A liquid crystal display module area load difference brightness compensation method

Abstract

The present application relates to a kind of liquid crystal display module area load difference's brightness compensation method, to solve the problem of existing compensation scheme response lag, low energy efficiency and weak environmental adaptability.The core scheme is: acquisition and analysis under multiple dynamic display scene area gray scale, texture and brightness time sequence characteristics, construct time sequence foresight fingerprint library, and train light phase mapping neural network with this, realize the nonlinear mapping of area characteristics to driving parameter;With model quantization and embedded coprocessor deep coupling;Then according to environmental parameters, dynamically fine-tune driving waveform, finally realize feedforward type brightness compensation in millisecond time scale.System embedded check window and online fine-tuning mechanism, effectively suppresses prediction error and adaptive optimization, significantly improves display quality, response speed and long-term stability.

Description

Technical Field

[0001] This invention relates to the field of liquid crystal display driving and image compensation control technology, and in particular to a brightness compensation method for regional load differences in a liquid crystal display module. Background Technology

[0002] In high-dynamic display scenarios (such as sports events, games, and video transitions), the abrupt changes in load between display areas (i.e., brightness, content complexity, or current consumption of different display areas) place higher demands on the display driving system for compensation and response speed. Existing mainstream LCD module brightness uniformity and compensation control technologies typically rely on methods such as area detection, load feedback, lookup table matching, and step-by-step adjustment to adjust display parameters under Mini LED backlighting or OLED self-emissive architectures. The industry mainstream approach generally employs the following workflow: First, a detection module (such as a TCON or backlight control unit) collects load change information for each area frame by frame. Then, through a preset lookup table, linear weight superposition, multi-area block verification, PWM brightness adjustment, or a closed-loop fine-tuning system based on brightness feedback, the display driving parameters are corrected, thereby achieving a certain degree of brightness compensation and area consistency control. Furthermore, some advanced solutions introduce image content analysis and multi-frame statistical methods, which improve compensation accuracy to some extent.

[0003] The existing solutions described above are suitable for conventional static or slow-moving scenarios, and can basically guarantee the brightness uniformity and energy efficiency of the display screen for most consumer display terminals. However, with the development of new technologies such as high refresh rates, high dynamic range (HDR) display modes, and local dimming, LCD display modules have exposed the following key technical bottlenecks in practical applications.

[0004] First, the traditional "detection-calculation-execution" closed-loop compensation structure exhibits significant response latency. When large-scale, cross-regional, high-frequency changes occur in the displayed content (such as explosion effects in games or sports scene transitions), the closed-loop feedback mode relies on serial processing such as sampling, data transfer, calculation, and feedback, making it difficult to complete parameter compensation in real time under millisecond-level time constraints. This leads to phenomena such as brightness lag, color flicker, or short-term surges in energy consumption in local areas, significantly impacting the user's visual experience.

[0005] Secondly, static schemes such as lookup tables and weighted overlay have limited generalization capabilities. Since these techniques are essentially based on a "known-mapping-correction" logical structure, they are prone to problems such as "lookup dimension explosion," inaccurate template matching, and lack of forward-looking adjustment strategies when dealing with high-dimensional dynamic scenes. They cannot perceive and respond to potential load trend changes in the next few consecutive frames, resulting in a mismatch between the compensation effect and the dynamic requirements of the real image content.

[0006] Furthermore, while methods such as multi-region segmentation correction and regional anchor point compensation improve the fine-grained spatial compensation capability, they significantly increase computational resource consumption, making it difficult to balance high efficiency and low-cost hardware implementation. In addition, mechanisms such as feedback-based brightness difference adjustment and PWM signal frequency linkage are not adaptable enough to changes in the external environment (such as temperature and panel aging) and long-term accumulated errors, which can easily lead to compensation inaccuracies and reduced system energy efficiency in complex working environments.

[0007] In summary, existing technologies struggle to achieve unified control of millisecond-level response, trend prediction, and energy efficiency robustness in highly dynamic display scenarios with sudden changes in regional load. Therefore, the market and industry urgently require a liquid crystal display driver parameter compensation method that can break the traditional closed-loop mechanism, possess regional load trend prediction capabilities, and offer low latency and high versatility. Especially in new display architectures such as Mini LED and OLED, display modules are required not only to accurately sense the evolution of regional load but also to generate driver parameter adjustment commands in advance based on the patterns of highly dynamic content, achieving faster response and higher display quality.

[0008] In summary, how to establish a liquid crystal display module compensation control method that uses regional load dynamic characteristics to predict forward, overcomes the traditional closed-loop feedback delay, and achieves millisecond-level efficient drive parameter adjustment under limited hardware resources has become a core problem that urgently needs to be solved in the current field of liquid crystal display driving and image compensation control technology. It is also an important technical defect that this patent strives to overcome, providing a clear direction for the innovation and application of subsequent technical solutions. Summary of the Invention

[0009] This application provides a brightness compensation method for regional load differences in a liquid crystal display module, aiming to solve one of the problems or issues of the prior art mentioned in the background section.

[0010] This application provides a brightness compensation method for regional load differences in a liquid crystal display module, specifically including: S1: To address the regional load mutation characteristics of LCD modules under different high dynamic display scenarios, the grayscale distribution entropy, spatial gradient amplitude, and temporal brightness change rate of each sub-region in a typical content sequence are collected. The above three types of spatiotemporal coupling feature vectors are then associated and labeled with the actual load jump amplitude and actual load jump direction of the corresponding sub-region in subsequent frames to construct a regional load dynamic fingerprint library with temporal foresight.

[0011] S2: Based on the spatiotemporal coupling feature vector in the regional load dynamic fingerprint database and the actual load jump data, train a lightweight phase mapping neural network model, abstract the driving parameter adjustment action into a driving parameter phase encoding with three orthogonal dimensions including adjustment start timing, adjustment intensity slope and adjustment decay rhythm, and establish a nonlinear mapping relationship from the spatiotemporal coupling feature vector to the driving parameter phase encoding.

[0012] S3: The trained lightweight phase-mapped neural network model is quantized to integer precision and deployed to a dedicated coprocessor of the display timing controller, enabling it to instantaneously generate the corresponding driving parameter phase code based on the input spatiotemporal coupling feature vector before each frame of the image enters the compensation control stage.

[0013] S4: The phase mapping neural network model is used to receive the spatiotemporal coupling feature vector output by the current frame image analysis module. The integrated lightweight phase mapping neural network model is used for inference calculation to output the phase encoding of the driving parameters that characterize the combination of adjustment start timing, adjustment intensity slope and adjustment decay rhythm, which serves as the control basis for subsequent signal synthesis.

[0014] S5: Based on the phase encoding of the driving parameters, the locally stored driving waveform template library is called in the phase decoder built into the compensation control module to match the pulse shape, duty cycle sequence and voltage swing combination corresponding to the phase command, and fine-tuned in combination with the current panel temperature and aging coefficient environmental parameters to generate a preliminary driving signal that adapts to the current load change trend.

[0015] S6: Apply the initial driving signal to the corresponding sub-region driving circuit of the liquid crystal display module to perform brightness compensation action, so as to realize the forward response of driving parameters based on trend prediction within the millisecond time scale when the regional load changes rapidly, and eliminate the compensation delay caused by the traditional closed-loop feedback mechanism.

[0016] S7: Set a verification window at each preset frame number, collect the feedforward output result and compare it with the actual brightness data detected by the closed-loop feedback, calculate the deviation value between the two, and determine whether the deviation value exceeds the set stability threshold in order to identify whether there is a cumulative error in the feedforward prediction.

[0017] S8: If the deviation value is determined to exceed the stability threshold, the online fine-tuning mechanism of the model is triggered, and the weight parameters of the lightweight phase mapping neural network model are corrected and updated using the deviation data in the current verification window. Otherwise, the existing model parameters are maintained to ensure the accuracy of the phase encoding of the driving parameters and the energy efficiency robustness of the system during long-term operation.

[0018] The brightness compensation method for regional load differences in a liquid crystal display module provided in this application has the following beneficial effects: (1) By constructing a regional load dynamic fingerprint database and combining it with a lightweight phase mapping neural network, this solution realizes the feedforward prediction and phase-based control of the regional brightness response trend in Mini LED backlight or OLED self-emissive modules, effectively overcoming the problems of high system latency and compensation lag in the traditional "detection-computation-execution" serial architecture. Existing technologies generally rely on real-time load detection results for feedback adjustment, which makes it difficult to respond to local brightness changes in a timely manner in high dynamic scenarios (such as fast-paced sports events, game explosion effects, etc.) due to the excessively long processing link, easily causing visual degradation phenomena such as halo trailing and contrast distortion. This invention establishes a load evolution fingerprint dataset containing multi-dimensional features such as grayscale distribution entropy, spatial gradient magnitude, and temporal brightness change rate in the offline stage, and labels the load jump behavior of subsequent frames, enabling the model to have the ability to anticipate the regional load evolution path. When running online, it can quickly output the phase code for adjusting the driving parameters based on the features of the current frame without waiting for the complete load calculation, which significantly advances the timing of the compensation action and greatly compresses the system-level delay between content change and brightness response. In particular, it shows a significantly better response sensitivity and visual continuity than traditional methods in the dynamic transition process of hundreds of milliseconds.

[0019] (2) The complex driving parameter adjustment is abstracted into three orthogonal dimensions of behavior patterns: "adjustment start timing", "intensity slope" and "attenuation rhythm" by adopting a phase mapping mechanism, and low-overhead deployment is achieved, thereby improving the expression efficiency and generalization ability of the control strategy without increasing hardware resources. Unlike the traditional approach of directly outputting voltage / current values ​​or relying on lookup tables to match fixed compensation curves, this invention models the driving response as a combinable and compressible phase state, which enables the originally high-dimensional nonlinear parameter control behavior to be represented in a structured way, greatly reducing model complexity and storage overhead. At the same time, the phase command can be decoded and restored into a specific driving signal in the compensation control module through the local waveform template library, and dynamically fine-tuned in combination with environmental variables such as panel temperature and aging coefficient, taking into account both adjustment accuracy and adaptability. In addition, the model is embedded in the TCON coprocessor after INT8 quantization, which can complete instantaneous inference before each frame of image enters the compensation process, avoiding the cumulative delay caused by multi-level buffering and repeated calculations in traditional closed-loop systems, significantly improving the energy efficiency robustness and operating efficiency of the system, especially suitable for the real-time requirements of high refresh rate and high resolution display scenarios.

[0020] (3) A cross-frame consistency verification mechanism is introduced to integrate limited feedback information while maintaining the feedforward dominance, thus constructing a hybrid control system that combines predictive foresight and long-term stability. This effectively prevents brightness drift or color deviation caused by the accumulation of feedforward errors. Since pure feedforward systems may produce prediction bias due to insufficient training data coverage or extrapolation in extreme scenarios, this scheme sets a periodic verification window to perform a weighted comparison between the feedforward output and the actual feedback result. Only when the deviation exceeds a preset threshold is a lightweight online fine-tuning of the model parameters triggered. This retains the high-speed response advantage of the feedforward system while effectively suppressing system drift. This mechanism avoids computational oscillations caused by frequent loop updates and avoids the resource consumption caused by continuous reliance on brightness difference feedback or iterative optimization in traditional methods, further enhancing the reliability and durability of the system in complex usage environments.

[0021] In summary, this solution establishes a feedforward control paradigm based on load evolution patterns as input, driving response characteristics as constraints, and phase mapping as a bridge. This achieves a technological leap from "responsive correction" to "trend-based preset," significantly improving the brightness response speed, compensation accuracy, and system stability of display modules under high dynamic content without requiring additional hardware investment. It breaks through the bottlenecks of existing technologies in latency control and energy efficiency balance, providing reliable technical support for the image quality performance of high-end display devices in highly real-time application scenarios such as video, gaming, and medical imaging. Attached Figure Description

[0022] Figure 1 This is the main flowchart of a brightness compensation method for regional load differences in a liquid crystal display module; Figure 2 This is a sub-flowchart of a brightness compensation method for regional load differences in a liquid crystal display module; Figure 3 This is another sub-flowchart of a brightness compensation method for regional load differences in a liquid crystal display module. Detailed Implementation

[0023] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0024] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0025] like Figure 1 As shown, this application provides a brightness compensation method for regional load differences in a liquid crystal display module, specifically including: S1: To address the regional load mutation characteristics of LCD modules under different high dynamic display scenarios, collect the grayscale distribution entropy value, spatial gradient amplitude, and temporal brightness change rate of each sub-region in a typical content sequence. Then, associate and label the above three types of spatiotemporal coupling feature vectors with the actual load jump data of the corresponding regions in subsequent frames to construct a regional load dynamic fingerprint library with temporal foresight.

