A multi-display mode driving control method for COG liquid crystal module

Abstract

The application discloses a multi-display mode driving control method for a COG liquid crystal module, and belongs to the technical field of display driving, and comprises the following steps: dividing a current display picture into a plurality of display sub-region blocks, and constructing a block description vector; establishing a short-time reference block description and a background reference block description, and determining a short-time change amount and a background deviation amount; extracting a time sequence change feature; constructing a region-level hidden state recursive model, and obtaining a region hidden state probability; determining a background update coefficient, and updating the background reference block description; extracting a picture-level aggregation amount, and determining a current corresponding picture state; selecting a target driving action from candidate driving actions, and generating a driving parameter; collecting an execution feedback amount, and updating model parameters and function weights. Through the construction of a double reference memory and a region hidden state recursive model, the application achieves the problem of fine determination of local and periodic dynamic pictures, and solves the problems of coarse screen determination granularity and insufficient local change recognition in the prior art.

Description

Technical Field

[0001] This invention relates to the field of display driving technology, and in particular to a multi-display mode driving control method for COG liquid crystal modules. Background Technology

[0002] Currently, display devices are increasingly widely used in portable, embedded, and smart terminals. COG (Chip On Glass) liquid crystal modules have become the mainstream choice due to their high integration and compact structure. For example, COG liquid crystal modules can be used in display devices such as TFT-LCD displays, thin-film transistor liquid crystal displays, liquid crystal displays, TFT LCD screens, segment LCDs, and small-size liquid crystal display panels. As display content becomes increasingly rich and diverse, liquid crystal modules need to be able to adapt to different types of display scenarios to meet the requirements of visual effects and system performance.

[0003] Existing LCD driving systems analyze input video data or displayed image patterns through a timing controller, switching between normal display mode and self-refresh mode to control the driving mode of the display panel. They also manage the frame refresh process based on image changes, ensuring orderly operation of the display process in different display scenarios. For COG display modules using TFT-LCD or thin-film transistor liquid crystal display structures, the above process is typically completed by a display controller working in conjunction with the module driving circuit, controlling partial or full-screen refresh based on changes in the displayed content. Furthermore, existing systems, such as those in TFT LCD screens or small-sized LCD panels, typically detect image updates in the display signal and, in conjunction with clock and synchronization signals, uniformly schedule the scanning timing, data loading order, and driving cycle. They incorporate buffer units or internal storage units in the driving circuit to maintain the current display state when the image content does not change significantly, thus preserving and repeating display data. Some systems also identify areas of change in the displayed content and, combined with a preset driving strategy, select the appropriate refresh mode to complete driving control under different display modes, achieving stable operation of the LCD panel in various working scenarios.

[0004] For example, Chinese invention patent CN103903581B discloses a liquid crystal display device and its driving method, comprising: at least one or more source driver ICs for driving multiple data lines formed in a panel; a timing controller for generating a power control signal for changing the level of the driving voltage applied to the source driver ICs according to a pattern of an image output to the panel; and a driving voltage generator for generating a first driving voltage or a second driving voltage according to the power control signal to drive the source driver ICs, wherein the first driving voltage and the second driving voltage have different levels.

[0005] For example, Chinese invention patent CN104795031B discloses a display device and a method for driving the display device, comprising: a display panel including coupled data lines and scan lines; a data driver configured to supply data voltage to the data lines based on digital video data; a scan driver configured to supply scan signals to the scan lines; a timing controller configured to select a normal mode and a self-refresh mode based on a panel self-refresh enable signal, wherein in the normal mode, the data driver and the scan driver are driven at a first frame rate, and in the self-refresh mode, the data driver and the scan driver are driven at a second frame rate lower than the first frame rate; and a power supply configured to supply driving voltage to the data driver, the scan driver, and the timing controller and to transmit DC power supply voltage to their external source, wherein the transmission of DC power supply voltage is blocked during a blank period in the self-refresh mode.

[0006] The above-mentioned technology has at least the following technical problems: In existing technologies, multi-display mode driving systems receive display mode trigger information, determine the target display mode, and select the corresponding driving parameter group to control the frame rate, driving state, and image processing flow of the COG LCD module to achieve switching between different display modes. However, in actual display processes, the screen content does not always appear in a uniform, whole-screen manner, but often exhibits a state where local area changes coexist with a relatively stable overall image, such as cursor blinking, number scrolling, local icon changes, or small window content refresh. For such mixed display scenarios involving static images, dynamic images, and static backgrounds with local dynamics, existing technologies typically still rely primarily on the overall screen state for judgment, lacking an effective distinction between local changes and the overall screen state. This makes it difficult to fully identify local dynamic content such as small-area animations, cursor blinking, or number jumps, resulting in a coarse-grained determination of the current screen state and an inaccurate reflection of the true content characteristics in mixed display scenarios.

[0007] Furthermore, existing technologies typically rely solely on overall screen disturbances, simple inter-frame differences, or preset triggering strategies to switch between high and low frame rates and select modes when adjusting refresh rates or switching display modes. Because their judgment of screen changes is mostly at a general and comprehensive level, it is difficult to stably reflect the range of changes, the process of changes, and the differences in state over time in different areas. Therefore, the accuracy of the system's judgment of the current screen state is still limited, making it difficult to provide a stable and fine-grained control basis for switching multiple display modes and adjusting corresponding drive parameters. This, in turn, affects the display mode selection of the display system in mixed display content scenarios, making it difficult for the system to balance local dynamic response speed, overall screen stability, and energy efficiency optimization. Summary of the Invention

[0008] To address the technical problem in existing technologies where it is difficult to distinguish image states in complex scenes, thus affecting image stability, this invention provides a multi-display mode driving control method for COG liquid crystal modules. The technical solution is as follows: A multi-display mode driving control method for COG LCD modules is provided, the method comprising: The current display screen is acquired and divided into multiple sub-regions. A block description vector is constructed for each sub-region. A short-time reference block description and a background reference block description are established for each sub-region. Based on the block description vector at the current time, a first change and a second change are determined. Temporal change features are extracted based on the first and second changes. Based on the first and second changes, the temporal change features, and the state association information of neighboring sub-regions, a region-level hidden state recursive model is constructed to obtain the region hidden state probability corresponding to each sub-region. Background update coefficients are determined based on the region hidden state probabilities, and the corresponding background reference block descriptions are updated. Screen-level aggregation quantities are extracted based on the region hidden state probabilities, and the screen state corresponding to the current display screen is determined. Based on the screen state, a target driving action is selected from multiple candidate driving actions, and driving parameters are generated to control the display module to execute the display drive. After executing the display drive, execution feedback quantities are collected, and the parameters in the region-level hidden state recursive model and the cost function weights corresponding to the candidate driving actions are updated based on the execution feedback quantities.

