age” subset is set at approximately 50 years old, with its support range covering 35 to 65 years old; and the center point for the “Old Age” subset is set at approximately 75 years old, with its support range covering 55 to 100 years old. The membership functions of adjacent subsets overlap at the boundaries of their support ranges. For example, around age 25, the membership degree of “Childhood” gradually decreases to 0 while the membership degree of “Youth” gradually increases to 1, thus achieving a smooth transition and fuzzy mapping between different age stages.

[0043] For local time T, its universe of discourse is defined as 0:00 to 23:00 (a complete date cycle). The fuzzy subsets "Night," "Day," and "Evening" are also divided using triangular membership functions. The membership function center point of the "Night" subset is set at approximately 3:00, with its main support interval covering 22:00 to 8:00 the next day, representing typical sleep and low-light periods. The center point of the "Day" subset is set at approximately 13:00, with its support interval covering 8:00 to 18:00, representing work and activity periods with high ambient light. The center point of the "Evening" subset is set at approximately 19:00, with its support interval covering 16:00 to 23:00, representing the transition period with diminishing sunlight and predominantly artificial lighting. The membership functions of these subsets also overlap at the boundaries of their support intervals (such as near 8:00, 18:00, and 22:00) to ensure smooth and continuous changes in control output at time boundaries, avoiding abrupt changes.

[0044] For ambient illuminance E (with a universe of discourse of [0, 1000] lux), its fuzzy subsets "Dark," "Moderate," and "Bright" are also quantized using triangular membership functions. Specifically, the center point of the membership function for the "Dark" subset is set at approximately 100 lux, with an effective support range covering 0 to 300 lux, representing a dimly lit environment with insufficient illumination; the center point for the "Moderate" subset is set at approximately 400 lux, with a support range covering 200 to 600 lux, representing a typical artificial lighting or soft natural light environment; and the center point for the "Bright" subset is set at approximately 800 lux, with a support range covering 500 to 1000 lux, representing a brightly lit environment with sufficient light. The membership functions of adjacent subsets overlap at the boundaries of their support ranges (e.g., near 200-300 lux and 500-600 lux) to achieve a smooth transition of illuminance intensity between fuzzy subsets and ensure the continuity of control output.

[0045] The collected LCD panel environmental data and the user's predicted age value (age) are substituted into their respective membership functions to obtain the corresponding membership vectors, thus completing the fuzzification transformation of all input variables.

[0046] Fuzzy inference is then performed, with the inference process based on a pre-defined fuzzy rule base. (Include (The first fuzzy rule), Rules The form is: if yes And age is and yes and yes So, blue light modulation amount yes ,in , , , These are fuzzy subsets of ambient light intensity, predicted user age, local time, and geographic location factors, respectively. This is a fuzzy subset of the blue light modulation amount. Blue light modulation amount The set of language variables is defined as {significantly reduced, appropriately reduced, maintained, appropriately enhanced, significantly enhanced}, with corresponding membership functions as follows: , , , , Its domain The percentage change in blue light intensity adjustment ranges from -50% to 50%. The membership function parameters for the blue light intensity adjustment are as follows: Significant reduction: support range [-50%, -10%], center point -30%; Moderate reduction: support range [-30%, 10%], center point -10%; Maintain: support range [-10%, 10%], center point 0%; Moderate enhancement: support range [-10%, 30%], center point 10%; Significant enhancement: support range [10%, 50%], center point 30%.

[0047] It should be noted that the fuzzy rule base R is manually defined and arranged based on prior knowledge and expert experience in the fields of display comfort and visual health. Each rule adopts an "if-then" form, where the "if" part (antecedent) is composed of fuzzy subsets corresponding to the four input variables: ambient light intensity E, predicted user age (age), local time T, and geographic location factor L, connected by a logical "AND". The "then" part (consequence) specifies the target fuzzy subset corresponding to the blue light adjustment amount U. For example, a typical rule can be expressed as: "If the ambient light intensity is 'dark', the user age is 'elderly', the local time is 'night', and the geographic location factor is 'home', then the blue light adjustment amount is 'significantly reduced'." The rule base is designed to systematically cover various typical usage scenarios. The total number of rules N is much smaller than the total number of permutations and combinations of the fuzzy subsets of the input variables. It is a refined, complete, and conflict-free set of empirical rules formed by merging similar scenarios and eliminating contradictory or invalid combinations.

[0048] Based on the current input LCD panel environmental data and the user's predicted age. Calculate the first Rules activation intensity And through the fuzzy subset of the conclusion of the activation intensity clipping rule. This yields the activated output fuzzy set; The activation output fuzzy sets corresponding to each rule are aggregated to obtain the aggregated membership function value. .

[0049] Finally, defuzzification is performed, and the centroid method is used to aggregate the membership function values. Converted to precise blue light adjustment. The calculation formula is: ; in, The first displayed adjustment result is shown, where u is the blue light adjustment variable. This represents the aggregate membership function value.