[0026] S2: Based on the spatiotemporal coupling feature vector in the regional load dynamic fingerprint database and the actual load jump data, train a lightweight phase mapping neural network model, abstract the driving parameter adjustment action into a driving parameter phase encoding with three orthogonal dimensions including adjustment start timing, adjustment intensity slope and adjustment decay rhythm, and establish a nonlinear mapping relationship from the spatiotemporal coupling feature vector to the driving parameter phase encoding.

[0027] S3: The trained lightweight phase-mapped neural network model is quantized to integer precision and deployed to a dedicated coprocessor of the display timing controller, enabling it to instantaneously generate the corresponding driving parameter phase code based on the input spatiotemporal coupling feature vector before each frame of the image enters the compensation control stage.

[0028] S4: The phase mapping neural network model is used to receive the spatiotemporal coupling feature vector output by the current frame image analysis module. The integrated lightweight phase mapping neural network model is used for inference calculation to output the phase encoding of the driving parameters that characterize the combination of adjustment start timing, adjustment intensity slope and adjustment decay rhythm, which serves as the control basis for subsequent signal synthesis.

[0029] S5: Based on the phase encoding of the driving parameters, the locally stored driving waveform template library is called in the phase decoder built into the compensation control module to match the pulse shape, duty cycle sequence and voltage swing combination corresponding to the phase command, and fine-tuned in combination with the current panel temperature and aging coefficient environmental parameters to generate a preliminary driving signal that adapts to the current load change trend.

[0030] S6: Apply the initial driving signal to the corresponding sub-region driving circuit of the liquid crystal display module to perform brightness compensation action, so as to realize the forward response of driving parameters based on trend prediction within the millisecond time scale when the regional load changes rapidly, and eliminate the compensation delay caused by the traditional closed-loop feedback mechanism.

[0031] S7: Set a verification window at each preset frame number, collect the feedforward output result and compare it with the actual brightness data detected by the closed-loop feedback, calculate the deviation value between the two, and determine whether the deviation value exceeds the set stability threshold in order to identify whether there is a cumulative error in the feedforward prediction.

[0032] S8: If the deviation value is determined to exceed the stability threshold, the online fine-tuning mechanism of the model is triggered, and the weight parameters of the lightweight phase mapping neural network model are corrected and updated using the deviation data in the current verification window. Otherwise, the existing model parameters are maintained to ensure the accuracy of the phase encoding of the driving parameters and the energy efficiency robustness of the system during long-term operation.

[0033] Step S1: For the regional load change characteristics of the liquid crystal display module under different high dynamic display scenarios, collect the gray-scale distribution entropy value, spatial gradient amplitude, and temporal brightness change rate of each sub-region in the typical content sequence, and associate and label the above three types of spatiotemporal coupling feature vectors with the actual load jump data of the corresponding region in subsequent frames to construct a regional load dynamic fingerprint library with temporal foresight.

[0034] The term "regional load abrupt change feature" refers to the sharp fluctuation of display driving load in a specific region due to excessively high complexity or rapid changes in local image content. The regional load dynamic fingerprint database stores a sample set composed of numerous pairs of spatiotemporal coupling feature vectors and actual load jump data. The actual load jump data includes the actual load jump amplitude and the actual load jump amplitude direction. The "typical content sequence" refers to a continuous stream of video frame data collected from high-dynamic display scenarios encountered in actual applications of liquid crystal display modules, collected to fully induce and characterize the regional load abrupt change features, in order to construct the regional load dynamic fingerprint database. The "subsequent frames" refer to several frames that are temporally consecutive and immediately adjacent to the video frame (or group of frames) currently being used to extract the spatiotemporal coupling feature vector within the typical content sequence.

[0035] Step S1 specifically includes: S1.1: Obtain the original frame data of typical content sequences of the liquid crystal display module in high dynamic display scenarios such as fast cuts in sports events, explosion effects in games, and video transitions. Perform sub-region gridding processing on the original frame data to generate a set of sub-region image blocks containing independent pixel matrices, which serve as the basic execution object for subsequent feature extraction.

[0036] Frame sequence signals are acquired from the LCD module in high dynamic range display scenarios such as fast cuts in sports events, explosion effects in games, and video transitions. The input conditions are the original frame video data for each scenario and the corresponding scanning timing information of the display module. Timing synchronization processing is performed on the original frame video data to ensure that the frame start marker of the video signal is strictly aligned with the row scan start signal of the display module in the time domain, ensuring that subsequent region division can map pixel positions according to the actual physical display layout. A grid partitioning algorithm is applied to the synchronized frame data. Based on the physical pixel resolution of the LCD panel and the distribution of region processing units in the compensation control module, the entire frame image is divided into several sub-region grids. The grid size is determined by constraints that satisfy both the spatial resolution requirements of feature extraction and the data throughput capability of the phase mapping neural network model. Pixel index mapping processing is performed on the generated grid structure, storing the pixels in each sub-region grid as a two-dimensional array matrix according to row and column coordinates, forming a data block set containing independent pixel matrices. The data block set undergoes boundary consistency verification to remove sub-regions with missing or distorted pixels caused by scene switching or edge cropping, ensuring that all subsequent feature extraction objects possess complete data dimensions. Through the above processing method, the original frame data is transformed into an image block set containing independent pixel matrices for each sub-region, establishing a unified and standardized basic execution object for subsequent extraction of spatiotemporally coupled feature vectors such as grayscale distribution entropy, spatial gradient magnitude, and temporal brightness change rate.

[0037] For example, on a 3840×2160 resolution LCD module, raw frame data is acquired for fast-moving scenes in sports events. The frame rate is set to 120 frames per second, and the scan timing start signal frequency is 135kHz. During grid partitioning, the entire frame is divided into grid units of 160 pixels wide and 120 pixels high, resulting in 24×18=432 sub-regions. After synchronization, each sub-region matrix contains 19200 pixels (160×120). During index mapping, data is organized in row-major order, ensuring that the pixel in the i-th row and j-th column corresponds to an index in the two-dimensional array. ,in Represents pixel value, For line numbers, The column number is used. After boundary consistency verification, the pixel loss rate in the edge sub-regions caused by scene transitions was detected to be [percentage missing]. Exceeding the set threshold The sub-region is then removed. The output is a set of 410 valid image patches, providing spatially and temporally consistent input data for the next sub-step to perform Shannon entropy calculation, Sobel operator convolution, and differential temporal statistics.

[0038] S1.2: Based on the set of image blocks in the sub-region, the Shannon entropy calculation algorithm, Sobel operator convolution operation and differential temporal statistics method are used respectively to perform gray-level distribution discreteness analysis, edge texture intensity detection and brightness transient rate measurement in parallel, so as to generate three basic spatiotemporal coupled feature vectors that characterize the current frame state characteristics: gray-level distribution entropy value, spatial gradient magnitude and temporal brightness change rate.

[0039] The input execution object is the set of sub-region image blocks output by step S1.1. This set is a two-dimensional data cell array composed of independent pixel matrices generated by gridding the original frame data.

[0040] For each sub-region image block, the Shannon entropy calculation algorithm is applied, and discreteness analysis is performed using the pixel gray-level probability distribution function. The calculation formula is as follows: in Let be the probability of the i-th gray level. is the gray-level distribution entropy value, ensuring gray-level discreteness is quantized, and n is the total number of gray levels in the image.

[0041] Sobel operator convolution is performed on each sub-region image patch to extract gradient components in the horizontal and vertical directions. The convolution kernel matrix is ​​used to detect edge texture intensity, and the gradient magnitude is calculated using the following formula: in and These are the horizontal and vertical gradient responses, respectively.

[0042] For the same sub-region of a continuous frame sequence, a differential temporal statistics method is applied. The average grayscale values ​​of the current frame and the previous frame are read, and the formula for calculating the luminance transient rate is as follows: in The average grayscale value of the current frame. The average grayscale value of the previous frame. This represents the time-series brightness change rate.

[0043] The grayscale distribution entropy, spatial gradient magnitude, and temporal brightness change rate are stored as three types of basic spatiotemporal coupled feature vectors according to sub-regions, and index identifiers are retained in the data structure for subsequent sliding window correlation analysis.

[0044] Through the parallel processing described above, the set of sub-region image blocks in step S1.1 is transformed into three types of basic feature data that are quantifiable and indexable, thereby achieving an accurate representation of the current frame state and providing stable feature input for the generation of temporal labels in step S1.3.

[0045] For example, in a fast-slicing scene of a sports event at a 1920×1080 resolution, the sub-region grid is set to a 16×16 pixel matrix. The grayscale quantization level in the Shannon entropy calculation is 256, and the probability distribution uses normalized histogram data. The entropy calculation results range from 4.5 to 7.8. The Sobel convolution kernels in the edge texture intensity detection are... The transpose of the matrix in the vertical direction yields a maximum gradient magnitude of approximately 220. In the calculation of the luminance transient rate, the difference in the average grayscale value between adjacent frames is between 15 and 40, corresponding to a transient rate of... The feature vector set output by this process can effectively predict the magnitude of load jumps in subsequent sliding window analysis, significantly improving the accuracy of trend prediction driving parameter encoding.

[0046] S1.3: The evolution trajectory of the basic spatiotemporal coupling feature vector in the continuous frame sequence is analyzed by using the sliding window tracking mechanism. The difference and derivative values ​​of the load state of the sub-regions between adjacent frames are calculated to generate two types of time-series label data: the actual load jump amplitude and the actual load jump direction, which represent the load change trend in the next one to three frames.

[0047] The input conditions are three basic spatiotemporal coupling feature vectors: grayscale distribution entropy value, spatial gradient magnitude, and temporal brightness change rate, after Shannon entropy calculation, Sobel operator convolution operation, and differential temporal statistical processing. Furthermore, these feature vectors have clear spatial index labels after being divided into sub-region grids.

[0048] The basic feature vectors are arranged in order of frame period. The window length is set to four frames by initializing the sliding window tracking mechanism, so that each window covers the feature data set of the current frame and the next three frames, thus ensuring the time extension of trend analysis.

[0049] Within the sliding window, feature difference calculation is performed between adjacent frames. The deviation operator is used to extract the change in brightness distribution dispersion for the grayscale distribution entropy value, the edge intensity difference is calculated for the spatial gradient magnitude, and the transient frequency difference is calculated for the temporal brightness change rate.

[0050] The first derivative operation is performed using the difference results within the window to obtain the rate of change of each characteristic value over time, which serves as an indicator of the speed of load state change, and the direction of change is further determined by sign determination.

[0051] The magnitude of load variation is quantified using a mathematical formula, which is: Where A represents the load variation range. The entropy difference of the gray-scale distribution. For spatial gradient magnitude difference, The sum of squares and the square root of the time-series brightness change rate difference are used to obtain the comprehensive load jump amplitude.

[0052] By combining the derivative results and using the threshold grading method, the direction of change is encoded into a binary label in the direction of brightening or dimming, so as to generate the two types of time-series label data required for trend prediction of the next one to three frames.

[0053] By using a sliding window tracking mechanism and a differential / derivative joint operation method, the basic spatiotemporal coupling feature vector from the previous step is transformed into actual load jump amplitude and direction data with trend prediction function, thereby realizing quantitative and qualitative characterization of future load mutations.

[0054] For example, when the display module processes a sports event video containing rapid camera cuts, the system sets the grayscale distribution entropy value to the Shannon entropy statistical value of a 256-pixel matrix in each sub-region, the spatial gradient magnitude to the mean edge intensity extracted by the Sobel operator, and the temporal brightness change rate to the normalized value of the difference between the mean brightness values ​​of two frames. The sliding window length is set to four frames, and a difference operation is performed on the current frame and the next three frames to obtain... =0.15, =0.08, =0.12. Substitute these difference values ​​into the comprehensive amplitude formula to calculate... = =0.213. The derivative calculation results show that both the grayscale and brightness change rates are positive, while the spatial gradient magnitude change is negative. Based on the directional encoding rule, the directional label is determined to be brightening. The label data of the next one to three frames are used in subsequent feature fusion to construct a regional load dynamic fingerprint database. Its trend prediction capability significantly improves the response speed and brightness adjustment accuracy of the compensation control module.

[0055] S1.4: Based on the basic spatiotemporal coupling feature vector and the time-series label data, perform multidimensional feature tensor splicing and semantic alignment fusion processing to map the spatial texture features, temporal evolution features and load mutation results of the current moment to a unified record structure, so as to generate a spatiotemporal coupling feature vector and a load jump correlation tuple with causal correlation attributes.