[0009] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: 1. This invention provides a multi-display mode driving control method for COG liquid crystal modules. It performs block-level streaming description of the current display screen, establishes dual-reference memory, extracts temporal change features, and constructs a region-level hidden state recursive model. This enables probabilistic determination of the static steady state, transient change state, continuous dynamic state, and periodic dynamic state of each region block. This allows for high-precision determination of static images, dynamic images, and mixed display scenes where static backgrounds and local dynamics coexist. Simultaneously, it can capture minute local changes and periodic fluctuations in real time, providing the system with richer region-level feature information, reducing the false judgment rate, and improving the reliability of display content recognition. This effectively solves the problem in existing technologies where systems typically use the entire screen state as the judgment unit, lack sensitivity to local changes, and cannot fully recognize small-area animations, cursor blinking, or number jumps, resulting in coarse-grained screen state determination. 2. This invention constructs a screen-level state determination model based on regional-level hidden states, and combines a cost optimization mechanism to select display modes and driving parameters. Simultaneously, it performs closed-loop correction on execution feedback quantities, achieving comprehensive analysis of local area change characteristics, time-dimensional characteristics, and continuous / periodic dynamics. Furthermore, it can dynamically adjust refresh strategies and driving parameters according to the actual needs of different display modes, achieving a balance between screen response speed, display effect, and power consumption. This guides subsequent display mode selection and driving parameter adjustments, taking into account local dynamic response speed, overall screen stability, and energy efficiency optimization. Ultimately, it improves the display performance and user experience of COG LCD modules in complex mixed scenarios, effectively solving the problem in existing technologies where display systems cannot accurately determine the current screen state by combining regional change characteristics and temporal characteristics of the displayed content when driving refresh rates or switching modes. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 A flowchart of a method for driving and controlling multiple display modes of a COG liquid crystal module is provided in an embodiment of this application; Figure 2 A schematic diagram of the dual-reference memory mechanism provided in the embodiments of this application; Figure 3 This is a block diagram of a region-level hidden state recursive model provided in an embodiment of this application; Figure 4 This is a screen-level state determination diagram provided in the embodiments of this application; Figure 5 This is a schematic diagram of the control action selection and feedback closed loop provided in the embodiments of this application. Detailed Implementation

[0012] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that embodiments of this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure. In the description of embodiments of this disclosure, the term "comprising" and similar terms should be understood as open-ended inclusion, i.e., "including but not limited to". The term "based on" should be understood as "at least partially based on". The term "one embodiment" or "this embodiment" should be understood as "at least one embodiment". The terms "first", "second", etc., may refer to different or the same objects. To make the technical problems, technical solutions, and advantages of this invention clearer, a detailed description will be provided below in conjunction with the accompanying drawings and specific embodiments.

[0013] like Figure 1 The diagram shown is a flowchart of a method for driving and controlling multiple display modes of a COG liquid crystal module according to an embodiment of this application. The method includes the following steps: S1: Divide the entire frame of the current display into multiple manageable display sub-region blocks, and then construct a block region vector for each sub-region block. When generating the current display frame, the main controller divides the entire frame into several display sub-region blocks according to a fixed grid or UI (User Interface) control boundaries. For fixed block division, the current screen can be mapped into more than M regular blocks according to the display resolution; for areas with known UI control boundaries, such as when the system knows the current control layout, including the status bar area, temperature number area, icon area, menu button area, and pop-up area, the control area can be treated as an irregular block. In this embodiment, the system divides according to fully fixed blocks, with each block being 8×8 pixels in size. It should be noted that technicians can freely choose the block division method according to the actual situation during actual operation, and this embodiment does not impose any constraints on this. The system establishes a block division mapping table during initialization, which includes the ID of each block i, the coordinates of its upper left corner (x, y, y), and the coordinates of its upper left corner (x, y, y). i y i Width and height (w) i h i The block type and neighborhood relationships provide fundamental information for subsequent steps in processing temporal and spatial features. For each block, a block description vector V is constructed at the current time t. i In this embodiment, the vector (t) contains four types of features: mean brightness, grayscale or color dispersion, edge energy, and compact signature. First, the mean brightness μ within the block is calculated. i (t), the specific formula is as follows: , Where Y(x,y,t) represents the normalized brightness value of the pixel with coordinates (x,y) in the current block at time t, and its value ranges from [0,1] to N. i Let N be the total number of pixels within the block, satisfying N i ≥1 and a positive integer, in fixed block partitioning mode, N i =w i ×h i In the 8×8 fixed-block scenario of this embodiment, N i =64; When using irregular block partitioning, N i This represents the actual number of valid pixels within the irregular block. When building the block partitioning map, the system only retains regions containing at least one valid pixel, thus avoiding a denominator of 0. This mean reflects the overall brightness variation of the block. Next, the grayscale or color dispersion σ is calculated. i (t) reflects the dispersion and structural complexity of pixels within a block, helping to distinguish whether the current block's brightness change is an overall change or a change in the block's structure. It is expressed in variance form, and the specific formula is as follows: , Where, σ i The smaller (t) is, the more uniform the pixel distribution within the block; σ i The larger the value of (t), the more pronounced the brightness variations, edge structures, or local textures within the area. Since some display screens require color images, the variances of the three primary colors (Red, Green, and Blue) are calculated separately and then summed with weights. Those skilled in the art should understand and be able to implement this method; however, this embodiment will not provide a detailed explanation. Next, the edge energy E is calculated. i (t), used to perceive changes in local structure within a block, such as text, icon outlines, or number strokes, helps the system recognize number jumps, cursor appearance or disappearance, small icon changes, and local structure switching. It can be represented in the form of gradient energy, and the specific formula is as follows: , Among them, G x (x, y, t) represents the gradient response in the horizontal direction, G y (x, y, t) represents the gradient response in the vertical direction, which can be calculated using conventional image processing operators. Finally, the compact signature H of the generated block is... i (t) preferably employs CRC signatures, partial projection hashes, or color histogram signatures to provide lightweight content fingerprints and enhance the ability to distinguish between numbers, icons, or partial pattern variations. For example, a compact signature H can be generated using partial projection hashes. i(t) The system reads all pixel brightness or color values ​​within the current block according to a preset scanning order, forming a pixel sequence for the current block. This pixel sequence is then projected onto L preset projection templates to obtain L projection statistics. Each projection statistic is then compared to its corresponding threshold. If a projection statistic is greater than or equal to its threshold, that statistic is recorded as 1; otherwise, it is recorded as 0, thus generating a binary signature H of length L. i (t). The above four features together constitute the block description vector V. i (t), denoted as V1(t) = [μ i (t), σ i (t), E i (t), Hi(t)]. The system uses streaming computation to statistically process pixels within a block. That is, when the master controller outputs pixel data, it does not need to save the entire frame, but instead accumulates the total brightness, sum of squares, gradient, or structural response in real time, while simultaneously generating a hash or CRC signature. Block description vector V i After V(t) is formed, it is written to the current frame block description buffer, resulting in the description set of all blocks in the entire frame {V1(t), V2(t), ..., V...}. M (t)}, where M is the total number of blocks.