[0050] Blue light regulation This is the first display adjustment result. In practical applications, the driving current of the blue LED in the backlight module of the LCD panel is adjusted by this percentage value, or the gain of the blue channel in the display color lookup table is adjusted to achieve precise adjustment of the blue light intensity output by the LCD panel.

[0051] Step S4: Collect user feedback data, use the user feedback data as a training set to train the gradient boosting tree model, obtain a trained gradient boosting tree model, input the first display adjustment result into the trained gradient boosting tree model, and output the second display adjustment result, thereby realizing intelligent display control of the LCD panel.

[0052] Step S4 receives the first display adjustment result from step S3. The system also includes the precise input data upon which this result is generated, such as ambient light intensity E, predicted user age (age), local time T (in hours), and geographic location factor L. Simultaneously, the system collects user feedback data, obtained through the user interface; specifically, the system provides interactive elements that allow users to manually fine-tune the current display effect, and records the target adjustment value set by the user when adjustments are made. The feedback value f is defined as the deviation between the user's target value and the first adjustment result of the system's current application, i.e., f = - .

[0053] The feedback value of each feedback event, together with the corresponding LCD panel environmental data and the user's age prediction value, constitutes a training sample, represented as an input feature vector. and its corresponding target label feedback value Maintain a local storage capacity of Let D be a first-in-first-out (FIFO) data buffer used to store the M most recently collected samples, j=1,2,...,M. Then the j-th sample is... M≤ .

[0054] The gradient boosting tree model is trained using all samples in buffer D. This model is an additive model, and its final predicted output is... K decision trees The combination is given as: ,in The learning rate controls the contribution of each tree to the final result. Let x represent the prediction value of the k-th decision tree given the input feature vector x, where K is the number of decision trees.

[0055] It should be noted that step S4 uses the gradient boosting tree model primarily because of its unique advantages in handling small-scale, structured feedback data in this scenario. This model can automatically learn the complex nonlinear relationships and interactions between environmental data, user attributes, and personal preferences without requiring manual feature design. Compared to deep neural networks, it is less prone to overfitting with small datasets, and its results offer better interpretability (e.g., feature importance), facilitating the analysis of the impact of different factors on personal preferences. Furthermore, its high training and inference efficiency makes it suitable for localized incremental learning on terminal devices. It can achieve continuous and robust personalized calibration of general fuzzy rules using sparse user feedback data while protecting user privacy, ultimately achieving an optimal balance between control precision, generalization ability, and computational cost.

[0056] Each decision tree The input feature vector x is mapped to a leaf node by recursively partitioning the feature space. This leaf node contains a fixed predicted value. The model is trained to minimize a regularized loss function over all training samples. For the objective, the loss function is defined as: ; in, It is the loss function of the gradient boosting tree model. This represents the true target adjustment value of the j-th training sample. This represents the predicted target adjustment value for the j-th training sample. express and The mean squared error is given by K, where K is the number of decision trees and k is the index of the decision tree. This is a regularization term for the complexity of the k-th decision tree, used to control model complexity and prevent overfitting. = * + * * , Let be the penalty coefficient for leaf nodes, and Leaf represent the total number of leaf nodes in the k-th decision tree. The L2 regularization coefficient for the leaf weights. This represents the output weight of the leaf node in the k-th decision tree.

[0057] The training uses a forward stepwise algorithm, and a new decision tree is built in the k-th iteration. To fit the current model The negative gradient on the training samples, and then... Add to the model, i.e. The training process continues until the preset number of trees is reached. Alternatively, the training can stop after the loss converges, resulting in a well-trained model. ().

[0058] After model training or updates are completed, when new display adjustment instructions need to be generated, step S4 receives the latest first display adjustment result from step S3. and its input feature vector . This vector Input to the trained gradient boosting tree model The model outputs a predicted user personalized preference offset for the current context. , The feedback prediction value. The final second display shows the adjustment result. It is obtained by adding the first displayed adjustment result to the feedback prediction value, that is This value This refers to the personalized blue light adjustment amount, which is sent to the driving circuit of the LCD panel for execution. Periodically, or after a certain amount of data has accumulated in the buffer, the entire buffer is used for retraining to achieve continuous adaptive learning. All data processing and model training are performed locally on the device. User's original feedback data is removed from temporary storage after being added to the buffer, and the buffer D is updated according to the first-in, first-out principle, without permanently retaining user data.

[0059] Furthermore, based on the above method embodiments, the present invention also provides a control system, including a memory, a processor, and a computer program stored in the memory, which is adapted to be loaded and executed by the processor to implement the above-described intelligent display control method for driving image recognition of a liquid crystal panel.