[0056] Based on the acquired spatiotemporal coupled feature vectors and time-series label data, the input structure of the multidimensional feature tensor fusion task is defined, clarifying the row and column mapping relationships between gray-level distribution entropy, spatial gradient magnitude, and temporal brightness change rate, corresponding to the spatial texture feature submatrix, temporal evolution feature submatrix, and future load mutation prediction submatrix, respectively. Dimensional alignment is performed on the spatial texture feature submatrix and temporal evolution feature submatrix, and bilinear interpolation is used to achieve precise smoothing of the numerical domain at phase sampling points to eliminate matrix boundary misalignment caused by resolution differences. The load jump amplitude matrix and load jump direction matrix in the time-series label data are encoded and transformed, and mapped into a composite vector using direction encoding. The two-dimensional component of the formal generation direction is weighted and combined with the amplitude value to form the future load mutation vector. The aligned spatial texture matrix, temporal evolution matrix, and future load mutation vector are then tensor-concatenated separately. The structure generates a comprehensive feature tensor, in which Represents the spatial texture matrix. Represents the time evolution matrix, This represents the future mutation vector. Using a semantic alignment fusion algorithm, each element in the comprehensive feature tensor is serialized and stored according to its timestamp label. A causal mapping hash function is then used to establish a one-to-one correspondence between the current features and future mutation trends, generating a spatiotemporally coupled feature vector with causal correlation attributes and a load mutation correlation tuple. Through multi-dimensional feature tensor concatenation and semantic alignment fusion processing, the result of the previous step is transformed into a unified record structure containing spatial, temporal, and predictive causal information, achieving a standardized description of the regional load mutation trend.

[0057] For example, on a Mini LED display device with a resolution of 3840×2160, for a fast-cut sports event scene, the input basic spatiotemporal coupled feature vector contains a gray-level distribution entropy matrix of size 64×64, a spatial gradient magnitude matrix of size 64×64, and a temporal brightness change rate sequence length of 3 frames; the load jump amplitude value in the temporal tag data is 0.8, and the direction angle is 45°. The sampling interval for bilinear interpolation during dimension alignment is set to 0.25 pixels, and the direction encoding is... Calculated Multiplying this by the amplitude value 0.8 yields the future load mutation vector. The combined feature tensor after tensor concatenation has a size of 64×64×5, where channels 1 to 3 represent the spatial texture features, temporal evolution features, and two components of the future mutation vector, respectively; channel 4 represents the amplitude value; and channel 5 represents the timestamp. After semantic alignment and fusion, the number of associated tuples is 4096. Each tuple contains the current frame feature set and the load jump prediction data for the next three frames. Validation results show that this associated tuple can significantly improve prediction accuracy in phase-mapped neural network training and maintain stable causal matching at the millisecond time level.

[0058] S1.5: The spatiotemporal coupling feature vector and the load jump correlation tuple are denoised, cleaned, and normalized. The processed data are then classified, indexed, and stored according to the display scene type and load change level to construct a regional load dynamic fingerprint database with temporal foresight, which serves as the sole input data source for training a lightweight phase mapping neural network model.

[0059] Step S2: Based on the spatiotemporal coupling feature vectors in the regional load dynamic fingerprint database and the actual load jump data, a lightweight phase mapping neural network model is trained. The driving parameter adjustment action is abstracted into a driving parameter phase encoding containing three orthogonal dimensions: adjustment initiation timing, adjustment intensity slope, and adjustment decay rhythm. A nonlinear mapping relationship is established from the spatiotemporal coupling feature vectors to the driving parameter phase encoding. Specifically, this includes: Tensor recombination is performed on the spatiotemporal coupling feature vectors in the regional load dynamic fingerprint database to construct a multidimensional input feature matrix. The load jump amplitude and load jump direction in the actual load jump data are discretized and encoded to generate a standardized target phase label sequence.

[0060] Tensor recombination is performed on the spatiotemporal coupled feature vectors from the regional load dynamic fingerprint database. A multidimensional tensor matrix construction algorithm is called to perform dimensional arrangement adjustment on the original feature vectors in the spatial dimension, temporal dimension, and feature category dimension. This ensures that the numerical matrices corresponding to the gray-scale distribution entropy, spatial gradient magnitude, and temporal brightness change rate have complete row and column correspondence and index mapping in the unified feature matrix.

[0061] The multidimensional feature matrix after dimensional arrangement adjustment is normalized. The mean-variance normalization method is used to distribute the values ​​of different feature channels within the same scale range, thereby eliminating the influence of feature dimension differences on the convergence speed of neural network training.

[0062] Obtain the actual load jump data corresponding to the above feature matrix, call the discretization encoding processing module, segment and map the continuous load jump amplitude values ​​according to the predefined threshold range, and encode them as discrete amplitude labels in the form of classification index. Map the positive and negative states of the load jump direction as binary values ​​of direction labels, so that both amplitude and direction are converted into standardized encoding structures.

[0063] A label alignment algorithm is used to ensure that the discretized load jump amplitude label and direction label are consistent with the feature matrix in the time dimension, and the two are combined into a single target label vector.

[0064] The sequence shaper is invoked to combine the feature matrix and the target label vector into a structured input sample set, forming a supervised learning data pair corresponding to a set of spatiotemporally coupled feature vector inputs and a set of target phase label outputs at each time step.

[0065] Through the above processing method, the result of the previous step is transformed into a multi-dimensional input feature matrix and a standardized target phase label sequence, thereby realizing the construction of the structured input sample set and supervision signal required for the training stage of the lightweight phase mapping neural network.

[0066] For example, in a fast-slicing sports event scenario using a Mini LED LCD display module, the grayscale distribution entropy value output by the regional load dynamic fingerprint library is 14.6, the spatial gradient amplitude is 3.85, and the temporal brightness change rate is 0.42. When constructing the three-dimensional tensor, the spatial dimension grid is 4×4, the temporal dimension length is 3 frames, and the feature category dimension includes the above three types of features. A normalization transformation maps the grayscale distribution entropy value to... The obtained normalized value, spatial gradient magnitude mapped to The obtained normalized value, the temporal brightness change rate is mapped to The obtained normalized value. The actual load jump amplitude is 12.4, which is encoded as amplitude label "2" in the preset amplitude threshold range [0,5], [5,10], [10,15]; the actual load jump direction is positive, which is encoded as direction label "1". After label alignment, the target phase label vector [2,1] is formed. Each sample in the final generated input sample set contains a 4×4×3 feature matrix and a target label vector [2,1]. The neural network trained in this scenario can significantly improve the prediction accuracy of the start timing of the driving parameter adjustment and the direction consistency of the output phase encoding.

[0067] S2.2: Based on the input sample set and the supervision signal, a lightweight phase mapping neural network model with a multilayer perceptron architecture is constructed. The backpropagation algorithm is used to iteratively optimize the network weight parameters to minimize the loss function value between the predicted phase code and the target phase label sequence, thereby obtaining an initial phase mapping model with nonlinear fitting capability.

[0068] Based on the multidimensional input feature matrix and target phase label sequence obtained by S2.1, a lightweight phase mapping neural network model with a multilayer perceptron architecture is constructed and the weight parameters are initialized. The number of nodes in each hidden layer, the type of activation function, and the regularization method are set to ensure that the nonlinear expressive power of the model matches its embedded deployability.

[0069] The input feature matrix is ​​fed into the network layer by layer using a feedforward propagation mechanism. After passing through the weight matrix product and activation function transformation, a continuous numerical output vector of the prediction driving parameter phase encoding is obtained.

[0070] The element-wise error between the predicted output and the target phase label sequence is calculated, and the mean squared error is used as the loss function. The expression for the loss function is as follows: ,in For the sample size, For the first Predicted values ​​for each sample, This corresponds to the target label value.

[0071] The gradient of the loss function is calculated layer by layer using the backpropagation algorithm. The chain rule is used to propagate the error of the output layer back to each hidden layer. The corresponding weight matrix and bias vector are updated in combination with the learning rate parameter.

[0072] At the end of each iteration, the weight update magnitude is pruned to prevent numerical overflow and gradient explosion by limiting the gradient range, thus ensuring the stability of the training process.

[0073] Repeat the above feedforward calculation and backpropagation process until the loss function value converges to the preset threshold to obtain an initial phase mapping model that has nonlinear fitting ability and performs stably on the training sample set.

[0074] Through the above iterative optimization process, the structured input sample set and supervision signal are transformed into network weight configuration with the ability to accurately predict the phase encoding of driving parameters, thus achieving the expected technical effect of constructing a nonlinear mapping model from spatiotemporally coupled feature vectors to driving parameter phase encoding.

[0075] For example, in the high dynamic range scenario of a Mini LED LCD display module, the input feature matrix is ​​set to 128×64, containing 128 samples, each with a 64-dimensional spatiotemporal coupled feature vector; the target phase label sequence is a 128×3 matrix, corresponding to the adjustment of the start-up timing, voltage change rate, and signal tail coefficient. The multilayer perceptron is configured with a 64-node input layer, two hidden layers of 48 and 32 nodes respectively, using ReLU activation, and a 3-node output layer. The loss function is mean squared error, and the learning rate is set to... During training, batch size is used. The algorithm iterates for 200 rounds, updating weights using backpropagation and the Adam optimizer. The initial loss value is calculated using the formula. After 50 iterations, the loss decreased to The model output is nearly identical to the target label triplet in terms of adjusting the start-up timing, voltage change rate, and signal tailing coefficient. In embedded operation, the delay of the predicted drive phase command is less than 2 milliseconds, the response speed of brightness compensation is significantly improved, and the energy efficiency is greatly enhanced.

[0076] S2.3: The output layer activation function of the initial phase mapping model is subjected to constraint mapping processing to force the continuous numerical output to be mapped to the decoupled space composed of three orthogonal dimensions: adjustment initiation timing, adjustment intensity slope, and adjustment decay rhythm. This generates the original driving parameter phase code representing the combined state of the driving parameter adjustment actions. The combined state of the driving parameter adjustment actions is a driving waveform configuration mode uniquely determined by the specific parameter values ​​in the three orthogonal dimensions of adjustment initiation timing, adjustment intensity slope, and adjustment decay rhythm within a driving cycle.

[0077] S2.4: Perform generalization capability verification processing based on the correlation between the original driving parameter phase encoding and the actual load jump data to remove overfitted nodes and prune redundant connections, thereby obtaining a lightweight phase mapping neural network model that meets the requirements of embedded deployment, ensuring that the model reduces computational complexity while maintaining accuracy.

[0078] Based on the correlation matrix established by the initial driving parameter phase encoding and actual load jump data, the input conditions for verifying the generalization ability of the lightweight phase-mapped neural network model are: the current set of model weight parameters, network topology information, training sample set partitioning scheme, and embedded deployment computing power constraints. A hierarchical hold-out method is used to construct the partitioning rules for the training set and independent validation set, with the original driving parameter phase encoding and the corresponding load jump amplitude and direction serving as the supervision reference signal for the validation set. The model's forward inference program on the validation set is called to calculate the standardized squared error between the output phase encoding result and the reference signal. The following mean squared error calculation formula is used for performance quantification: ,in Indicates the number of validation samples. To predict the phase-coded value, The target phase encoding value is used. The statistical distribution of the absolute values ​​of the weight gradients of each layer's nodes on the validation set is compared with a preset generalization threshold. Overfitting nodes whose gradients remain in the low-sensitivity range for extended periods and whose validation error contribution does not exceed the threshold are removed to reduce model complexity. After node removal, dependency analysis is performed on the connection paths of each layer to identify redundant connections that can be pruned. Weight sparsity is then performed, setting connections with weights close to zero that do not affect the main output characteristics to zero, thereby compressing the model parameter size. Multiple rounds of validation set inference are performed on the model after node removal and connection pruning to evaluate the difference between its mean squared error and the original model error, ensuring that the compressed model meets the real-time requirements of embedded computing within a preset tolerance range of accuracy loss. By removing overfitting nodes and pruning redundant connections, the results of the previous step are transformed into a lightweight phase mapping neural network model that meets embedded deployment requirements and has continuous generalization capabilities, achieving a reduction in computational complexity and storage consumption while maintaining trend prediction accuracy.