[0014] S2: Establish a dual-reference memory for each region block, extract region change features, and determine the region change status of the current sub-region block in the current frame. For example... Figure 2 The diagram shown is a schematic of the dual-reference memory mechanism provided in an embodiment of this application. Combined with the current block description vector V output above... i (t), the system maintains two types of reference values ​​for each block: short-time reference and slow background reference, where the short-time reference V i F (t-1) is used to represent the block state of the previous frame, aiming to reflect the visual information of the block in the previous frame, and is used to detect short-term disturbances such as instantaneous jumps, flickering, or digit flips; slow background reference V i B (t-1) is used to represent the background state of the block over a longer time scale. Its purpose is to determine whether the current block has truly broken free from the long-term background, thus avoiding misinterpreting short-term flickering as content updates. For each block, the system calculates the distances relative to both the short-term reference and the slow background reference to obtain the first change, i.e., the short-term change D. i F (t), and the second change, namely the background deviation D. i B (t), the specific formula is as follows: , , Where D( To describe the distance function, a combination of weighted Euclidean distance, Manhattan distance, and hash Hamming distance can be used to achieve multi-feature fusion. Hash Hamming distance is preferred as an example, where numerical statistical features are normalized using absolute difference, and compact signature features are normalized using hash Hamming distance. Its specific construction can be expressed as: , , , , , Where a and b represent two block description vectors to be compared, representing the same region block at different times or under different reference states, for example, in calculating short-time change D. i F When (t), a represents V. i (t), b represents V i F (t-1), where ω1 is the weighting coefficient for normalizing the mean difference in brightness, ω2 is the weighting coefficient for normalizing the difference in grayscale or color dispersion, ω3 is the weighting coefficient for normalizing the difference in edge energy, and ω4 is the weighting coefficient for normalizing the Hamming distance of the compact signature hash. The sum of the weights must be 1. R μ R is the normalized scale for the mean brightness. σ R is a normalized scale for grayscale or color dispersion. E Let R be the normalized scale for the edge energy, where R is the normalized scale for the edge energy. μ >0、R σ >0、R E >0, preferably taking the preset upper bound, full-scale value, or nominal maximum change scale corresponding to the characteristic difference, so that d μ (a,b),d σ (a,b),d E (a, b) falls within the interval [0, 1], where L is the hash length and Hamming represents the Hamming distance, used to measure the difference between two binary signatures. If CRC signatures are used, the hash distance can be simplified to a binary judgment: 0 for signatures that match and 1 for those that don't. This allows numerical statistical features to reflect the degree of change, while the signature reflects changes in content identity. It should be noted that the specific values ​​of ω1, ω2, ω3, and ω4 can be adjusted by technicians according to actual conditions. For example, setting ω1=0.25, ω2=0.15, ω3=0.4, and ω4=0.2, increasing ω1 when focusing more on brightness, and increasing ω3 or ω4 when focusing more on structural changes, is not a constraint in this embodiment.

[0015] During system initialization, in the first frame after the system display has stabilized, let V... i F (0) and Vi B (0) values ​​are all related to V i (0) are the same, and auxiliary state variables are established, such as a stable counter C. i stable Continuous Deviation Counter C i change and dynamic confidence level P i dyn This is used to guide slow background adaptive updates, ensuring that both types of references start from the initial frame. At each frame t, the system first obtains the current block description vector V. i (t), and then calculate the short-time change D. i F (t) and background deviation D i B (t), and provide a preliminary state interpretation. If D i F (t)>T F And D i B (t) <T B If D is a short-term fluctuation, then it is determined to be a short-term fluctuation, and the background is not updated in this case; if D i F (t)>T F And D i B (t)>T B If D is a true update candidate, then it is determined to be a genuine update candidate; i F (t) < T F And D i B (t) <T B If the background value is large, it is considered a stable background, and the background reference can slowly absorb the current value; if D... i F (t) < T F And D i B (t)>T B If T is found to be in a stable state after the update, then it is determined to be in a stable state. F For judging short-term changes, T B To set a background deviation threshold, the threshold can be a fixed value or adaptively set according to the block type, region location, or control category. For example, a fixed T value can be set. F =0.08, T B =0.12. Then update the auxiliary state variables; when the background stabilizes, C... i stable When C increases and continues to deviate from the background i change Increases, but does not increase or is reset to zero during short-term fluctuations. Update short-term reference V. iF (t) and slow background reference V i B (t), where the short-time reference can be directly replaced by the current value V. i (t), while the slow background reference needs to be updated adaptively according to the current state, which will be implemented later.

