Dynamic backlight control method and system for high-performance liquid crystal display module

Abstract

The present application relates to the technical field of liquid crystal display, specifically to a dynamic backlight control method and system of high-efficiency liquid crystal display module, comprising the following steps: S1, collecting image data, human eye fixation point distribution data and preference information for image brightness and color combination; S2, using the collected data to train a CNN-RNN hybrid model; S3, based on the sensitivity of human eye to brightness change, introducing a JND threshold, and designing a brightness weighting function; S4, weighting and averaging the image brightness according to the CIE 1931 brightness curve; S5, using a Sobel operator to detect the brightness mutation area; the present application effectively reduces the influence of brightness mutation on the visual system by introducing a JND threshold and designing a reasonable brightness weighting function and dynamic compensation strategy, thereby reducing visual fatigue; the Sobel operator is used to detect the brightness mutation area, and a dynamic compensation coefficient is applied to finely adjust the pixel value, realizing pixel-level brightness optimization, thereby maintaining high contrast and delicate image quality, and reducing energy consumption.

Description

Technical Field

[0001] This invention relates to the field of liquid crystal display technology, specifically to a dynamic backlight control method and system for high-performance liquid crystal display modules. Background Technology

[0002] With the rapid development of display technology, liquid crystal display (LCD) modules have become a key display technology with wide application. As an electro-optical display device, LCD modules have advantages such as wide viewing angle and low power consumption. However, there is still room for improvement in their picture performance in the high and low brightness ranges, especially in dynamic backlight control. The core of dynamic backlight control is to adjust the backlight brightness in real time according to the brightness distribution of the picture and the environmental conditions. For example, in dark scenes, the backlight is reduced to enhance the details in the dark areas, and in bright scenes, the backlight is increased to optimize the contrast. This adjustment not only improves the picture quality, but also effectively reduces visual fatigue.

[0003] Because the human eye has different sensitivities to different colors, traditional fixed backlight systems cannot flexibly adjust the backlight brightness according to the changes in brightness of the animated image and the human eye's color sensitivity, resulting in increased energy consumption and a poor visual experience. Summary of the Invention

[0004] The purpose of this invention is to provide a dynamic backlight control method and system for high-performance liquid crystal display modules to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a dynamic backlight control method for a high-performance liquid crystal display module, comprising the following steps:

[0006] S1. Image data is acquired through the image data acquisition module, and the YOLO algorithm is used to introduce deeper network layers for real-time object detection; human eye perception preference data acquisition module is used to acquire human eye gaze point distribution data and preference information for image brightness and color combination.

[0007] S2. The model training module uses the collected data to train a CNN-RNN hybrid model to predict human visual perception preferences.

[0008] S3. The brightness weighting module, based on the human eye's sensitivity to brightness changes, introduces a JND threshold and designs a brightness weighting function to weight the image brightness. The formula is:

[0009]

[0010] Where L is the pixel brightness value, and W(L) is the brightness weighting coefficient;

[0011] S4. The color space conversion and brightness weighted averaging module performs a weighted average of the image brightness based on the CIE 1931 brightness curve;

[0012] S5, the brightness change detection and dynamic compensation module uses the Sobel operator to detect brightness change areas and adjusts the pixel values ​​of the change areas according to the dynamic compensation coefficient;

[0013] First, calculate the dynamic compensation coefficient:

[0014]

[0015] Where ΔL is the local brightness difference, representing the brightness difference between the abrupt change region and the surrounding region; L avg α is the local average brightness value, representing the average brightness of the abrupt change region and the surrounding region; α is the compensation intensity coefficient, which controls the strength of the compensation.

[0016] Then, adjust the pixel values ​​in the abrupt change region according to the dynamic compensation coefficient:

[0017] L compensated =L original ·K comp

[0018] Among them, L compensated Adjusted pixel value, L original K represents the original pixel value. comp For dynamic compensation coefficients;

[0019] Pixel-level brightness optimization is achieved by adjusting pixel values ​​in abrupt change regions based on a dynamic compensation coefficient.