[0079] For example, in a Mini LED high dynamic range display scenario, the regional load dynamic fingerprint database contains 40,000 sample records. Each record contains a three-dimensional feature vector of grayscale distribution entropy, spatial gradient magnitude, and temporal brightness change rate, as well as corresponding load jump amplitude and direction labels. The initial model has a total of 560,000 weight parameters, and the embedded coprocessor's computing power is constrained to an execution latency of no more than 0.8 milliseconds per frame. In the validation set, the generalization threshold is set to an absolute gradient value below 0.0005 and a contribution rate below 0.5, and 3,200 overfitted nodes are removed. Dependency analysis reveals that redundant connections account for 15%, and pruning is performed, reducing the total number of model parameters to 470,000 and storage usage to 84% of the original. The mean squared error of the pruned model on the validation set is calculated as follows: The calculated result is 0.0032, which is only a slight increase from 0.0030 before compression. In actual embedded operation, the single-frame inference latency is shortened to 0.65 milliseconds, which significantly improves the real-time performance of the system and maintains the trend prediction accuracy.

[0080] S2.5: The lightweight phase mapping neural network model is used to perform forward inference calculation on the new spatiotemporal coupled feature vector to output a high-precision driving parameter phase code, thereby completing the establishment of a nonlinear mapping relationship from the spatiotemporal coupled feature vector to the driving parameter phase code.

[0081] The input conditions are the parameters and network structure of the lightweight phase-mapped neural network model obtained after processing by S2.4, and the new spatiotemporal coupled feature vector data set passed in by the image analysis module.

[0082] The new spatiotemporal coupled feature vector is subjected to dimension verification and format normalization to ensure that the order of each feature component is consistent with the feature matrix in the training stage, and to eliminate numerical accuracy deviations caused by transmission.

[0083] The normalized feature vectors are input into the input layer of the lightweight phase-mapped neural network, and the internal multilayer perceptron weights and biases are invoked to perform matrix multiplication and nonlinear activation transformation to extract the nonlinear correlation features between the regional load evolution law and the driving response characteristics.

[0084] A forward inference scheduler is used to complete the linear transformation from the input layer to the hidden layer and the feature mapping from the hidden layer to the output layer in stages. The output results are subjected to a constraint projection operation, which forces the predicted values ​​to be mapped to a decoupled space with three orthogonal dimensions: adjusting the start timing, adjusting the intensity slope, and adjusting the decay rhythm.

[0085] During the mapping process, the normalization factor is calculated for each of the three-dimensional output vectors. ,in and This represents the maximum and minimum reference values ​​for each dimension during the training process, in order to eliminate the differences in the dimensions of different physical quantities.

[0086] The normalized 3D output vector is integrated into the phase encoding of the driving parameters, recorded as a structured data triple, and stored in the mapping engine cache.

[0087] Through the above-mentioned multi-stage matrix transformation, nonlinear mapping and normalization processing, the training model of the previous step is effectively combined with the new input features, and transformed into high-precision driving parameter phase encoding to realize the output of control data for trend prediction.

[0088] For example, in a high-dynamic game scenario using a Mini LED backlit LCD display module, the phase-mapping neural network model receives a spatiotemporal coupled feature vector containing a grayscale distribution entropy of 12.6, a spatial gradient amplitude of 8.3, and a temporal brightness change rate of 15.2. During the input preprocessing stage, this vector is integer-quantized and normalized according to training specifications, and the feature entropy value is standardized to obtain... ≈0.52. During the inference phase, a three-layer perceptron is used to extract temporal nonlinear features and output the original phase values ​​[4.5, 1.8, 0.3]. The normalization factor is calculated as the first dimension. ≈0.33, second dimension ≈0.5, third dimension =1.0. Normalization process yields the phase command triplet [1.49, 0.9, 0.3]. This triplet corresponds to the drive adjustment behavior mode of "early start + slow rise + short trailing". It significantly improves the response speed in the subsequent drive signal synthesis process and achieves millisecond-level output of brightness compensation in actual verification, effectively eliminating the flickering phenomenon of the display screen when the brightness changes.

[0089] like Figure 2 As shown, step S3 involves quantizing the trained lightweight phase-mapping neural network model to integer precision and deploying it to a dedicated coprocessor of the display timing controller. This enables the model to instantaneously generate the corresponding driving parameter phase code based on the input spatiotemporal coupling feature vector before each frame of the image enters the compensation control stage. Specifically, this includes: S3.1: Obtain the floating-point weight parameters and network topology data of the trained lightweight phase mapping neural network model, and use the fixed-point quantization algorithm to perform precision compression processing on the floating-point weight parameters, mapping the 32-bit floating-point number to an 8-bit integer value, so as to generate an integer precision model parameter set that is adapted to the embedded computing power constraints, thereby significantly reducing storage occupation and computation latency while maintaining nonlinear fitting ability.

[0090] The input conditions include a 32-bit floating-point weight parameter set and a network topology description file for a lightweight phase-mapped neural network model that has been trained and whose accuracy meets the requirements for embedded deployment.

[0091] Perform data traversal processing on the floating-point weight parameter set, and statistically analyze the numerical distribution characteristics of each layer of weights, including minimum, maximum and variance values, as a reference for setting the quantization interval.

[0092] Based on the statistical results, a segmented interval mapping table is constructed. A symmetric linear fixed-point quantization method is used to map the original 32-bit floating-point number to an 8-bit integer value. The weight of each floating-point number is calculated using the following quantization formula: in, The quantized integer value. These are the original floating-point weight values. The scaling factor is determined by the interval mapping table.

[0093] The scaling factor is optimized layer by layer, and adjustments are made based on the quantization error evaluation index and the principle of minimizing the mean square error. The value, and the corresponding formula is: in ( ) represents the maximum absolute value of the current layer weight, and 127 represents the maximum positive value of an eight-bit signed integer.

[0094] The quantized integer weight parameters are recombined with the unchanged network topology to generate an integer precision model parameter set, and memory alignment optimization is performed on the parameter set to ensure access efficiency in the coprocessor cache architecture.

[0095] By using fixed-point quantization compression, the floating-point weight data from the previous step is transformed into an integer precision model parameter set that adapts to embedded computing power constraints, thereby significantly reducing storage usage and inference computation latency while maintaining nonlinear fitting capabilities.

[0096] For example, when a display module deploys a lightweight phase-mapped neural network, its original model contains a 15-layer multilayer perceptron structure with a total of 1.2 million weights and a floating-point range between -0.85 and 0.92. During quantization, a scaling factor is calculated for the maximum value of each layer's weights. For instance, if the maximum absolute value of a weight in a hidden layer is 0.76, the corresponding scaling factor is... ≈ Each floating-point weight is divided by the scaling factor and rounded to obtain an eight-bit integer weight value. After quantization, the size of the integer model is reduced from 5.0MB to 1.25MB, the inference latency is reduced from 6.8ms to 3.1ms, and the prediction accuracy of the phase coding of the region driving parameter does not decrease significantly, ensuring that the coprocessor can complete millisecond-level instruction generation between each frame scan interval, and realize the feedforward response for brightness compensation.

[0097] S3.2: Based on the parameter set of the integer precision model and the instruction set architecture features of the dedicated coprocessor for the display timing controller, the model operators are refactored using compiler optimization techniques to achieve hardware affinity. The general neural network operations are converted into a parallel matrix multiplication instruction sequence dedicated to the coprocessor, thereby generating a binary firmware code package that can be executed efficiently on the dedicated coprocessor, achieving deep coupling between algorithm logic and hardware resources.

[0098] The input execution object consists of the integer precision model parameter set obtained by fixed-point quantization in step S3.1 and the instruction set architecture description file of the dedicated coprocessor for the display timing controller. This object contains the weight matrix, bias vector, topology index, and a set of parallel operation instructions supported by the coprocessor. Based on the integer precision model parameter set, an operator mapping table is established, decomposing general operators such as convolution, fully connected, activation, and normalization in the neural network model into basic units of matrix multiplication and vector operations. The mapping table is parsed one by one using the instruction set architecture description file to filter out the operation paths with hardware acceleration capabilities of the coprocessor, and marking unsupported operators that need to be implemented through software simulation. For hardware-accelerated matrix multiplication operations, compiler optimization methods are used to perform instruction rearrangement. With the help of loop unrolling and register binding strategies, the batch processing of matrix multiplication in the multilayer perceptron is converted into a parallel multiply-add instruction sequence dedicated to the coprocessor. In the process of generating matrix multiply-add instructions, block grouping is performed according to the dimensionality characteristics of the weight matrix, and the product result of a single block matrix is ​​calculated through the following operations: in, For block weight matrix units, For the corresponding input feature vector unit, This indicates element-level operations performed according to the coprocessor's parallel addition rules. For block indexing, a coprocessor-specific binary template is embedded into the matrix multiplication sequence after instruction rearrangement. Combined with optimized activation function call statements, a firmware instruction list that can be executed on hardware without waiting is generated. Linking and address allocation are performed on the generated firmware instruction list, ensuring that all instruction entities and weight data are physically contiguous in storage space, guaranteeing the pipelining characteristics of data loading and instruction execution. Through hardware affinity reconstruction, general neural network operations are transformed into a coprocessor-specific parallel matrix multiplication instruction sequence, achieving deep coupling between model execution logic and hardware resources. This provides directly deployable binary files for the subsequent firmware burning and real-time mapping engine construction in the S3.3 sub-step.

[0099] For example, for a display timing controller coprocessor of a certain LCD module, its parallel matrix multiplication instruction supports executing 4-way multiply-accumulate operations per cycle, with a maximum register count of 64. The weight matrix dimension in the integer precision model parameter set is 128×64, and the bias vector dimension is 128. During the compiler optimization stage, the matrix multiplication is divided into 32 blocks by row, each containing 4 rows. A loop unrolling strategy is used to continuously call 8 groups of coprocessor multiply-accumulate instructions in the instruction chain, achieving the processing of a matrix multiplication block in a single cycle. (In the formula...) In the middle, block index The values ​​range from 0 to 31, and the elements of the weight matrix are... and input feature elements All are stored in the coprocessor cache, corresponding to parallel addition operations. Physically, it is executed in parallel by the four ALUs of the coprocessor. After hardware affinity refactoring, the firmware instruction list achieved millisecond-level inference in load testing, and the inference latency per frame was reduced to about half that of the original general instruction execution, which significantly improved the real-time performance and stability of the phase encoding generation of the driver parameters.

[0100] S3.3: Receive the binary firmware code package and burn it into the non-volatile storage area of ​​the dedicated coprocessor for the display timing controller, and use the system initialization bootloader to perform integrity verification and memory address remapping processing on the loaded model data.

[0101] S3.4: Configure the data input interface of the phase mapping neural network model to establish a high-speed direct connection channel with the output bus of the image analysis module. Use the zero-copy data transmission mechanism to define the inflow format and timing alignment rules of the time-space coupled feature vector, so as to generate a low-latency data flow pipeline that can capture and preprocess input data in real time between each frame image scanning interval, eliminating the additional time overhead caused by data transportation.

[0102] The input conditions are that the dedicated coprocessor of the display timing controller has been programmed and is in a ready state, and the image analysis module transmits a spatiotemporal coupled feature vector containing grayscale distribution entropy value, spatial gradient amplitude and timing brightness change rate through the output bus.

[0103] Based on the characteristics of the coprocessor hardware registry read interface, the physical layer communication protocol parameters of the receiving bus are defined, including data width, clock frequency, and the state sequence of start and stop bits, to ensure that the input interface has a seamless electrical match with the output bus of the image analysis module.

[0104] The high-speed direct connection channel construction logic module in the coprocessor firmware is invoked to set the mapping relationship between the direct channel buffer pointer and the bus address, so that the spatiotemporal coupling feature vector is directly written to the coprocessor cache during transmission without passing through the system main memory, thus achieving direct data delivery under zero-copy conditions.

[0105] The inflow format of the time-space coupled feature vector is defined on the direct connection channel, including the sequential arrangement of vector components, the fixed-point encoding bit width of each component, and the position identifier of the fixed-point decimal places. This format is bound to the parser of the data receiving module at the hardware description language (HDL) level to ensure that the input data is consistent with the internal representation of the model inference unit.

[0106] A timing alignment rule is constructed by measuring the phase position of the current display scan line cycle through a hardware timer and synchronizing the capture time of the time-space coupled feature vector with the idle phase of the scan line gap. This ensures that the feature data is strictly matched with the display refresh cycle before entering the inference unit, thereby avoiding compensation delay caused by timing drift.

[0107] By combining the zero-copy transmission mechanism with the timing alignment rules, the model loading result from the previous step is transformed into a low-latency data pipeline with real-time capture and preprocessing capabilities, enabling high-speed input of spatiotemporal coupled feature vectors between each frame scan and eliminating data transport overhead.