[0016] S3: For each region block, extract temporal variation features along the time dimension. Based on the current block description vector V output above... i (t), Current block short-term change D i F (t) and the current block background deviation D i B (t), the system constructs the activity tag A for each block. i (t), serving as the fundamental input for persistence, periodicity, and spatial diffusion analysis. When D i F (t)>T F Or D i B (t)>T B At that time, let A i (t) being 1 indicates that the current block is active in this frame; otherwise, A is... i When (t) is 0, it indicates that the current block is basically stable.

[0017] First, persistence features are extracted to determine whether a block remains active for multiple consecutive frames. In this embodiment, persistence C is calculated by statistically analyzing the activity percentage of the most recent k frames. i (t) is achieved, and the specific formula is as follows: , Where τ represents the time index of the most recent k frames, τ∈{max(1, t-k+1), ..., t}, k≥1 and is a positive integer, used to count the proportion of block activity, A i (τ) represents the activity flag of the i-th block in frame τ. It is set to 1 when the block is active in frame τ, and 0 otherwise. When C i When (t) is closer to 1, it indicates that the current block has been active for most of the last k frames; when C i When (t) is close to 0, it indicates that the change only occurs in a few frames, which is a transient jump. Besides the method described in this embodiment, the persistence feature can also be obtained by calculating and normalizing the number of consecutive active frames. The persistence features obtained by both methods have no impact on subsequent derivation and calculation, and technicians can freely choose according to actual conditions; this embodiment does not impose any constraints on this. Next, periodic features are extracted to identify regular dynamics such as cursor blinking, icon breathing, or periodic jumps in local numbers. Periodic feature R i(t) can be obtained from the nearest sliding window of length W, for D i F The (t) sequence is obtained by performing autocorrelation or frequency domain analysis, and the specific formula is as follows: , Where Cor represents the correlation coefficient, and W is a positive integer with a value ≥ 2 and covers at least one complete cycle, typically preferably 1.5 to 2 complete cycles. For example, for fast local periodic dynamics, W is 8 to 32 frames; for low-to-medium frequency flickering or breathing-like dynamics, W is 48 to 128 frames. τ represents the time delay, is a positive integer, and 1 ≤ τ. min <τ max <W. When the variance of the window involved in the calculation is 0, the correlation value is preferably taken as 0. Blocks with high periodicity indicate that their activity exhibits a stable rhythm, while blocks with low periodicity may be random perturbations or one-off switching. Finally, spatial diffusion features are extracted to measure whether the activity of a block is isolated or belongs to the expansion and change of a larger region. Spatial diffusion feature S i (t) can be calculated through the proportion of activity in the neighborhood. For the i-th block, a neighborhood set N(i) is defined. In this embodiment, an 8-neighborhood is used, namely the eight blocks above, below, left, right, upper left, lower left, upper right, and lower right of the i-th block. The calculation is performed on these eight blocks: , Where j represents the index of each adjacent region block in the neighborhood set N(i). The more active blocks within a neighborhood, the better S... i The larger the value of (t), the more regional the change in the current block becomes; if the proportion of activity in the neighborhood is low, the change in the block may be an isolated point disturbance or noise. It should be noted that when the i-th block is a boundary block or corner block, the number of neighborhood blocks is less than 8. In this case, the neighborhood set is determined by the actual number of adjacent blocks and normalized to the actual number of neighborhoods. The system uses the previously calculated short-term change D of the current block... i F (t), Current block background deviation D i B (t), and the persistence feature C calculated in this step. i (t), periodicity characteristic R i (t) and spatial diffusion characteristics S i (t) Packed into block observation vector Y i (t), denoted as: , S4: Construct a region-level hidden state recursive model based on the block observation vector Y from the previous steps. i (t) Determine the hidden state of the current sub-region block. For example... Figure 3The diagram shown is a block diagram of a region-level hidden state recursive model provided in an embodiment of this application. For each region block i, a hidden state vector X is defined. i (t), denoted as: , Where, x i S (t) represents the static steady-state intensity, which corresponds to D in this embodiment. i F (t) small, D i B (t) Cases with small size, low persistence, low periodicity, and low spatial diffusion, including typical background blocks, stable text blocks, and stable icon blocks; x i T (t) represents the intensity of transient changes, reflecting the degree of short-term jumps or transient disturbances in the block. In this embodiment, it corresponds to situations where the current frame changes significantly but does not deviate from the background for a long time, has low persistence, and weak spatial diffusion, such as cursor flashing, a single number flip, a frame jitter, and refresh transition frames; x i C (t) represents the intensity of the continuous dynamic state, reflecting the degree to which the block remains active across multiple consecutive frames. In this embodiment, it corresponds to areas with high persistence, significant background deviation, and potential linkage with neighboring regions, such as local animations, menu expansion, or scrolling areas; x i P (t) represents the intensity of the periodic dynamic state, reflecting the periodic dynamic activity of the block. In this embodiment, it corresponds to a high periodicity index and a stable rhythmic change, such as cursor blinking, breathing light animation, or periodic digital refresh. The block observation vector Y... i (t) is used as input, and a recursive model of the hidden state of the region is constructed based on these inputs. The hidden state of each block is recursively updated, and the specific formula is as follows: , The first part is A×X i (t-1) represents the state continuation term, indicating the inertial continuation of the block's state from the previous time step to the current state. For example, a statically stable block usually remains statically stable in the next frame, and continuous dynamic blocks and periodically dynamic blocks also continue their inertia. A represents the state continuation matrix, used to control the degree of historical preservation of the region's state; the second part is B×Y. i (t) represents the current observation driver, used to apply the observation evidence from the current frame to the state update, such as D. i F When (t) is large, the value of the transient state vector component increases, C i (t) increases the value of the continuous dynamic state vector component when it is high, R i(t) When the value of the periodic dynamic state vector component is high, the value of the static steady state vector component is increased; when the overall change is small, the value of the static steady state vector component is increased. B represents the observation driving matrix, which is used to control the block observation vector Y. i (t) Driving ability for hidden state updates; Part III Γ×∑ j∈N(i) ω ij ×(X j (t-1)-X i (t-1)) represents the neighborhood spatial coupling term, used to reflect the spatial consistency of the block state influenced by the states of neighboring blocks. Γ represents the neighborhood coupling parameter, used to control the propagation strength of the neighborhood state difference term on the current block state, and its value is generally between 0.1 and 0.5. In a preferred embodiment, Γ is set to 0.3. N(i) is the neighborhood set of the i-th block, X j (t-1)-X i (t-1) represents the difference in state between neighboring block j and current block i at the previous time step, to avoid single-block noise being misjudged as region update, ω ij The neighborhood weight represents the weight of the current block. The larger the weight, the greater the impact of changes in the neighborhood state on the current block. For example, if the weight of the block above is 0.3, it means that changes in the block above have a moderate impact on the current block. It should be noted that the above values ​​are examples in this embodiment. In actual operation, those skilled in the art can freely choose according to the actual situation, and this embodiment does not impose any restrictions on this.