[0020] Furthermore, in S2, a CNN-RNN hybrid model is trained using a dynamic learning rate adjustment strategy, and the training process is monitored using tools such as TensorBoard, including the following steps:

[0021] S21. Use the isnull() function to check for missing or outlier values ​​in image data and human eye perception preference data. In image data, missing values ​​are usually manifested as the absence of specific pixel values. The missing pixel value is inferred by interpolating the existing pixel values ​​using linear interpolation. Outlier values ​​are detected using the statistical method Z-score and replaced with the median. Ensure that the human eye gaze point distribution data is aligned with the image data in time and space.

[0022] S22. Normalize the image data by scaling its pixel values ​​to the range of [0,1] or [-1,1] to accelerate the training speed of the neural network; standardize the human eye perception preference data, such as gaze point distribution, brightness preference, and color preference, so that its mean is 0 and its variance is 1.

[0023] S23. Divide the dataset into training set, validation set and test set, with a ratio of 70% training set, 15% validation set and 15% test set.

[0024] S24. Use the deep learning framework TensorFlow to build a hybrid RNN-CNN-RNN model with a convolutional neural network (CNN) as the main structure and a recurrent neural network (RNN) as the main structure. The CNN is responsible for extracting spatial features from video frames, while the RNN is responsible for processing the changes of these features in the time series. The combination of the two enables the model to process the spatial features and temporal dependencies of the video at the same time, thereby improving the performance of the model.

[0025] S25. Use the model.fit() method to train the model, input the training data into the RNN-CNN-RNN hybrid model, perform forward and backward propagation, and update the network parameters. During the model training process, adopt a dynamic learning rate adjustment strategy, such as exponential decay, to optimize the training effect. At the same time, use tools such as TensorBoard to monitor the training process of the model, observe the changes in the loss function and evaluation metrics, and ensure the stability and effectiveness of the model training.

[0026] S26. After each round of training, use the model.evaluate() method to evaluate the model's performance on the test dataset, and monitor the loss function value and accuracy metrics.

[0027] S27. Input the test set data into the trained neural network, calculate the loss function value, accuracy, precision, recall, and F1 score on the test set to evaluate the model's predictive ability and generalization performance.

[0028] A CNN-RNN hybrid model was trained that can accurately predict human visual perception preferences, providing strong support for dynamic backlight control of high-performance liquid crystal display modules.

[0029] Furthermore, in step S4, the image brightness is weighted and averaged according to the CIE 1931 brightness curve. The specific steps are as follows:

[0030] S41. Normalize the RGB values ​​of the image pixels to the range of 0-1:

[0031] S42. Perform Gamma correction on the normalized RGB values ​​to compensate for the non-linear response of the display device;

[0032] S43. Use a transformation matrix to convert the Gamma-corrected RGB values ​​into tristimulus values ​​in the XYZ color space;

[0033] S44. Extract the Y component as the brightness value, and perform a weighted average based on the CIE 1931 brightness curve to obtain the weighted average brightness value of the entire image.

[0034] S45. Normalize the weighted average brightness values ​​to the range of 0-1, and generate a grayscale image as the output brightness image based on the normalized brightness values.

[0035] The dynamic backlight control system for high-performance liquid crystal display modules includes:

[0036] S1, Image Data Acquisition Module, adopts the YOLO algorithm to introduce deeper network layers for real-time object detection. It directly predicts bounding boxes and class probabilities through a single forward propagation. The acquired image data provides a foundation for subsequent object detection and perceptual preference analysis.

[0037] S2, Human Eye Perception Preference Data Acquisition Module, which includes an eye-tracking device and a psychological questionnaire system. The eye-tracking device is used to collect human eye fixation point data to construct a dataset containing image stimulus-fixation point distribution and dwell time. At the same time, the psychological questionnaire is used to collect human preference information on image brightness and color combination.

[0038] S3, Model Training Module: This module uses the deep learning framework TensorFlow to build a CNN-RNN hybrid model, which combines a convolutional neural network (CNN) as the main structure with a recurrent neural network (RNN). The collected data is used to train the CNN-RNN hybrid model to predict the human eye's perceptual preferences for different images.