[0108] For example, in a Mini LED display module with a resolution of 3840×2160 and a refresh rate of 240Hz, the coprocessor's data input interface is set to a 32-bit width, a 2GHz clock frequency, a logic high start bit, and a logic low stop bit. The direct-connect channel buffer size is configured to 2048 bytes, with mapped addresses from 0x8000 to 0x87FF, to avoid sharing conflicts with the system's main memory. The inflow format of the spatiotemporal coupling feature vector is defined as follows: grayscale distribution entropy value in the first 8 bits, spatial gradient magnitude in the middle 8 bits, and temporal brightness change rate in the last 16 bits. All three are encoded using 8-bit fixed-point encoding, with 3 decimal places. The timing alignment rule utilizes the nanosecond-level resolution of the coprocessor's internal timer to measure the scan line period. The trigger phase for capturing the spatiotemporal coupled feature vector is set to 80% of the row period, meaning data is injected during the last 20% of the time window before each row of pixels finishes rendering. With this configuration, the input latency of the spatiotemporal coupled feature vector from the image analysis module to the coprocessor inference unit is reduced to 1.3 microseconds, significantly improving the real-time performance of the phase encoding generation of the driving parameters and ensuring that the compensation control response is completed within a millisecond timescale.

[0109] S3.5: Start the online inference service of the phase mapping neural network model. Through the internal inference scheduler and set frame synchronization to trigger interrupt, use the pipeline parallel execution strategy to coordinate the concurrent operation of the three stages of feature reading, model inference and instruction output, so as to generate the real-time closed-loop capability of instantaneously outputting the corresponding driving parameter phase encoding based on the input spatiotemporal coupling feature vector within a millisecond time scale, and complete the final construction from offline model to online real-time control core.

[0110] An independent buffer is initialized and allocated to form a three-stage pipelined operation framework. The frame synchronization trigger interrupt signal is bound to the display scan line timing reference, and trigger accuracy is maintained through hardware clock counting. Upon interrupt triggering, the write pointer of the buffer is immediately locked to ensure the integrity of the feature data. The spatiotemporally coupled feature vector captured by the interrupt trigger is sent to the feature reading stage, directly mapped to the coprocessor cache using a zero-copy transfer mechanism, achieving parallelization of data reading and inference preparation. Inference stage operator scheduling is initiated simultaneously with the feature reading completion flag being set. During the inference stage, the input features are nonlinearly mapped using parallel matrix multiplication operations of a multilayer perceptron. A pipelined mechanism reserves an instruction encapsulation area before computation is complete, allowing instruction formatting and inference operations to occur simultaneously, reducing waiting latency. After the inference output generates the phase-encoded latent variable of the driving parameters, this latent variable is passed to the instruction encapsulation stage. Field encoding, timing alignment, and interface protocol matching are performed according to the phase instruction triplet rules, generating a phase-encoded data packet of the driving parameters that can be directly parsed by the compensation control module.

[0111] For example, in a Mini LED LCD display module with a resolution of 3840×2160 and a refresh rate of 120Hz, the coprocessor inference scheduler needs to... The entire pipeline operation is completed within seconds. Frame synchronization interrupts are bound to the start signal of each line scan. The buffer capacity is configured to 256KB for storing quantized feature tensors. The zero-copy mapping during the feature reading phase takes approximately [time missing]. Milliseconds; the inference phase uses INT8 matrix multiplication, performing a 4096×1024 weight matrix product per frame, which takes [time missing]. Milliseconds. The time consumed during the instruction encapsulation stage to perform protocol encoding on the three orthogonal dimensions of the inference output. Milliseconds, total latency controlled within Within milliseconds. This configuration enables the phase encoding of driving parameters to be generated in advance within one frame period and output to the compensation control module. Verification results show that in video transition scenes that quickly switch between high and low brightness, the visible delay of brightness compensation is significantly reduced, and the energy efficiency of the display panel is greatly improved.

[0112] like Figure 3 As shown, step S4 involves: receiving the spatiotemporal coupling feature vector output by the current frame image analysis module using the phase mapping neural network model; performing inference calculations through the internally integrated lightweight phase mapping neural network model; and outputting the phase encoding of driving parameters representing the combined state of adjusting the start-up timing, adjusting the intensity slope, and adjusting the attenuation rhythm, which serves as the control basis for subsequent signal synthesis. Specifically, this includes: S4.1: Obtain the spatiotemporal coupling feature vector output by the current frame image analysis module, which includes grayscale distribution entropy value, spatial gradient magnitude, and temporal brightness change rate. Perform integer quantization preprocessing operation on the spatiotemporal coupling feature vector to generate a quantized spatiotemporal coupling feature vector tensor, which serves as the standardized input data for the lightweight phase mapping neural network model.

[0113] The system acquires a spatiotemporally coupled feature vector from the current frame image analysis module, containing grayscale distribution entropy, spatial gradient magnitude, and temporal brightness change rate. A high-speed direct connection is used to transfer this feature vector from the image analysis module's output bus to the input buffer of the real-time processing framework, which deploys a phase-mapped neural network model, using a zero-copy method. Based on coprocessor computing power constraints, a quantization preprocessing unit performs fixed-point mapping on the floating-point feature values, converting the 32-bit floating-point numbers element-wise into 8-bit integer values, forming a basic feature matrix adapted to the hardware execution format. The basic feature matrix is ​​then dimensionally scanned to detect the dynamic range of each feature field and calculate normalization coefficients. A linear scaling algorithm is used to map the feature values ​​to a preset integer closed interval to ensure the numerical stability of the input tensor. A multidimensional rearrangement process is performed on the normalized integer feature matrix, arranging the grayscale distribution entropy, spatial gradient magnitude, and temporal brightness change rate in the tensor according to the model input order, and assigning a fixed index position to each feature to match the structured requirements of the neural network input port. During quantization, a quantization error vector is calculated, and the degree of quantization accuracy preservation is evaluated using the following formula: in, These are the original floating-point eigenvalues. For the corresponding quantized integer value, The total number of features is given. Based on the quantization error vector analysis results, the normalization coefficients and mapping intervals are adjusted to ensure that the quantization process remains balanced across different feature channels. Through the above quantization, normalization, and rearrangement processes, the original spatiotemporal coupled feature vectors from the previous step are transformed into standardized quantized spatiotemporal coupled feature vector tensors, enabling efficient inference input preparation for lightweight phase-mapped neural network models on coprocessors.

[0114] For example, in a fast-slicing sports event scene of a Mini LED LCD display module, the current frame image analysis module outputs a grayscale distribution entropy value of 2.85, a spatial gradient amplitude of 15.7, and a temporal brightness change rate of 0.42. These floating-point values ​​are converted to eight-bit integers via fixed-point mapping: the grayscale distribution entropy value is mapped to the integer 2^31, the spatial gradient amplitude to the integer 124, and the temporal brightness change rate to the integer 108. When calculating the normalization coefficients, the grayscale entropy channel coefficient is set to 81.05, the spatial gradient channel coefficient to 8.25, and the brightness change rate channel coefficient to 256.0. After scaling, the values ​​of each channel stabilize within the closed interval of 0 to 255. After rearranging according to the model input order, a quantized feature matrix of length 3 [2^31, 124, 108] is formed. Using the formula... Calculate the quantization average error when When the value is equal to 0.0047, the quantization accuracy is determined to meet the preset threshold. The output quantized spatiotemporal coupled feature vector tensor is input to the lightweight phase mapping neural network, which can significantly improve the feature reading speed and maintain a significant improvement in the prediction accuracy of the driving parameters during the inference stage.

[0115] S4.2: Based on the quantized spatiotemporal coupled feature vector tensor, the lightweight phase mapping neural network model deployed in the dedicated coprocessor of the display timing controller is invoked to perform forward inference operations. The nonlinear transformation capability of the multilayer perceptron structure inside the model is used to extract the implicit correlation between the regional load evolution law and the drive response characteristics, so as to output the phase-encoded latent variable of the drive parameter containing the original values ​​of three orthogonal dimensions.

[0116] S4.3: Perform multi-dimensional decoupling mapping processing on the latent variables of the phase encoding of the driving parameters. According to the predefined phase encoding rules, the original values ​​are mapped to the time offset representing the adjustment start timing, the voltage change rate representing the adjustment intensity slope, and the signal tail coefficient representing the adjustment decay rhythm, so as to generate a structured driving parameter phase encoding triplet with clear physical meaning.

[0117] The original values ​​of the three orthogonal dimensions of the phase-encoded latent variables of the driving parameters are obtained as the input reference for the decoupling mapping process, and a correspondence matrix with the predefined phase encoding rules is established. The latent variable values ​​are then fed into the time mapping module, voltage slope mapping module, and tail coefficient mapping module, respectively. The piecewise linear transformation functions of each module are used to convert the values ​​into physically meaningful parameters. In the time mapping module, based on the system timing reference frequency and frame period length, the latent variables are converted into time offsets through multiplication coefficients to ensure that the conversion result is synchronized with the scan line. The formula is used... in, This is the time offset. As the periodic benchmark coefficient, This is a time-dimensional latent variable. For the voltage slope mapping module, the latent variable value is multiplied by the preset voltage dynamic range and divided by the adjustment time constant to generate the voltage change rate, using the formula... in, The rate of change of voltage. This is the maximum voltage swing. As a latent variable of intensity, The slope adjustment time constant. For the tail coefficient mapping module, the latent variable is input into the exponential decay function to obtain the signal tail coefficient, as shown in the formula: in, This is the trailing coefficient. To decay the latent variable, This represents the system decay time constant. The three physical parameters are standardized in terms of data type and unit format to generate structured driving parameter phase-encoded triplets, and their matching degree with the physical model is verified. Through multi-dimensional decoupling mapping and physical parameter transformation, the latent variable results from the previous step are transformed into time offsets, voltage change rates, and signal tailing coefficients with clear physical meaning that can directly guide waveform synthesis, thus realizing the accurate driving parameter compilation capability of the compensation control module.

[0118] For example, in a high dynamic range display device using Mini LED backlighting, the three latent variables obtained from the inference output of the lightweight phase-mapped neural network model are time dimension 0.45, voltage intensity dimension 0.78, and decay dimension 0.32. The system timing reference coefficients are set to a period of 16.6 ms per frame, a maximum voltage swing of 5.0 V, a slope adjustment time constant of 4.0 ms, and a decay time constant of 8.0 ms. The time dimension latent variables are then expressed using the formula... The time offset was found to be 7.47 ms; the voltage intensity latent variable was then calculated using the formula... The voltage change rate was found to be 0.975 V / ms; the decay dimension latent variable was then calculated using the formula... The trailing coefficient was found to be 0.960. In multi-scenario testing, after this triplet was applied to the compensation control module, the accuracy of the driving waveform synthesis was significantly improved, and the brightness compensation response time was compressed to less than 8ms, meeting the brightness stability requirements for high dynamic content playback.

[0119] S4.4: Based on the structured driving parameter phase encoding triplet, perform timing alignment and format encapsulation operations to integrate the time offset, voltage change rate and signal tail coefficient into a driving parameter phase encoding data packet that conforms to the compensation control module interface protocol, so as to serve as the direct control basis for the subsequent phase decoder to call the driving waveform template library and perform signal synthesis.

[0120] Step S5: Based on the phase encoding of the driving parameters, the locally pre-stored driving waveform template library is called in the phase decoder built into the compensation control module to match the pulse shape, duty cycle sequence, and voltage swing combination corresponding to the phase command. Fine-tuning is then performed based on the current panel temperature and aging coefficient environmental parameters to generate a preliminary driving signal adapted to the current load change trend. The driving waveform template library is a structured waveform parameter database pre-stored in the local memory of the compensation control module. Its core function is to decode and map the abstract driving parameter phase encoding generated upstream into a set of specific electrical signal parameters that can be directly used in the driving circuit. Each entry in the driving waveform template library stores a standard driving waveform template, which uniquely corresponds to one or more specific driving parameter phase codes (or phase index keys).

[0121] Specifically, it includes: S5.1: Obtain the adjustment start timing feature value, adjustment intensity slope feature value, and adjustment decay rhythm feature value from the phase encoding of the driving parameters. Use a multi-dimensional vector space mapping algorithm to jointly encode the feature values ​​of the three orthogonal dimensions to generate a phase index key value with unique identification, which will serve as the accurate addressing basis for subsequent waveform retrieval.

[0122] The input drive parameter phase-encoded data packet contains characteristic values ​​for adjusting the start-up timing, adjusting the intensity slope, and adjusting the decay rhythm, all of which are structured numerical parameters.

[0123] Normalized timing coding is performed on the feature value of the adjustment start timing to map the time offset to a uniform millisecond reference scale, ensuring consistency in cross-scenario comparisons.

[0124] The slope characteristic value of the regulation intensity is standardized by a scaling factor, and the voltage change rate is converted into a dimensionless slope coefficient so as to maintain the same physical meaning in different voltage drive ranges.