[0018] Due to the recursive update of the hidden state vector X i (t) represents the unnormalized state strength, therefore it needs to be transformed into a probability distribution using Softmax, as shown in the following formula: Among them, Π i (t) represents the hidden state probability vector of the i-th region block at time t, Π i S (t) represents the probability that the i-th block belongs to a statically stable state, Π i T (t) represents the probability that the i-th block belongs to a transient state, Π i C (t) represents the probability that the i-th block belongs to a continuous dynamic state, Π i P (t) represents the probability that the i-th block belongs to a periodic dynamic state. This probability is obtained by normalizing the state strength vector output from the region-level hidden state recursive model. Specifically, the previous hidden state vector X... i(t) is recursively derived from the state continuation term, the observation-driven term, and the neighborhood spatial coupling term. The state continuation term reflects historical state inertia, the observation-driven term reflects evidence of changes in the current frame, and the neighborhood coupling term reflects spatial consistency constraints. For the i-th region block at time t, the current observation input is formed based on its degree of change relative to the previous time step, its deviation relative to the background reference, its continuous activity level, its periodic repetition level, and its neighborhood linkage level. Combining the state results of the region block at the previous time step and the state results of neighboring blocks, static steady state intensity, transient change state intensity, continuous dynamic state intensity, and periodic dynamic state intensity are generated, respectively. The closer the match with a certain type of state characteristic, the higher the corresponding state intensity. Specifically, the degree of change is determined by comparing the difference between the current block description and the block description at the previous time step; the degree of deviation is determined by the strength of the deviation from the long-term background reference; the degree of continuous activity is determined by whether the current region block remains active for multiple consecutive frames; the degree of periodic repetition is determined by whether the changes in the current region block repeat according to a stable rhythm; and the degree of neighborhood linkage is determined by whether the changes in the current region block form a connected expansion or regional transition with adjacent region blocks. For example, when a region block changes significantly in the current frame but has low persistence and weak neighborhood linkage, the transient change state intensity component is relatively large. After normalization, Π i T (t) will dominate; when a region block remains active for multiple consecutive frames and its neighborhood shows an expansion trend, the sustained dynamic state intensity component is larger, then Π i C (t) dominates; when the changes in a certain region block are periodically repeated, the periodic dynamic state probability Π i P (t) increases. The reason why it needs to be controlled by the hidden state probability is that the state of the block cannot be accurately determined by simply relying on the difference of the current frame, and different types of blocks have significantly different requirements for background updates. For example, for static background areas, slow updates will cause background change tracking to lag; for local animation or periodically flickering areas, fast updates will mistakenly absorb the foreground into the background, resulting in afterimages or misjudgments.

[0019] S5: Adaptively control the background update coefficients based on the hidden state probabilities to achieve dynamic updates of the block-level background. Specifically, the background update coefficient α for each block at time t... i (t) is calculated from the four types of hidden state probabilities output in S4, and its expression can be written as: , Where clip means to use α i The range of (t) is limited to [α]. min α max Within the interval, α min Represents α i The lower bound of (t), α maxRepresents α i The upper bound of (t), where k1 represents the probability Π i C The weights of (t), k2 represents the probability Π i P The weights of (t), k3 represents the probability Π i S The weights of (t) are generally determined by satisfying k3 > k1, k3 > k2, with k1 and k2 being negative and k3 being positive. This ensures that a higher probability of static stability results in a faster update, while a higher probability of continuous dynamic or periodic dynamic changes results in a slower update. The update coefficient for transient changes is mainly determined by the baseline term α0, falling between the two. For example, k1 = -0.1, k2 = -0.15, k3 = 0.2. In practice, technicians can freely set the weight values ​​according to the actual situation; this embodiment does not impose any restrictions on this. If a region is in a stable state, it indicates that the current observation in that region is relatively consistent with the long-term background, and short-term disturbances and continuous dynamic components are weak. Therefore, the background update coefficient can be appropriately increased to allow the background reference to track the real slowly changing background more quickly. If a region is in a continuous or periodic dynamic state, it indicates that the region still contains stable motion components or repetitive changing components. If it is written into the background too quickly, the dynamic foreground may be mistakenly absorbed into the background. Therefore, the background update coefficient should be reduced. If a region is in a transient state, it indicates that the region is more likely to correspond to short-term transitions, single jumps, or transient disturbances. Therefore, its background update speed should be between that of a stable state and a continuous or periodic dynamic state, neither absorbing too quickly nor freezing excessively. In addition, since transient changes usually correspond to short-term transitions or single jumps, their background update requirements are between those of a stable state and a continuous or periodic dynamic state. They do not need to significantly accelerate background absorption or significantly suppress background updates. Therefore, in this embodiment, Π is not included. i T (t) is not written as an independent term in the background update coefficient formula, but rather the intermediate update level corresponding to the transient change is represented by the benchmark update term α0. Subsequently, this background update coefficient is used to perform an exponential sliding update on the slow background reference, as shown in the following formula: , Therefore, when the probability of static stability of the region block is high, α i When α is larger, the background absorbs the current observation more quickly; when the probability of continuous or periodic dynamics of the region block is high, α i (t) is relatively small, and historical information is preserved in the background to avoid foreground contamination. In actual implementation, the update coefficients are usually subjected to freeze protection, gradual recovery, and abnormal rollback to ensure the stability and reliability of background updates. Freeze protection refers to directly freezing several frames when the periodic dynamic probability or continuous dynamic probability is extremely high; gradual recovery refers to gradually restoring α after the dynamic ends. i(t); Abnormal rollback refers to the rollback to the old background reference if local mis-absorption causes residual abnormalities. It should be noted that this step is different from the previously proposed slow background reference concept. This step further clarifies the specific update rules of the slow background reference. This rule can achieve adaptive updates of different blocks in the same image, enabling fast tracking of the static background while protecting the dynamic foreground from mis-absorption.