[0039] S4, Brightness Weighting Module, is used to design a brightness weighting function based on the human eye's sensitivity to brightness changes, introduce the JND threshold, and perform weighted processing on the image brightness to avoid overexposure.

[0040] S5, Color Space Conversion and Luminance Weighted Averaging Module, is used to convert the RGB values ​​of image pixels into tristimulus values ​​in the XYZ color space, extract the Y component as the luminance value, and perform a weighted average.

[0041] S6, the brightness change detection and dynamic compensation module, uses the Sobel operator to detect brightness change areas and adjusts the pixel values ​​of the change areas according to the dynamic compensation coefficient to achieve pixel-level brightness optimization.

[0042] Furthermore, in S1, the YOLO algorithm introduces deeper network layers by using Darknet-53 as a feature extraction network. The Darknet-53 network further extracts image features by adding convolutional layers and residual blocks. The specific steps are as follows:

[0043] S11. In the YOLO algorithm, the original feature extraction network is replaced with a deeper Darknet-53 network.

[0044] S12. In the Darknet-53 network, residual blocks pass the input directly to the output through skip connections.

[0045] S13. Use the exponential decay strategy to optimize training performance. Decrease the learning rate exponentially to optimize the training process. The formula is: lrt = lro · γ (t / T)

[0046] Where lrt is the learning rate of the t-th iteration, lro is the initial learning rate, γ is the decay rate, which controls the speed at which the learning rate decreases, and T is the decay period, which means that the learning rate decays once every T iterations.

[0047] S14. Train the network thoroughly using a large-scale dataset;

[0048] S15. After training, the model is evaluated and tuned by calculating metrics such as loss function value and accuracy to assess the model's performance.

[0049] S16. Deploy the trained model to the image data acquisition module.

[0050] Compared with existing technologies, the beneficial effects of this invention are as follows: By introducing a JND threshold and designing a reasonable brightness weighting function and dynamic compensation strategy, the impact of brightness abrupt changes on the visual system is effectively reduced, thereby reducing visual fatigue; the brightness weighting function can significantly reduce brightness abrupt changes and overexposure, thereby improving the overall image quality; the design of the brightness weighting function fully considers the perceptual characteristics of the human eye, avoiding the stimulation of the visual system by brightness abrupt changes and overexposure, which helps to reduce visual fatigue experienced by users when viewing the screen for a long time; under the premise of ensuring image quality, a reasonable brightness weighting strategy can reduce the energy consumption of the liquid crystal display module, especially when using a nonlinear compression strategy in high-brightness areas, which can significantly reduce the power consumption of the backlight module, thereby improving the energy efficiency of the entire display system; by introducing a JND threshold, the degree of adjustment of backlight brightness can be more accurately controlled, thereby optimizing image quality, especially when processing complex scenes and dynamic images, the application of the JND threshold can significantly improve the detail and contrast of the image; and by detecting brightness abrupt change areas through the Sobel operator and applying dynamic compensation coefficients to finely adjust pixel values, pixel-level brightness optimization is achieved, thereby maintaining high contrast and delicate image quality, reducing energy consumption, and significantly improving the user's visual experience. Attached Figure Description

[0051] Figure 1 This is a flowchart of the method of the present invention;

[0052] Figure 2 This is a flowchart of the training process for the CNN-RNN hybrid model of the present invention;

[0053] Figure 3 This is a flowchart of the brightness-weighted average process of the present invention. Detailed Implementation

[0054] The technical solution of the present invention will now be described in detail through specific embodiments.

[0055] Please see Figure 1-3 This invention provides a dynamic backlight control method for a high-performance liquid crystal display module, comprising the following steps:

[0056] S1. Image data is acquired through the image data acquisition module, and the YOLO algorithm is used to introduce deeper network layers for real-time object detection; human eye perception preference data acquisition module is used to acquire human eye gaze point distribution data and preference information for image brightness and color combination; the data is used for subsequent neural network training to predict human eye perception preferences for different images.