[0125] An exponential curve is fitted to the characteristic value of the adjustment decay rhythm, and the signal tail coefficient is mapped to the time-domain decay constant, taking into account both amplitude reduction and time delay control indicators.

[0126] The three orthogonal eigenvalues ​​mentioned above are input into a multidimensional vector space mapping algorithm to construct a three-dimensional coordinate vector. (in Normalized startup timing values, For the standardized slope coefficient, (where the decay constant is used) and hash compression coding function Generate a globally unique phase index key value.

[0127] Write the phase index key value into the address register as a precise location identifier for subsequent retrieval of the drive waveform template library.

[0128] By using a multi-dimensional vector space mapping process, the result of the previous step is transformed into a unique phase index key value with a clear physical meaning, thereby achieving precise addressing and optimized matching efficiency for drive waveform retrieval.

[0129] For example, in a Mini LED high dynamic range display scenario, the received driving parameter phase-encoded triplets consist of a time offset of 5ms, a voltage change rate of 0.8 V / ms, and a signal tail coefficient of 0.25. Normalized timing encoding is performed on the time offset, with a reference frame period of 16.67ms, corresponding to the normalized value... ≈0.299. The voltage change rate is standardized using a scaling factor. Assuming the maximum change rate in the driving range is 1.5 V / ms, the corresponding standardized slope coefficient is... ≈0.533. Perform exponential curve fitting on the signal tail coefficient, assuming the fitting formula is... ≈1.386. Combine the three into a three-dimensional vector. Input a hash compression encoding function, assuming the function form is: The output key value is 0x7FA923B4. After this key value is written to the address register, the corresponding reference pulse shape is matched in the drive waveform template library. The reference pulse shape is a cubic spline smooth waveform, the duty cycle sequence is [0.4, 0.6], and the voltage swing combination is [2.8V, 3.3V]. This achieves millisecond-level matching and accurate retrieval, effectively improving waveform search performance and ensuring that the physical characteristics of the drive signal are consistent with the phase command.

[0130] S5.2: Receive the phase index key value, perform a fast matching operation in the locally pre-stored drive waveform template library based on the hash lookup mechanism, extract the reference pulse shape data, reference duty cycle sequence data and reference voltage swing data that completely correspond to the phase index key value, and construct an initial drive waveform set containing basic timing and amplitude information.

[0131] The input conditions include a uniquely identifying phase index key value generated by S5.1 and a drive waveform template library data structure stored locally in the compensation control module. For the phase index key value, a hash key mapping table is established to map the key value to the address index area of ​​the drive waveform template library, achieving the initial stage of address positioning. Based on the address index area, a hash lookup operator is invoked to execute an open addressing or chained hash matching strategy to calculate the hash function value of the key value. ,in A phase index key is used to ensure that the search path completes index location within constant time complexity. Using the found index position, the corresponding reference pulse shape data object is extracted from the drive waveform template library. This object contains an array of pulse waveform sampling points and timing description parameters such as rising and falling edges. Reference duty cycle sequence data is read in parallel. This sequence, composed of duty factors, defines the proportional characteristics of the PWM drive signal's working cycle. Simultaneously, reference voltage swing data is retrieved. This data is stored as a voltage scalar sequence, representing the amplitude control range of the drive signal. The reference pulse shape data, reference duty cycle sequence data, and reference voltage swing data are combined according to the output specifications of the drive waveform template library into a unified initial drive waveform set. This set structurally contains basic timing and amplitude information, providing a complete waveform input reference for subsequent environmental parameter fine-tuning. A hash-based fast matching process is used to transform the results of the previous step into a complete initial drive waveform set, achieving high-speed and accurate waveform template retrieval.

[0132] For example, in a highly dynamic game scenario, the compensation control module receives a 256-bit fixed-length binary code as the phase index key generated by S5.1. The template library contains 1024 records, each containing a pulse shape of 512 sampling points, 8 duty cycle sequences, and 16 voltage swing scalars. The system configures a hash function. The key value is mapped to record number 512 in the library. The reference pulse shape sampling array for this record defines precise rise time characteristics of 3μs and fall time characteristics of 2μs on the time axis, with a duty cycle sequence of [0.4, 0.5, 0.45, 0.55, 0.5, 0.6, 0.55, 0.5] and a voltage swing scalar range of [4.5V, 5.0V, 5.5V, 6.0V, 6.5V, 7.0V, 7.5V, 8.0V, 8.5V, 9.0V, 9.5V, 10.0V, 10.5V, 11.0V, 11.5V, 12.0V]. During the fast hash retrieval process, the 256-bit key value is processed and directly located to the aforementioned record, achieving instantaneous extraction of the reference data. The output initial drive waveform set has a complete structure, including accurate waveform sampling points, duty cycle and voltage amplitude, and has the ability to provide a consistent benchmark for environmental correction algorithms. In the verification, this set can significantly improve the matching accuracy and response speed in the drive signal synthesis stage.

[0133] S5.3: Collect the real-time reading of the panel temperature sensor of the current LCD module and the cumulative working time aging coefficient. Use a multivariate linear interpolation compensation algorithm to fuse and calculate the real-time reading of the panel temperature and the cumulative working time aging coefficient to generate a dynamic correction factor matrix that characterizes the current physical environment, so as to quantify the influence weight of environmental drift on the driving effect.

[0134] S5.4: Input the initial drive waveform set and the dynamic correction factor matrix, and use the adaptive gain scaling transformation technology to apply the dynamic correction factor matrix to the reference pulse shape data, reference duty cycle sequence data and reference voltage swing data, perform point-by-point amplitude correction and timing offset adjustment, and output the refined drive waveform parameter set after environmental characteristic adaptation.

[0135] Input the initial driving waveform set and the dynamic correction factor matrix, define the amplitude sampling point in each pulse shape data and each duty cycle segment in the duty cycle sequence as independent correction objects, and establish a mapping table between amplitude coefficient and timing offset.

[0136] The amplitude correction components corresponding to temperature drift and aging effect in the dynamic correction factor matrix are applied to the amplitude sampling points of the reference pulse shape data. The point-by-point correction value is generated using the adaptive gain scaling factor calculation function, which is in the form of a weighted product, where the weights are set according to the rate of change of environmental parameters.

[0137] A phase synchronization calibration algorithm is used to apply the timing offset correction component to the reference duty cycle sequence. Timing alignment is achieved by inserting or deleting micro-cycles in each duty cycle segment, while maintaining the phase continuity of the original waveform.

[0138] For voltage swing data, a combination of gain scaling and upper limit saturation logic is used to first calculate the target voltage gain after environmental correction, and then limit the voltage to ensure that it does not exceed the safety threshold of the drive circuit. The calculation formula is as follows: in The reference voltage swing, For the voltage gain component corresponding to the dynamic correction factor matrix, Maximum safe voltage.

[0139] The data, after amplitude correction, timing offset adjustment and voltage gain limiting, is integrated into a three-dimensional structure combining pulse shape, duty cycle sequence and voltage swing, and a refined set of drive waveform parameters adapted to the characteristics of the current physical environment is output.

[0140] By using adaptive gain scaling transformation and multi-factor correction processing, the initial drive waveform set from the previous step is transformed into refined drive waveform parameters that take into account the effects of environmental drift and aging characteristics, thereby achieving stable output of the drive signal in the actual working environment.

[0141] For example, in a machine with an operating temperature of Celsius, cumulative working hours In the Mini LED LCD display module, the amplitude sampling value of the reference pulse shape is... The average duty cycle sequence is volts. %, voltage swing reference value The amplitude gain component in the dynamic correction factor matrix is... The timing offset component is Micro-period, voltage gain component is Safe maximum voltage Volt. Applying the amplitude gain to the reference amplitude sample value yields the correction value. The time offset component is applied to the duty cycle sequence and adjusted to an average value while maintaining phase continuity. %; after calculating the voltage swing gain, we obtain The voltage is set and the limiting condition is met. The refined drive waveform parameter set of the output can significantly improve the stability of brightness compensation in actual environment and ensure that the drive circuit still maintains safe and efficient operation under high temperature aging conditions.

[0142] S5.5: Integrate the pulse width modulation signal and analog voltage level information in the refined drive waveform parameter group, use the digital-analog mixed signal synthesis engine to perform waveform reconstruction and smoothing filtering, and generate a preliminary drive signal that finally adapts to the current regional load change trend and eliminates environmental interference, so that the subsequent drive circuit can directly perform brightness compensation action.

[0143] The input includes a refined set of drive waveform parameters adapted to environmental characteristics, including pulse width modulation (PWM) signal data and analog voltage level data. The PWM signal data and analog voltage level data are phase-synchronized and merged in the time domain to generate a composite waveform sequence that maintains consistency between rising and falling edges. This merged composite waveform sequence is input to a mixed-signal synthesis engine. Multi-channel parallel interpolation is used to match the amplitude of the digital domain PWM signal and the analog domain voltage level at the sampling node, forming a continuous waveform transition segment. A low-pass smoothing filter module is configured in the synthesis engine based on a preset filter coefficient matrix to perform frequency domain suppression processing on the continuous waveform transition segment, reducing high-frequency noise components and mitigating environmental electromagnetic interference. The filtered waveform is output to a waveform reconstruction unit, where a template-matching-based envelope recovery algorithm fine-tunes the overall waveform shape to ensure that the waveform amplitude and timing characteristics fully meet the requirements of the drive circuit. The reconstructed waveform is compared with the original characteristics of the refined drive waveform parameter set for correlation verification. After confirming compliance with the brightness compensation control logic requirements, a preliminary drive signal that adapts to the current area's load variation trend and eliminates environmental interference is generated. By synthesizing and smoothing mixed-signal processing, the refined driving waveform parameter set from the previous step is transformed into a preliminary driving signal that can be directly applied to the driving circuit, thereby achieving efficient brightness compensation based on trend prediction.

[0144] For example, in a Mini LED LCD display module, the pulse width modulation signal of the refined drive waveform parameter group is set to the duty cycle. ,cycle Milliseconds, peak analog voltage level set to The mixed-signal synthesis engine employs a four-channel parallel processing architecture, with a sampling rate of [missing information - likely a specific value]. For kilohertz, the interpolation operation uses the cubic spline interpolation formula: ,in In the interval The cubic polynomial interpolation result is obtained, with interpolation coefficients determined by the PWM duty cycle distribution and the peak voltage level. The cutoff frequency of the low-pass filter module is set to... kilohertz, the filter coefficient matrix is ​​generated by the Chebyshev II filter design, and the passband ripple amplitude does not exceed Decibels. The waveform reconstruction unit uses an envelope recovery algorithm to adjust the slope of the pulse rising edge to... Volts / microseconds and maintain the trail duration as Milliseconds. When the output initial drive signal was verified in the drive circuit simulation, its correlation coefficient with the target waveform was significantly improved, the waveform distortion was significantly reduced, and it was able to stably perform brightness compensation actions within a millisecond response time.

[0145] Step S6: Apply the initial driving signal to the corresponding sub-region driving circuit of the liquid crystal display module to perform brightness compensation, so as to achieve a trend-predictive drive parameter feedforward response within a millisecond timescale when the regional load changes rapidly, eliminating the compensation delay caused by the traditional closed-loop feedback mechanism. Specifically, it includes: S6.1: Obtain the initial drive signal containing the pulse shape, duty cycle sequence and voltage swing combination, as well as the current panel temperature and aging coefficient environmental parameters. Use the timing synchronization alignment algorithm to perform frame period phase calibration processing on the initial drive signal to generate a synchronized drive command stream with strict timing reference, ensuring that subsequent execution actions and display scan line addresses are accurately matched.

[0146] The initial drive signal input includes three types of waveform parameters: pulse shape, duty cycle sequence, and voltage swing combination, as well as environmental data on the aging coefficient corresponding to the current panel temperature and cumulative working time.

[0147] The above preliminary driving signal is sampled using a high-precision clock source and a frame period index matrix is ​​generated. The waveform sampling points within the frame period are then mapped to the time reference coordinate system corresponding to the display scan line address.

[0148] Based on the dynamic correction factor generated by the panel temperature and aging coefficient, the peak position of the waveform and the pulse width in the frame period index matrix are adjusted to eliminate the influence of the physical environment on the synchronization characteristics.

[0149] A timing synchronization alignment algorithm is used to compare the phase of the adjusted waveform sampling points with the horizontal scanning trigger signal of the display controller, calculate the phase offset, and complete the phase calibration using the following formula: in As a reference phase standard, To measure the phase of the waveform.