[0020] S6: After obtaining the region-level hidden state, the block-level probability is further converted into whole-frame semantic features, providing reliable input for subsequent display mode selection to determine the display scene of the entire current screen and to determine the screen-level state. For example... Figure 4 The diagram shown is a screen-level state determination diagram provided in an embodiment of this application. This step extracts screen-level aggregation quantities based on the probability distribution of region blocks belonging to static, transient, continuous dynamic, and periodic dynamic states. Through a screen-level hidden state model, it obtains screen states with semantic distinguishability, such as fully static, locally transient, locally continuous dynamic, locally periodic dynamic, region expansion transition, or global switching. For example, when dozens of small blocks are flashing, the control strategy differs significantly from that of a large expanding region: the former may use local periodic refresh, while the latter may require transition enhancement or a larger area refresh. When extracting screen-level features, the system first summarizes the block-level state probabilities into screen-level aggregation quantities. In this embodiment, the screen-level aggregation quantities include the area ratio of various block states, the number of dynamically connected regions, the area ratio of the largest dynamically connected region, the rate of change of the center of gravity of the dynamic region, and the block state distribution entropy. The area ratio is obtained by weighted summation of each block by area, as shown in the following formula: , , , , Among them, a i Let ρ represent the area of ​​the i-th block. S ρ represents the probability of a block's statically stable state. T ρ represents the probability of a block's transient state change. C ρ represents the probability of the block's persistent dynamic state. P This represents the probability of a block's periodic dynamic state. Calculating the number of dynamically connected regions first requires generating a dynamic block mask, then summing the transient probability, persistent dynamic probability, and periodic dynamic probability of each block. If the sum exceeds a preset threshold τ... act If a value is found to be dynamic, the block is considered a dynamic block and denoted as 1; otherwise, it is denoted as 0. After obtaining the dynamic block masks for all blocks, connected component analysis is performed on the block mesh to calculate the number N of dynamically connected regions. c (t), used to distinguish whether dynamic blocks are scattered or expanded in patches. The area percentage of the largest dynamically connected region A. max(t) By calculating the area of ​​all dynamically connected regions in the frame, the ratio of the largest region's area to the total area of ​​the entire frame is used to distinguish whether the dynamic block is a scattered small area, a large dynamic area, or even a full-screen transition. Calculating the rate of change of the center of gravity of a dynamic region first requires calculating the coordinates of its center of gravity. These coordinates are obtained by weighting the geometric centers of gravity of all dynamic blocks by area or dynamic probability. Specifically, each dynamic block has an area and possesses transient, continuous, and periodic dynamic probabilities. These probabilities are summed to obtain the block's weight. Then, the geometric center of each dynamic block is multiplied by its weight, and the sum of all blocks is divided by the sum of all their weights to obtain the weighted center of gravity of the dynamic region in the current frame. Furthermore, the rate of change of the center of gravity v... c (t) represents the Euclidean distance between the centroid of the dynamic region in the current frame and the centroid of the dynamic region in the previous frame, divided by the inter-frame time interval. The block state distribution entropy H is calculated using the area proportions of the four types of block states in the entire frame: stable blocks, transient blocks, persistent dynamic blocks, and periodic dynamic blocks, which occupy different areas of the image. The entropy is obtained by multiplying each proportion by its logarithm, summing the results, and taking the negative value. Entropy H reflects the degree of state mixing in the entire frame; a high H indicates the coexistence of multiple states or a transitional scene, while a low H indicates a simple and stable image state. The above image-level aggregates are packaged to form an image-level feature vector, denoted as: , Next, the frame-level features are mapped to the hidden state vector of the current frame state through the frame-level hidden state model, denoted as z(t): , Among them, z fs (t) indicates that the current screen-level hidden state is a completely static screen, z lt (t) indicates that the current scene-level hidden state is a local transient change scene, z ld (t) indicates that the current screen-level hidden state is a local continuous dynamic image, z lp (t) indicates that the current screen-level hidden state is a local periodic dynamic frame, z re (t) indicates that the current screen-level hidden state is a region expansion transition screen, z gs z(t) represents the current screen-level hidden state as a global screen transition. Since the screen-level hidden state vector z(t) obtained through recursive updates is an unnormalized state, it needs to be transformed into a probability distribution using Softmax. The specific formula is as follows: , This probability output avoids the pattern jitter that can occur with hard-label classification. The criteria for determining different scene states and their correspondence with feature vectors are clear; for example, a completely static scene corresponds to ρ. S High, N c (t) small, Amax (t) low, v c (t) low and H low; local transient changes correspond to ρ T Rising but short duration, small dynamic area, N c (t) small; local continuous dynamic image corresponds to ρ C Maintain high resolution across multiple frames, A max (t) Concentration; Local periodic dynamic images correspond to ρ P High, dynamic regions are fixed and repeat periodically; the transitional scenes corresponding to the dynamic regions show a clear trend of interconnected expansion; A max (t) rises, v c (t) height and H height; global screen switching corresponds to A. max (t) is extremely large, the largest connected region is close to the entire screen and most blocks migrate synchronously.