[0057] S2. The model training module uses the collected data to train a CNN-RNN hybrid model to predict human visual perception preferences. The model training employs a dynamic learning rate adjustment strategy and uses tools such as TensorBoard to monitor the training process, including the following steps:

[0058] S21. Use the isnull() function to check for missing or outlier values ​​in image data and human eye perception preference data. In image data, missing values ​​are usually manifested as the absence of specific pixel values. Therefore, use linear interpolation to interpolate existing pixel values ​​to infer the value of the missing pixel. Use the statistical method Z-score to detect outliers and replace them with the median. Ensure that the human eye gaze point distribution data is aligned with the image data in time and space.

[0059] S22. Normalize the image data by scaling its pixel values ​​to the range of [0,1] or [-1,1] to accelerate the training speed of the neural network; standardize the human eye perception preference data, such as gaze point distribution, brightness preference, and color preference, so that its mean is 0 and its variance is 1.

[0060] S23. Divide the dataset into training set, validation set and test set, with a ratio of 70% training set, 15% validation set and 15% test set.

[0061] S24. Use the deep learning framework TensorFlow to build a hybrid RNN-CNN-RNN model with a convolutional neural network (CNN) as the main structure and a recurrent neural network (RNN) as the main structure. The CNN is responsible for extracting spatial features from video frames, while the RNN is responsible for processing the changes of these features in the time series. The combination of the two enables the model to process the spatial features and temporal dependencies of the video at the same time, thereby improving the performance of the model.

[0062] S25. Use the model.fit() method to train the model, input the training data into the RNN-CNN-RNN hybrid model, perform forward and backward propagation, and update the network parameters. During the model training process, adopt a dynamic learning rate adjustment strategy, such as exponential decay, to optimize the training effect. At the same time, use tools such as TensorBoard to monitor the training process of the model, observe the changes in the loss function and evaluation metrics, and ensure the stability and effectiveness of the model training.

[0063] S26. After each round of training, use the model.evaluate() method to evaluate the model's performance on the test dataset, and monitor the loss function value and accuracy metrics.

[0064] S27. Input the test set data into the trained neural network, calculate the loss function value, accuracy, precision, recall, and F1 score on the test set to evaluate the model's predictive ability and generalization performance.

[0065] A CNN-RNN hybrid model that can accurately predict human visual perception preferences was trained, providing strong support for dynamic backlight control of high-performance liquid crystal display modules.

[0066] S3. The brightness weighting module, based on the human eye's sensitivity to brightness changes, introduces a JND threshold and designs a brightness weighting function to weight the image brightness. This process applies weights to areas with low brightness (<50 cd / m²). 2 Linear weighting is used in the bright areas (>200 cd / m²). 2 To avoid overexposure, nonlinear compression is introduced. The formula is:

[0067]

[0068] Where L is the pixel brightness value, and W(L) is the brightness weighting coefficient;

[0069] By introducing a JND threshold and designing a reasonable brightness weighting function and dynamic compensation strategy, the impact of brightness abrupt changes on the visual system is effectively reduced, thereby reducing visual fatigue.

[0070] S4. The image brightness is weighted and averaged according to the CIE 1931 brightness curve using the color space conversion and brightness weighted averaging module. The specific steps are as follows:

[0071] S41. Normalize the RGB values ​​of image pixels from the range of 0-255 to the range of 0-1:

[0072]

[0073] S42. Perform Gamma correction on the normalized RGB values, typically using Gamma = 2.2, to compensate for the non-linear response of the display device:

[0074]

[0075] S43. Use a transformation matrix to convert the gamma-corrected RGB values ​​to tristimulus values ​​in the XYZ color space:

[0076]

[0077] S44. Extract the Y component as the luminance value, and perform a weighted average based on the CIE 1931 luminance curve to obtain the weighted average luminance value of the entire image:

[0078]

[0079] Where F is the total number of pixels in the image, Y i Let Y be the Y component value of the i-th pixel;

[0080] S45. Normalize the weighted average brightness value to the range of 0-1 to provide a unified brightness scale for subsequent dynamic backlight control:

[0081]

[0082] Where Ymax is the maximum value of the Y component in the XYZ color space, L avg It is a brightness-weighted average;

[0083] Based on the normalized brightness values, a grayscale image is generated as the output brightness image;

[0084] It achieves accurate weighted averaging of image brightness and spatiotemporal correlation matching between backlight brightness and image content, while taking into account the differences in human eye sensitivity to different colors.