[0150] The calculated phase offset is applied to the pulse shape data and duty cycle sequence, and the waveform data points are adjusted by interpolation resampling to ensure strict alignment with the line scan timing.

[0151] Synchronization instruction encapsulation is performed on the adjusted waveform and voltage swing combination, converting each drive signal cycle into a unique synchronized drive instruction stream and caching it to a high-speed storage unit for direct call by subsequent execution modules.

[0152] By using a timing synchronization alignment process, the initial drive signal from the previous step is transformed into a synchronized drive instruction stream with a strict timing reference, thereby achieving precise matching between the drive execution action and the display scan line address.

[0153] For example, in a Mini LED display module with a resolution of 3840×2160 and a frame rate of 240Hz, the input initial drive signal pulse width is 4.2μs, the duty cycle is 45%, the voltage swing is 3.8V, the panel temperature is 42℃, and the aging factor is 0.87. The reference period length of the sampling timing index matrix is ​​4.167ms. The current waveform phase is measured through phase comparison. 35°, reference phase The angle is 38°. Calculate the phase offset using the formula. The offset is applied to the waveform data, and the pulse peak position is advanced to match the line scan trigger time. The resampled synchronized drive command stream has strict frame period alignment characteristics, maintaining a phase fluctuation of less than 0.5° at a frame rate of 240Hz, ensuring that the subsequent drive circuit loading process achieves high-precision brightness compensation response within the millisecond time scale of regional load changes.

[0154] S6.2: Receive the synchronized drive command stream and extract the voltage swing combination data from it. Perform nonlinear predistortion mapping processing based on the digital-to-analog conversion characteristic curve of the source driver to generate digital grayscale voltage codewords that correct the transmission link loss, providing a high-precision digital quantity control basis for physical level output.

[0155] S6.3: Input digital grayscale voltage codeword and duty cycle sequence data, and use high-speed pulse width modulation signal synthesis logic to perform multi-channel parallel waveform reconstruction processing to generate an analog drive voltage waveform with a specific rising edge slope and holding time, so as to realize the physical electrical signal reproduction of the load change trend of the area to be loaded.

[0156] S6.4: Obtain the analog driving voltage waveform and combine it with the impedance matching parameters of the sub-region driving circuit. Then, perform charge injection rate optimization processing through the gate conduction timing control unit to generate a transient electric field excitation signal that matches the response characteristics of liquid crystal molecules, thereby eliminating the liquid crystal deflection hysteresis effect caused by sudden changes in regional load.

[0157] The analog drive voltage waveform with a specific rise slope and hold time generated in S6.3 is obtained. The built-in impedance parameter register is called to obtain the real-time impedance matching parameters of the corresponding sub-region drive circuit, forming a dataset that correlates waveform amplitude and impedance characteristics.

[0158] The amplitude data of the simulated drive voltage waveform and the impedance characteristics are correlated and input to the gate turn-on timing control unit. The turn-on phase interval and turn-off delay interval are determined according to the impedance matching condition, and the turn-on timing baseline matrix is ​​constructed.

[0159] An optimization algorithm for charge injection rate is applied to the conduction timing baseline matrix to achieve adaptive adjustment of charge mobility by controlling the combination of width and amplitude of gate drive pulses.

[0160] A transient electric field calculation model is used to calculate the transient electric field amplitude parameters based on the optimized charge injection rate and the response time constant of the liquid crystal molecules to the electric field intensity. The formula is as follows: in, The transient electric field intensity, This is the optimized peak driving voltage. The electric field efficiency coefficient is obtained by the charge injection rate optimization algorithm. The thickness is the liquid crystal layer.

[0161] The transient electric field excitation waveform is generated based on the calculated transient electric field amplitude parameters. By superimposing a preset rise segment and decay segment through a waveform modulator, the deflection of liquid crystal molecules under high dynamic load change scenarios can quickly follow the predicted trend.

[0162] Through the above processing method, the analog driving voltage waveform of S6.3 is converted into a transient electric field excitation signal, thereby achieving the adaptation of the liquid crystal molecule response characteristics and eliminating the liquid crystal deflection hysteresis effect caused by sudden changes in regional load.

[0163] For example, in a high-speed game switching scenario using a Mini LED backlit LCD display module, the impedance matching parameters of the acquisition sub-region driving circuit are 5.2Ω, the liquid crystal layer thickness is 4.5μm, the peak value of the analog driving voltage waveform generated by S6.3 is 12.6V, and the rise time slope is 0.85 V / μs. The electric field efficiency coefficient is 0.93 obtained by optimizing the charge injection rate through the gate conduction timing control unit. The electric field strength is calculated as follows: The output value is approximately 2.604 V / μm, corresponding to the transient electric field excitation signal being maintained for 5μs in the gradual rise phase and 8μs in the decay phase. The response time of liquid crystal molecule deflection is reduced to approximately 60% of that in the original environment, and the visual delay of brightness compensation is significantly reduced, ensuring the continuity and stability of brightness changes in high dynamic range images.

[0164] S6.5: Apply a transient electric field excitation signal to the corresponding sub-region pixel electrode of the liquid crystal display module, and use the principle of liquid crystal dielectric anisotropy to perform dynamic adjustment of light transmittance, so as to complete the brightness compensation output based on trend prediction within the millisecond time scale when the regional load changes rapidly, and completely eliminate the visual compensation delay caused by the traditional closed-loop feedback mechanism.

[0165] The input conditions include the transient electric field excitation signal after charge injection rate optimization, the physical position index of the pixel electrode of the corresponding sub-region of the liquid crystal display module, and the dielectric anisotropy coefficient of the liquid crystal material.

[0166] Electrode mapping processing is performed on the transient electric field excitation signal. Based on the two-dimensional coordinate matrix of the pixel electrodes in the sub-region, the spatial distribution of the signal is mapped point by point to the electrode array address, so as to achieve accurate loading of electric field excitation in the spatial domain.

[0167] Before applying the signal to the electrode, the response time of the orientation change of the liquid crystal molecules is predicted using a dielectric anisotropy model. The prediction results are used as the basis for adjusting the amplitude and duration of the electric field signal to ensure that the driving signal is synchronized with the rotation process of the liquid crystal molecules.

[0168] The change in target transmittance is calculated using the electric field transmittance dynamic adjustment formula: in, The deflection angle change is calculated from the change in the deflection angle of liquid crystal molecules under the action of an electric field. The formula for the deflection angle change is: in, Where is the dielectric constant. The anisotropy coefficient of the liquid crystal is... To apply electric field strength, The thickness of the liquid crystal layer. Where is the elastic constant. The current temperature. This is the critical temperature.

[0169] The above deflection angle calculation results are converted into light transmittance change values. The output voltage and scanning time are adjusted in real time by the drive circuit to achieve millisecond-level dynamic adjustment of light transmittance.

[0170] The loading process is monitored in a closed loop. The real-time curve of transmittance change is collected by the optical sensing unit and compared with the expected transmittance target. The amplitude and duration of the electric field excitation signal are dynamically adjusted so that the transmittance change trajectory strictly follows the trend prediction result.

[0171] By using electrode mapping, dielectric response prediction, and real-time light transmittance adjustment, the transient electric field excitation signal from the previous step is transformed into a precisely controllable dynamic light transmittance adjustment index, thereby achieving brightness compensation output that eliminates visual compensation delay.

[0172] For example, on a liquid crystal display module with a resolution of 3840×2160, a sub-region driving circuit with a diagonal size of 55 inches is selected, and the liquid crystal layer thickness is configured as follows: The dielectric anisotropy coefficient is set to elastic constant The critical temperature is configured as follows The applied electric field strength is set to... The deflection angle of the liquid crystal molecules is calculated using the deflection angle formula. Corresponding transmittance change value After electrode mapping loading, the error between the real-time light transmittance curve and the predicted trend target is within... Within milliseconds, the brightness of the display panel is significantly improved in scenarios with rapidly changing regional loads, eliminating visual compensation delay and enhancing display uniformity.

[0173] Step S7: Set a verification window at each preset frame number, collect the feedforward output result and compare it with the actual brightness data detected by the closed-loop feedback, calculate the deviation value between the two, and determine whether the deviation value exceeds the set stability threshold to identify whether there is a cumulative error in the feedforward prediction. The feedforward output result refers to the preliminary driving signal (or driving parameters) directly predicted and generated by the phase mapping neural network model based on the input video content features (temporal-spatial coupling feature vector), without correction by the actual brightness feedback of the current frame.

[0174] Specifically, it includes: S7.1: Perform modulo operation on the frame counter inside the display timing controller to generate a periodic verification window trigger signal. When the frame count value reaches the preset verification interval threshold, lock the current frame as the verification reference frame, and synchronously collect the expected brightness distribution map corresponding to the preliminary drive signal generated by the feedforward output module in this frame and the actual brightness distribution map fed back by the light sensor array, forming a dual-channel verification data pair containing the expected brightness distribution map and the actual brightness distribution map with spatiotemporal coordinate alignment.

[0175] The frame counter register inside the display timing controller is read to obtain the current accumulated frame count value as the time base signal input.

[0176] The time reference signal is input to the modulo operation processing unit, and an integer division operation with a preset verification interval threshold as the modulus is used to generate a periodic verification window trigger signal. The modulus is set according to the matching relationship between the display refresh rate and the target verification frequency.

[0177] The periodic verification window trigger signal is sent to the frame locking logic unit to perform a locking operation on the frame in the current frame sequence that is at the trigger period position, forming a verification reference frame index with a unique timing identifier.

[0178] Based on the verification reference frame index, the preliminary drive signal data packet generated by the frame is extracted in the feedforward output buffer module, and the mapping rendering unit is called to convert the drive signal into the expected brightness distribution map. This conversion process uses the panel drive characteristic curve and the correction coefficient matrix to complete the calculation of the light output value.

[0179] Simultaneously, within the closed-loop feedback acquisition subsystem, the synchronous scanning command of the light sensor array is invoked to perform pixel-by-pixel brightness measurement on the display screen corresponding to the calibration reference frame. Based on the system's row and column scanning sequence, the measurement results are remapped in coordinates and aligned in time sequence to generate an actual brightness distribution map.

[0180] The expected brightness distribution map and the actual brightness distribution map are respectively input into the dual-channel verification data construction unit. The pixel positions and timestamps of the two are completely matched by the spatial index mapping algorithm, and the dual-channel verification data pair containing spatiotemporal coordinate alignment information is output.

[0181] Through the above chain processing method, the execution result of the driving signal in the previous step is transformed into dual-channel brightness data with temporal consistency and spatial comparability, ensuring the input accuracy of subsequent deviation calculation and realizing the location and quantification of the cumulative error of feedforward prediction.

[0182] For example, in a Mini LED LCD display module with a refresh rate of 120Hz, the cumulative value of the frame counter is 1024, the preset verification interval threshold is set to 240 frames, and a modulo operation is performed on the frame counter value. The remainder is 64. The trigger signal is generated when the remainder equals 0. Locking is not triggered in this cycle. The corresponding frame is locked as the reference frame when the remainder returns to zero. The feedforward output buffer module is called to base the drive signal of this frame on the panel drive curve. Converted to the expected brightness distribution, where and These represent the slope and intercept components of the panel correction coefficient matrix, respectively. The closed-loop acquisition unit performs light intensity measurement with an integration time of 1 ms per pixel and outputs a matrixed actual brightness distribution. The spatial index mapping algorithm uses pixel coordinates ( , The key is to match the two image matrices one by one to generate dual-channel verification data pairs. Through this process, the module can complete the synchronous acquisition of dual-channel luminance data of the reference frame within 2ms in a 120Hz high dynamic range scene, providing high-precision input for subsequent deviation vector calculation and significantly improving the accuracy of cumulative error identification.

[0183] S7.2: Based on the expected brightness distribution map and the actual brightness distribution map in the dual-channel verification data pair, perform pixel-by-pixel grayscale difference calculation and regional weighted averaging to eliminate local noise interference and extract global brightness deviation features, and generate a comprehensive brightness deviation vector that characterizes the degree of loss in feedforward prediction accuracy. This vector quantifies the amplitude difference and temporal phase shift between the trend prediction driving parameters and the actual load demand.

[0184] The dual-channel verification data pair already formed in S7.1 is obtained, and the expected brightness distribution map and the actual brightness distribution map are strictly registered in the pixel coordinate space to ensure the spatial consistency of the difference operation.

[0185] The pixel-by-pixel grayscale difference calculation module is invoked. For each pair of corresponding pixels, the brightness value is subtracted to form a grayscale difference matrix containing the full resolution. The brightness value is calculated in the quantized grayscale level.