[0060] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, the phrase "comprising an element defined as..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0061] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A smart display control method for driving image recognition on a liquid crystal panel, characterized in that: Includes the following steps: Step S1: Acquire user image data of the LCD panel through a camera and perform preprocessing to obtain the user image of the LCD panel; Step S2: Input the user image of the liquid crystal panel into the improved VGG-Face model and output the predicted user age value; Step S3: Collect LCD panel environmental data. Based on the LCD panel environmental data and the user's age prediction value, adjust the display of the LCD panel using fuzzy logic control to obtain the first display adjustment result. Step S4: Collect user feedback data, use the user feedback data as a training set to train the gradient boosting tree model, obtain a trained gradient boosting tree model, input the first display adjustment result into the trained gradient boosting tree model, and output the second display adjustment result, thereby realizing intelligent display control of the LCD panel.

2. The intelligent display control method for image recognition driving of a liquid crystal panel according to claim 1, characterized in that: The process of acquiring user image data from the LCD panel via a camera and preprocessing it to obtain a user image of the LCD panel includes the following specific steps: The camera captures the original image of the user's face at a preset sampling frequency; the user's facial region is detected and located in the original image; the image of the user's facial region is standardized and enhanced to finally obtain the user image on the LCD panel.

3. The intelligent display control method for image recognition driving of a liquid crystal panel according to claim 2, characterized in that: The process of inputting the user image from the liquid crystal panel into the improved VGG-Face model and outputting a predicted user age value includes the following steps: Construct an improved VGG-Face model; input the user image of the liquid crystal panel into the improved VGG-Face model: ； Where age is the predicted age of the user. The output scalar value is the sigmoid value of the fully connected layer, and Sigmoid() is the Sigmoid activation function. S is a flattened one-dimensional vector, and S is a preset scaling factor.

4. The intelligent display control method for image recognition driving of a liquid crystal panel according to claim 3, characterized in that: The construction of the improved VGG-Face model includes the following steps: The improved VGG-Face model includes: an input convolutional layer, four feature extraction stages, and a fully connected layer; The input convolutional layer contains 64 convolutional kernels of size 3×3, with a stride of 1 and padding of 1; The four feature extraction stages are as follows: the first feature extraction stage includes two Fire modules and a pooling layer; the second feature extraction stage includes four Fire modules, residual connections, and a pooling layer; the third feature extraction stage includes eight Fire modules, residual connections, and a pooling layer; and the fourth feature extraction stage includes four Fire modules, residual connections, and a pooling layer. The fully connected layer contains one output neuron.

5. The intelligent display control method for image recognition driving of a liquid crystal panel according to claim 4, characterized in that: The collection of LCD panel environmental data includes the following steps: LCD panel environmental data includes ambient light intensity measured by an ambient light sensor. Local time value And the geographic location factor L.

6. The intelligent display control method for driving image recognition of a liquid crystal panel according to claim 5, characterized in that: The process of adjusting the display of the LCD panel based on the LCD panel's environmental data and the user's predicted age, using fuzzy logic control to obtain a first display adjustment result, includes the following steps: Construct a membership function between LCD panel environmental data and predicted user age values; Substitute the LCD panel environmental data and the user age prediction value into the corresponding membership function to obtain the membership vector of the LCD panel environmental data and the membership vector of the user age prediction value. The aggregate membership function value is calculated based on the membership vector of the LCD panel environmental data, the membership vector of the user age prediction value, and the preset fuzzy rule base. The first display adjustment result is calculated based on the aggregate membership function value.

7. The intelligent display control method for driving image recognition of a liquid crystal panel according to claim 6, characterized in that: The calculation of the first display adjustment result based on the aggregate membership function value includes the following specific steps: The centroid method is used to convert the aggregate membership function values ​​into the first display adjustment result. The calculation formula is as follows: ； in, The first displayed adjustment result is shown, where u is the blue light adjustment variable. This represents the aggregate membership function value.

8. The intelligent display control method for image recognition driving of a liquid crystal panel according to claim 7, characterized in that: The process of collecting user feedback data and using this data as a training set to train the gradient boosting tree model to obtain a trained gradient boosting tree model is as follows: Collect user feedback data, wherein the user feedback data is the user's feedback value f, f= - , Indicates the target adjustment value; The feedback values, along with the corresponding LCD panel environmental data and the predicted user age, form a training set. The forward stepwise algorithm is used to train the gradient boosting tree model using the training set, resulting in a trained gradient boosting tree model.

9. The intelligent display control method for image recognition driving of a liquid crystal panel according to claim 8, characterized in that: The process of inputting the first display adjustment result into the trained gradient boosting tree model and outputting the second display adjustment result is as follows: The initial adjustment result is then input into the trained gradient boosting tree model: ； in, This represents the feedback prediction value. () represents a gradient boosting tree model. This represents the input feature vector; The second display adjustment result is obtained by adding the first display adjustment result to the feedback prediction value: ； in, The second display shows the adjustment results.

10. A smart display control system for driving image recognition on a liquid crystal panel, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1-9.

Images