[0085] S5, the brightness change detection and dynamic compensation module uses the Sobel operator to detect brightness change areas and adjusts the pixel values ​​of the change areas according to the dynamic compensation coefficient;

[0086] First, calculate the dynamic compensation coefficient:

[0087]

[0088] Where ΔL is the local brightness difference, representing the brightness difference between the abrupt change region and the surrounding region; L avg α is the local average brightness value, representing the average brightness of the abrupt change region and the surrounding region; α is the compensation intensity coefficient, which controls the strength of the compensation.

[0089] The value of α is usually between 0.1 and 0.3. When α is larger, the compensation effect is more obvious, but it may lead to overcompensation; when α is smaller, the compensation effect is more gradual.

[0090] ΔL is calculated using a sliding window to determine the difference between the abrupt change region and the average brightness of its neighborhood.

[0091] L avg Gaussian weighting is used to calculate the average brightness within the window, thus enhancing the weight of the center pixel.

[0092] The dynamic compensation coefficient reflects the degree of backlight adjustment and is used for subsequent pixel-level brightness optimization;

[0093] Fine-tuning of pixel values ​​in varying regions ensures smooth image transitions.

[0094] Then, adjust the pixel values ​​in the abrupt change region according to the dynamic compensation coefficient:

[0095] L compensated =L original ·K comp

[0096] Among them, L compensated Adjusted pixel value, L original K represents the original pixel value. comp For dynamic compensation coefficients;

[0097] The pixel values ​​in the abrupt change area are adjusted according to the dynamic compensation coefficient to achieve pixel-level brightness optimization, maintaining high contrast and delicate image quality.

[0098] The dynamic backlight control system for high-performance liquid crystal display modules includes:

[0099] S1, Image Data Acquisition Module: This module employs the YOLO algorithm with deeper network layers for real-time object detection. It directly predicts bounding boxes and class probabilities through a single forward propagation, providing a foundation for subsequent object detection and perceptual preference analysis. The deeper network layers are introduced into the YOLO algorithm using Darknet-53 as the feature extraction network. Darknet-53 further extracts image features by adding convolutional layers and residual blocks, improving object detection performance. The specific steps are as follows:

[0100] S11. In the YOLO algorithm, the original feature extraction network is replaced with a deeper Darknet-53 network. Darknet-53 contains 53 convolutional layers and 5 max-pooling layers. Its structure can be divided into a front section, a middle section, and a back section. The front section mainly consists of convolutional layers and max-pooling layers, used to extract basic image features; the middle section mainly consists of residual blocks, used to enhance feature expression and propagation; the back section mainly consists of global average pooling layers and fully connected layers, used for classification and regression tasks.

[0101] S12. In the Darknet-53 network, residual blocks pass the input directly to the output through skip connections, which helps to alleviate the gradient vanishing problem in deep neural networks and improves the training efficiency and performance of the network. When integrating Darknet-53 into the YOLO algorithm, it is necessary to ensure the correct integration of residual blocks to maintain the stability and performance of the network.

[0102] S13. Use the exponential decay strategy to optimize training performance. Decrease the learning rate exponentially to optimize the training process. The formula is: lrt = lro · γ (t / T)

[0103] Where lrt is the learning rate of the t-th iteration, lro is the initial learning rate, γ is the decay rate, which controls the speed at which the learning rate decreases, and T is the decay period, which means that the learning rate decays once every T iterations.

[0104] The exponential decay strategy is used to optimize the training effect, so that the model continues to maintain a good convergence effect in the later stage of training.

[0105] S14. Train the network thoroughly using a large-scale dataset and iteratively optimize the network parameters so that the network can learn more complex features, thereby improving the accuracy and robustness of object detection.

[0106] S15. After training, the model is evaluated and tuned by calculating metrics such as loss function value and accuracy to assess the model's performance.