[0186] The region weighted average processing method is adopted. The pixel differences in the gray-level difference matrix are locally aggregated according to the preset weight mask. The weight mask is calculated based on the importance factor of the sub-region and the visual center weight coefficient, thereby eliminating the interference introduced by local noise or fluctuations in non-critical regions.

[0187] By using a signal denoising filter to perform spatial smoothing and high-frequency suppression operations on the weighted grayscale deviation data, the deviation outliers caused by occasional brightness spikes are further eliminated, so that the deviation data can maintain continuity in both spatial and temporal dimensions.

[0188] The filtered grayscale deviation data is input to the deviation vector generation module, which encodes the amplitude and phase information of the sub-region deviation, and quantifies the difference between the trend prediction driving parameters and the actual load demand through the following comprehensive deviation vector construction formula: in, For the first Brightness amplitude deviation in individual sub-regions This is the phase offset. Given the total number of sub-regions, the comprehensive brightness deviation vector is obtained by summing the results using operators and normalizing the summation by the number of regions. .

[0189] Through the above processing method, the dual-channel verification data pair from the previous step is transformed into a comprehensive brightness deviation vector that can fully reflect the difference in brightness amplitude and the temporal phase shift, thereby achieving a quantitative description of the degree of loss in feedforward prediction accuracy.

[0190] For example, in a liquid crystal display module with a resolution of 3840×2160, the range of the difference matrix between the expected brightness distribution map captured by the verification window and the actual brightness distribution map after registration is as follows: The weight mask is set to 0.08 for the central region and 0.03 for the edge region, within a range of 12 to +15 gray levels, to highlight differences in visually sensitive areas. After weighted averaging, the weighted difference in the central region is approximately 8.7 gray levels, and the weighted difference in the edge region is approximately 4.1 gray levels. A 3×3 Gaussian kernel is used for spatial smoothing with a standard deviation of 1.2, effectively eliminating high-frequency noise from isolated pixels. Amplitude deviation and phase shift are encoded as numerical pairs, for example, an amplitude deviation of 8.7 gray levels and a phase shift of 0.45 milliseconds. These are input into a formula to calculate the overall brightness deviation vector. When the number of sub-regions n is 12, the overall brightness deviation vector value is... The results indicate that within the current verification window, there is a significant amplitude and phase deviation between the trend prediction driving parameters and the actual load requirements. Subsequently, S7.3 will determine whether to trigger the online fine-tuning mechanism based on this, thereby significantly improving the brightness uniformity and compensation stability in high dynamic display scenarios.

[0191] S7.3: Using a stability threshold configuration table pre-stored in non-volatile memory, perform multi-dimensional boundary comparison analysis on the comprehensive brightness deviation vector to identify whether the deviation component exceeds the allowable dynamic fault tolerance range, and generate a stability judgment result containing an out-of-tolerance flag and an error level code. This result clearly indicates whether the current lightweight phase mapping neural network model has a cumulative prediction error that leads to a decrease in display quality.

[0192] S7.4: Based on the out-of-tolerance flag status in the stability determination result, execute conditional branch logic judgment. If the flag indicates that there is accumulated error, generate a model online fine-tuning enable instruction. If the flag indicates that it is normal, generate a model parameter hold instruction. This outputs the final verification decision signal used to control the subsequent model update process, ensuring that the computationally intensive weight correction process is only started when necessary to maintain system real-time performance.

[0193] S7.5: Based on the final verification decision signal, the comprehensive brightness deviation vector within the historical verification period is subjected to sliding window statistical archiving processing to construct a long-term error evolution trend dataset, and an error evolution trend report is generated to evaluate the aging characteristics of the display module and the impact of environmental drift. This provides data support for the subsequent dynamic adjustment strategy of the stability threshold configuration table, and realizes the adaptive long-term stable operation of the compensation control module.

[0194] Step S8: If the deviation value exceeds the stability threshold, an online model fine-tuning mechanism is triggered to correct and update the weight parameters of the lightweight phase mapping neural network model using the deviation data within the current verification window; otherwise, the existing model parameters are maintained to ensure the accuracy of the phase encoding of the driving parameters and the system's energy efficiency robustness during long-term operation. Specifically, this includes: S8.1: Obtain the comprehensive brightness deviation vector and the corresponding quantized spatiotemporal coupling feature vector tensor generated within the verification window. Use the gradient backpropagation calculation algorithm to perform partial derivative calculation of the loss function on the comprehensive brightness deviation vector to generate a weight update gradient matrix that characterizes the prediction error distribution of the current lightweight phase mapping neural network model, serving as the core driving basis for model parameter correction.

[0195] S8.2: Receive the weight update gradient matrix and combine it with the pre-stored momentum history state data. Use the adaptive moment estimation algorithm to perform first-order moment and second-order moment statistics fusion processing on the weight update gradient matrix to generate a standardized parameter adjustment step size vector containing the learning rate dynamic scaling factor, thereby eliminating the convergence oscillation problem caused by the difference in gradient scale between different network layers.

[0196] S8.3: Based on the standardized parameters, adjust the step size vector and the integer precision model parameter set currently residing in the cache, and use fixed-point differential update logic to perform layer-by-layer weight offset accumulation operation to generate a new set of weight parameters after error compensation correction, ensuring that the numerical precision of the model in the embedded coprocessor storage unit remains consistent and that no overflow truncation occurs.

[0197] S8.4: Perform a non-volatile memory atomic write operation on the new copyright weight parameter set, and use the integrity check code generation algorithm to perform cyclic redundancy check processing on the written data to generate a persistent model version file with power failure protection capability, so as to prevent the model parameters from being damaged due to power fluctuations or abnormal interruptions and affecting the normal operation of the display timing controller.

[0198] S8.5: Load the persistent model version file, and use the hot-switching mechanism to replace the running phase mapping neural network model instance (i.e., the old version) without interrupting the frame synchronization triggering interruption. Complete the reconstruction of the inference context to generate a real-time driving parameter phase encoding output stream with the latest error correction capability, and finally realize the adaptive evolution and energy efficiency robustness maintenance of the lightweight phase mapping neural network model in the long-term operation process.

[0199] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0200] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the element or object preceding “comprising” or “including” encompasses the element or object listed following “comprising” or “including” and its equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0201] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for brightness compensation based on regional load differences in a liquid crystal display module, specifically comprising: S1: To address the regional load abruptness characteristics of LCD modules under different high dynamic display scenarios, collect the spatiotemporal coupling feature vectors of each sub-region in a typical content sequence, and associate and label them with the actual load jump data of the corresponding sub-region in subsequent frames to construct a regional load dynamic fingerprint database. S2: Based on the spatiotemporal coupling feature vector in the regional load dynamic fingerprint database and the actual load jump data, train a phase mapping neural network model and establish a nonlinear mapping relationship from the spatiotemporal coupling feature vector to the phase encoding of the driving parameters; S3: Quantize the trained phase-mapped neural network model to integer precision and deploy it to the dedicated coprocessor of the display timing controller; S4: The phase mapping neural network model receives the spatiotemporal coupling feature vector output by the current frame image analysis module, performs inference calculations, and outputs the phase encoding of the driving parameters; S5: Based on the phase encoding of the driving parameters, call the pre-stored driving waveform template library, match the pulse shape, duty cycle sequence and voltage swing combination corresponding to the phase encoding of the driving parameters, and fine-tune it in combination with the current panel temperature and aging coefficient environmental parameters to generate a preliminary driving signal that adapts to the current load change trend. S6: Apply the initial driving signal to the corresponding sub-area driving circuit of the liquid crystal display module to perform brightness compensation.

2. The brightness compensation method for regional load differences in a liquid crystal display module according to claim 1, characterized in that, Following step S6, steps S7 and S8 are also included: S7: Set a verification window for each preset number of frames, collect the feedforward output results and compare them with the actual brightness data detected by the closed-loop feedback, calculate the deviation value between the two, and determine whether the deviation value exceeds the set stability threshold. S8: If the deviation value exceeds the stability threshold, the weight parameters of the phase mapping neural network model are corrected and updated using the deviation data in the current verification window; If the deviation value does not exceed the stability threshold, the existing model parameters are maintained.

3. The brightness compensation method for regional load differences in a liquid crystal display module according to claim 1, characterized in that, The spatiotemporal coupling feature vector includes: grayscale distribution entropy value, spatial gradient magnitude, and temporal brightness change rate.

4. The brightness compensation method for regional load differences in a liquid crystal display module according to claim 1, characterized in that, The actual load jump data includes: the magnitude and direction of the actual load jump.

5. A brightness compensation method for regional load differences in a liquid crystal display module according to claim 1, characterized in that, The phase encoding of the driving parameters includes characteristic values ​​of three orthogonal dimensions: adjustment initiation timing, adjustment intensity slope, and adjustment decay rhythm.

6. The brightness compensation method for regional load differences in a liquid crystal display module according to claim 1, characterized in that, The phase mapping neural network model is used to instantaneously generate the corresponding driving parameter phase code based on the input spatiotemporal coupling feature vector before each frame of image enters the compensation control stage.

7. A brightness compensation method for regional load differences in a liquid crystal display module according to claim 1, characterized in that, Step S2 specifically includes: Tensor reconstruction processing is performed on the spatiotemporal coupled feature vectors in the regional load dynamic fingerprint database to construct a multidimensional input feature matrix; The load jump amplitude and load jump direction in the actual load jump data are discretized and encoded to generate a standardized target phase label sequence. A phase mapping neural network model with a multilayer perceptron architecture is constructed based on the input feature matrix and the target phase label sequence. The network weight parameters are iteratively optimized using the backpropagation algorithm to obtain the initial phase mapping model. The output layer activation function of the initial phase mapping model is subjected to constraint mapping processing, and the continuous numerical output is forcibly mapped to the decoupled space composed of three orthogonal dimensions: adjustment start timing, adjustment intensity slope, and adjustment decay rhythm, to generate the original driving parameter phase code that represents the combined state of the driving parameter adjustment action. Based on the correlation between the original driving parameter phase encoding and the actual load jump data, a generalization capability verification process is performed to obtain the phase mapping neural network model.

8. A brightness compensation method for regional load differences in a liquid crystal display module according to claim 5, characterized in that, Step S4 specifically includes: The spatiotemporal coupling feature vector output by the current frame image analysis module is obtained, and an integer quantization preprocessing operation is performed on the spatiotemporal coupling feature vector to generate a quantized spatiotemporal coupling feature vector tensor that is adapted to the computing power constraints of the phase mapping neural network model. Based on the quantized spatiotemporal coupling feature vector tensor, the phase mapping neural network model is invoked to perform forward inference operations, extract the implicit correlation between the regional load evolution law and the driving response characteristics, and output the phase-encoded latent variables of the driving parameters. The phase-encoded latent variables of the driving parameters are subjected to multi-dimensional decoupling mapping processing, and the original values ​​are mapped to time offset, voltage change rate and signal tail coefficient according to the predefined phase encoding rules. Based on the time offset, voltage change rate, and signal tailing coefficient, timing alignment and format encapsulation operations are performed to obtain the phase encoding of the driving parameters.

9. A brightness compensation method for regional load differences in a liquid crystal display module according to claim 8, characterized in that, The time offset, voltage change rate, and signal tailing coefficient are used to characterize the adjustment start timing, adjustment intensity slope, and adjustment attenuation rhythm, respectively.

10. A brightness compensation method for regional load differences in a liquid crystal display module according to claim 9, characterized in that, Step S5 specifically includes: The feature values ​​of the adjustment initiation timing, adjustment intensity slope, and adjustment decay rhythm are jointly encoded using a multidimensional vector space mapping algorithm to generate a phase index key value with a unique identifier. The phase index key value is received, and a fast matching operation is performed in the locally pre-stored drive waveform template library based on the hash lookup mechanism to extract the reference pulse shape data, reference duty cycle sequence data and reference voltage swing data that are completely corresponding to the phase index key value, and to construct an initial drive waveform set containing basic timing and amplitude information. The system collects real-time readings of the panel temperature sensor of the current liquid crystal display module and the cumulative working time aging coefficient, and performs a fusion calculation on the real-time panel temperature readings and the cumulative working time aging coefficient to generate a dynamic correction factor matrix characterizing the current physical environment. Input the initial drive waveform set and the dynamic correction factor matrix, apply the dynamic correction factor matrix to the reference pulse shape data, reference duty cycle sequence data and reference voltage swing data, perform point-by-point amplitude correction and timing offset adjustment, and output a refined drive waveform parameter set; The pulse width modulation signal and analog voltage level information in the refined drive waveform parameter group are integrated, and waveform reconstruction and smoothing filtering are performed to generate the initial drive signal that finally adapts to the load change trend of the current area.

Images