[0021] S7: After obtaining the screen-level status, select the display mode and corresponding drive parameters based on the current screen status, and then control the COG module to execute the drive action. For example... Figure 5 The diagram shown illustrates the control action selection and feedback closed loop provided in this embodiment. After obtaining the screen state, the system performs cost optimization selection on multiple candidate control actions, including full-screen normal refresh, TE (Tearing Effect) synchronized local window refresh, background preservation plus dynamic ROI (Region of Interest) high-frequency refresh, layered periodic refresh, and full-screen transition enhancement refresh. Simultaneously, it outputs driving parameters such as local window coordinates, refresh cycle, TE synchronization start line, bus burst length, command batch size, and whether to enter local display mode. The same screen state corresponds to different optimal actions under different system or hardware constraints. For example, a continuous local dynamic screen is sometimes suitable for TE synchronized local window refresh, and sometimes for background preservation plus dynamic ROI high-frequency refresh. If bus usage is too high or the controller has limited support for local windows, it can revert to the full-screen normal refresh strategy.

[0022] First, a set of candidate control actions is defined, which includes five types of actions. The full-screen normal refresh mode is suitable for global switching or large-area expansion, where local updates are not beneficial and the controller or link is not suitable for complex local refresh scenarios; this is denoted as U1 in this embodiment. The TE synchronous local window refresh mode only refreshes locally changing windows under TE ticks, suitable for local transient changes, small-scale UI updates, and situations where the change area is clearly defined, with lower bus pressure; this is denoted as U2 in this embodiment. The background-preserving plus dynamic ROI high-frequency refresh mode is suitable for scenarios where static backgrounds and local continuous dynamics coexist, such as video windows, local animations, scrolling numbers, or indicator lights; this is denoted as U3 in this embodiment. The layered periodic refresh mode is mainly for local periodic dynamic scenes, refreshing the periodic dynamic layer rhythmically while refreshing the static layer at a low frequency; this is denoted as U4 in this embodiment. The full-screen transition enhanced refresh mode is suitable for area expansion transitions or global switching scenes; this is denoted as U5 in this embodiment. Each candidate action has a cost function, used to comprehensively evaluate the overall cost of performing the action in the current screen state. The cost function has a weighted combination structure, specifically expressed as: Among them, U k Indicates the motion control mode. T k This represents the refresh completion delay prediction, used to assess the time required for an action to complete the refresh. Specifically, the system first determines the window range or full-screen range to be refreshed based on candidate actions, then combines the display bit width, command overhead, and controller bus effective throughput to obtain the pure transmission time; subsequently, it adds the command issuance time, batch switching time, and TE synchronization waiting time to obtain the original completion delay prediction value for the refresh completion action, and then normalizes this prediction value to obtain the final value; B k This represents the bus occupancy cost, reflecting the consumption of transmission resources by an action. Specifically, the system calculates the total number of bytes required to be transmitted in the current cycle based on the refresh area, data bit width, command packet length, batch number, and burst transmission length corresponding to the candidate action, and then compares this result with the current system's available bus budget after normalization. k This represents the energy consumption cost, used to measure power consumption cost. Specifically, the system pre-establishes an action-level energy consumption estimation model, which is a weighted sum of refresh energy consumption per unit area, bus activity energy consumption per unit time, and mode-added energy consumption to obtain the original energy consumption estimate for the action, which is then normalized to obtain M. k This indicates the risk and cost of display artifacts or tearing, used to assess potential visual defects caused by the action. Display artifacts are a broader concept and can include visual defects such as ghosting, trailing, edge misalignment, localized transition anomalies, and visible residuals after refresh. Tearing is a specific risk, referring to screen breaks or misalignments that occur when the refresh writing process and the panel scanning process are not synchronized. Preferably, M... kThe artifact risk sub-item is weighted and synthesized from the artifact risk sub-item and the tear risk sub-item. The artifact risk sub-item is preferably determined based on one or more calculable quantities among the following: the non-coverage rate of the candidate refresh region over the current dynamic region, the outward expansion trend of the dynamic region boundary, the number of predicted residual blocks after the candidate action is executed, and the consistency deviation between the candidate action and the current region's hidden state result. The tear risk sub-item is preferably determined based on one or more calculable quantities among the following: the ratio between the data write duration of the candidate action and the available TE synchronization window length, the ratio between the expected data write amount of the candidate action and the available bandwidth of a single frame, the number of TE misses in the last 5-8 frames, and the consecutive occurrence length of TE misses. k The jitter cost of mode switching is used to penalize the control overhead and display instability caused by frequent switching actions in consecutive frames. It is determined based on whether the current candidate action is consistent with the action executed in the previous frame (i.e., the action type is the same), whether key control parameters such as refresh cycle, local window position and size, TE synchronization mode or starting line are consistent, the number of recent action switching, the magnitude of refresh cycle changes (i.e., window position offset, size change, and coverage change), and the magnitude of local window parameter changes (i.e., whether the synchronization mode has been switched and the amount of synchronization parameter change). Preferably, the system will use S... kThe original jitter intensity is divided into action type difference intensity, refresh cycle jump intensity, spatial window jump intensity, TE synchronization parameter jump intensity, and short-term switching frequency intensity. Action type difference intensity is determined by whether the candidate action type number matches the action type number of the previous frame. For example, if the current candidate action is the same as the action type executed in the previous frame, it is 0; if the action types are different, it is 1. The recent switching frequency is determined by the number of action type switches within the last 5 consecutive frames: 0 switches = 0, 1 switch = 1, 2 switches = 2, and at least 3 switches = 3. The refresh cycle jump intensity is determined by the ratio of the refresh cycle difference between the current candidate action and the action executed in the previous frame to the refresh cycle of the previous frame. For example, less than 5% = 0, 5% to 15% = 1, 15% to 30% = 2, and greater than 30% = 3. The spatial window jump intensity is calculated based on the offset distance of the window center coordinates and the percentage change in window area. For example: if the window center offset is no more than 1 block, the window area change is no more