[0107] S16. Deploy the trained model to the image data acquisition module;

[0108] The YOLO algorithm incorporates the Darknet-53 network to improve its feature extraction capabilities and detection accuracy. When detecting complex backgrounds or scenes with significant lighting variations, the Darknet-53 network can more robustly identify target objects through multi-layer nonlinear transformations. When detecting occluded or overlapping objects, the Darknet-53 network can more accurately locate target boundaries through multi-layer feature fusion. It can detect scene changes in images in real time and dynamically adjust the backlight control strategy based on these changes. The introduction of the Darknet-53 network into the YOLO algorithm enables the backlight control system to adapt to different scenes and lighting conditions, providing optimal display effects and user experience.

[0109] S2, Human Eye Perception Preference Data Acquisition Module, which includes an eye-tracking device and a psychological questionnaire system. The eye-tracking device is used to collect human eye fixation point data to construct a dataset containing image stimulus-fixation point distribution and dwell time. At the same time, the psychological questionnaire is used to collect human preference information on image brightness and color combination.

[0110] S3, Model Training Module: This module uses the deep learning framework TensorFlow to build a CNN-RNN hybrid model, which combines a convolutional neural network (CNN) as the main structure with a recurrent neural network (RNN). The collected data is used to train the CNN-RNN hybrid model to predict the human eye's perceptual preferences for different images.

[0111] S4, Brightness Weighting Module, is used to design a brightness weighting function based on the human eye's sensitivity to brightness changes, introduce the JND threshold, and perform weighted processing on image brightness to avoid overexposure and visual fatigue caused by sudden brightness changes.

[0112] S5, Color Space Conversion and Luminance Weighted Averaging Module, is used to convert the RGB values ​​of image pixels into tristimulus values ​​in the XYZ color space, extract the Y component as the luminance value, and perform a weighted average.

[0113] S6, the brightness change detection and dynamic compensation module, uses the Sobel operator to detect brightness change areas and adjusts the pixel values ​​of the change areas according to the dynamic compensation coefficient to achieve pixel-level brightness optimization.

[0114] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A dynamic backlight control method for a high-performance liquid crystal display module, characterized in that, Includes the following steps: S1. Image data is acquired through the image data acquisition module, and the YOLO algorithm is used to introduce deeper network layers for real-time object detection; human eye perception preference data acquisition module is used to acquire human eye gaze point distribution data and preference information for image brightness and color combination. S2. The model training module uses the collected data to train a CNN-RNN hybrid model to predict human visual perception preferences. S3. The brightness weighting module, based on the human eye's sensitivity to brightness changes, introduces a JND threshold and designs a brightness weighting function to weight the image brightness. The formula is: Where L is the pixel brightness value, and W(L) is the brightness weighting coefficient; S4. The color space conversion and brightness weighted averaging module performs a weighted average of the image brightness based on the CIE 1931 brightness curve; S5, the brightness change detection and dynamic compensation module uses the Sobel operator to detect brightness change areas and adjusts the pixel values ​​of the change areas according to the dynamic compensation coefficient; First, calculate the dynamic compensation coefficient: Where ΔL represents the local brightness difference; L avg α is the local average brightness value; α is the compensation intensity coefficient. Then, adjust the pixel values ​​in the abrupt change region according to the dynamic compensation coefficient: L compensated L original ·K comp Among them, L compensated Adjusted pixel value, L original K represents the original pixel value. comp For dynamic compensation coefficients; Pixel-level brightness optimization is achieved by adjusting pixel values ​​in abrupt change regions based on a dynamic compensation coefficient.

2. The dynamic backlight control method for a high-performance liquid crystal display module according to claim 1, characterized in that: In S2, a CNN-RNN hybrid model is trained using a dynamic learning rate adjustment strategy, and the training process is monitored using the TensorBoard tool, including the following steps: S21. Use the isnull() function to check for missing or outlier values ​​in image data and human visual perception preference data. S22. Normalize the image data by scaling its pixel values ​​to the range of [0,1] or [-1,1]; standardize the human visual perception preference data so that its mean is 0 and its variance is 1. S23. Divide the dataset into training set, validation set and test set, with a ratio of 70% training set, 15% validation set and 15% test set. S24. Use the deep learning framework TensorFlow to build a hybrid RNN-CNN-RNN model with a convolutional neural network (CNN) as the main structure and a recurrent neural network (RNN). S25. Use the model.fit() method to train the model, input the training data into the RNN-RNN hybrid model, perform forward and backward propagation, and update the network parameters. S26. After each round of training, use the model.evaluate() method to evaluate the model's performance on the test dataset, and monitor the loss function value and accuracy metrics. S27. Input the test set data into the trained neural network, and calculate the loss function value, accuracy, precision, recall, and F1 score on the test set to evaluate the model's predictive ability and generalization performance.

3. The dynamic backlight control method for a high-performance liquid crystal display module according to claim 1, characterized in that: In step S4, the image brightness is weighted and averaged according to the CIE 1931 brightness curve. The specific steps are as follows: S41. Normalize the RGB values ​​of the image pixels to the range of 0-1; S42. Perform Gamma correction on the normalized RGB values ​​to compensate for the non-linear response of the display device; S43. Use a transformation matrix to convert the Gamma-corrected RGB values ​​into tristimulus values ​​in the XYZ color space; S44. Extract the Y component as the brightness value, and perform a weighted average based on the CIE 1931 brightness curve to obtain the weighted average brightness value of the entire image. S45. Normalize the weighted average brightness values ​​to the range of 0-1, and generate a grayscale image as the output brightness image based on the normalized brightness values.

4. A dynamic backlight control system for a high-performance liquid crystal display module, characterized in that, include: S1, Image Data Acquisition Module, adopts the YOLO algorithm to introduce deeper network layers for real-time object detection, and directly predicts bounding boxes and class probabilities through a single forward propagation; S2, Human Eye Perception Preference Data Acquisition Module, which includes an eye-tracking device and a psychological questionnaire system. The eye-tracking device is used to collect human eye fixation point data to construct a dataset containing image stimulus-fixation point distribution and dwell time. At the same time, the psychological questionnaire is used to collect human preference information on image brightness and color combination. S3, Model Training Module: This module uses the deep learning framework TensorFlow to build a CNN-RNN hybrid model, which combines a convolutional neural network (CNN) as the main structure with a recurrent neural network (RNN). The collected data is used to train the CNN-RNN hybrid model to predict the human eye's perceptual preferences for different images. S4, Brightness Weighting Module, is used to design a brightness weighting function based on the human eye's sensitivity to brightness changes, by introducing a JND threshold, and to perform weighted processing on the image brightness. S5, Color Space Conversion and Luminance Weighted Averaging Module, is used to convert the RGB values ​​of image pixels into tristimulus values ​​in the XYZ color space, extract the Y component as the luminance value, and perform a weighted average. S6, the brightness change detection and dynamic compensation module, uses the Sobel operator to detect brightness change areas and adjusts the pixel values ​​of the change areas according to the dynamic compensation coefficient to achieve pixel-level brightness optimization.

5. The dynamic backlight control system for the high-performance liquid crystal display module according to claim 4, characterized in that: In S1, the YOLO algorithm introduces deeper network layers by using Darknet-53 as a feature extraction network. The Darknet-53 network further extracts image features by adding convolutional layers and residual blocks. The specific steps are as follows: S11. In the YOLO algorithm, the original feature extraction network is replaced with a deeper Darknet-53 network. S12. In the Darknet-53 network, residual blocks pass the input directly to the output through skip connections. S13. Use the exponential decay strategy to optimize training performance. Decrease the learning rate exponentially to optimize the training process. The formula is: lrt = lro · γ (t / T) Where lrt is the learning rate of the t-th iteration, lro is the initial learning rate, γ is the decay rate, which controls the speed at which the learning rate decreases, and T is the decay period, that is, the learning rate decays once every T iterations. S14. Train the network thoroughly using a large-scale dataset; S15. After training, the model is evaluated and tuned by calculating the loss function value and accuracy index to assess the model's performance. S16. Deploy the trained model to the image data acquisition module.

Images