than 10%, and the overlap rate of the front and rear windows is no less than 90%, then it is set to 0; if the window center offset is 1 to 2 blocks, or the window area change is 10% to 25%, or the overlap rate of the front and rear windows is 70% to 90%, then it is set to 1; if the window center offset is 2 to 4 blocks, or the window area change is 25% to 50%, or the overlap rate of the front and rear windows is 50% to 70%, then it is set to 2; if the window center offset is more than 4 blocks, or the window area change is more than 50%, or the overlap rate of the front and rear windows is less than 50%, then it is set to 3. The TE synchronization parameter jump intensity is obtained based on the proportion of the change in the synchronization start line to the synchronization start line of the previous frame. For example: if the synchronization mode has not switched and the change in the start line does not exceed 5% of the corresponding value in the previous frame, it is set to 0; if it has not switched but the change is between 5% and 15%, it is set to 1; if it has not switched but the change is between 15% and 30%, it is set to 2; if the synchronization mode has switched, or the change in the start line exceeds 30%, it is set to 3. The short-time switching frequency intensity is directly obtained by statistically analyzing the number of action type switching within the most recent N consecutive frames. The system performs weighted summation, direct accumulation, or table lookup synthesis of the above intensities to obtain the original mode switching jitter intensity, and then performs graded mapping, normalization processing, or amplitude limiting processing to obtain S. k Preferably, after obtaining the above items, the system directly sums them to obtain the original S. k When the original accumulated value is 0, S k Set to 0; when the original accumulated value is between 1 and 4, S k Take the lower value; when the original accumulated value is between 5 and 8, S k Take the middle value; when the original accumulated value is greater than 8, S k Take the higher value and apply a limiting effect. If the current candidate action is the same as the action in the previous frame, and the number of transitions in the last 5 consecutive frames is low, then S... kTake the smaller value or 0; if the action type changes, or although the action type remains the same but the refresh cycle, local window range, or TE synchronization parameters change significantly, then S k The corresponding increase. The above-mentioned sub-item assignment rules are only examples, and technicians can adjust them according to the actual situation. This embodiment increases S. k The values ​​of β1 and β2 are not constrained. β1 represents the weight of refresh completion delay prediction, β2 represents the weight of bus occupancy cost, β3 represents the weight of energy consumption cost, β4 represents the weight of artifact or tearing risk cost, and β5 represents the weight of mode switching jitter cost. For example, β1=0.3, β2=0.2, β3=0.1, β4=0.25, β5=0.15. Technicians can freely adjust these values ​​according to actual conditions during operation. This embodiment does not constrain specific values; it only requires ensuring that the sum of all weights in the formula is 1. After the total cost is calculated, the system selects the action with the minimum total cost as the final execution strategy, expressed as: After selecting an action, the system will generate specific driving parameters, including local window coordinates, refresh cycle, TE synchronization start line, bus burst length, command batch size, and whether to enter local display mode. Specifically, the local window coordinates are written to the local display window register group, including at least the horizontal start coordinate, horizontal end coordinate, vertical start coordinate, and vertical end coordinate; the refresh cycle is written to the frame cycle register, frequency divider register, line timing register, or refresh tick control field; the TE synchronization start line is written to the TE enable control bit, TE compare line register, or synchronization trigger line field; the bus burst length is written to the host interface burst length register, DMA (Direct Memory Access) burst transfer configuration field, or FIFO (First In First Off) threshold configuration field; the command batch size is written to the command queue depth register, transaction packet quantity field, or DMA descriptor batch parameters; and whether to enter local display mode is written to the local display enable bit, partial refresh mode command field, self-refresh or idle mode control bit, or ROI mode enable field. The names of the aforementioned registers or command fields may vary depending on the specific controller model and interface protocol, but their functions all correspond to hardware configurations such as window limitation, timing setting, synchronization triggering, bus scheduling, and mode switching. For example, if the background hold plus dynamic ROI high-frequency refresh mode is selected, the system will map the previously obtained dynamic area to the ROI and set a higher refresh frequency for the ROI, while setting a hold or low-frequency refresh strategy for the background, thereby achieving fine control. Furthermore, in this mode, the system preferably first converts the dynamic area into the start and end coordinates of the local window and writes them into the local display window register group, and then writes the refresh tick, frame period, or frequency division parameters corresponding to the ROI into the timing control register; then, the local display mode enable bit is enabled, and the TE synchronization start line is configured so that the write start point of the local refresh is aligned with the panel scan tick. After that, according to the current ROI area and refresh frequency requirements, the bus burst length and command batch size are set to control the data delivery rhythm of the host interface or DMA channel, and finally, the local refresh is started by the refresh trigger command, the video memory write command, or the local update start bit. For the background area, you can control whether it maintains the current display content or participates in the refresh at a lower frequency by using the background hold bit, low refresh frequency division configuration item, or background do not update flag.

[0023] S8: After the driving action is executed, the system collects execution feedback data, including actual write latency, TE misses, number of residual blocks after partial refresh, number of mode switches, and bus utilization, etc., to correct the weights of the region hidden state model parameter matrix and the action cost function. Based on the region hidden state recursive model, the system constructs an error signal according to the execution feedback and uses recursive least squares or exponential moving average estimation to correct matrix A, matrix B, and parameter Γ. For example, regarding matrix A, if the system frequently misjudges short-term transients as persistent dynamics during actual execution, it indicates that the inheritance of historical states is too strong, and the corresponding component in A should be appropriately reduced. If the system fails to maintain the true persistent dynamics, resulting in frequent state jitter, the corresponding component in A should be appropriately increased. Regarding matrix B, if the system is strongly pushed into a persistent dynamic state after only one short-term jump, local noise disturbance, or single-frame digital flip in the current frame, it indicates that the current observation drives the persistent dynamic state too strongly, and the corresponding component in B should be appropriately reduced. If the system already shows obvious background deviation, high persistence characteristics, or strong activity evidence in the current frame, but the model still cannot transition to a persistent dynamic state in time, it indicates that the current observation drives the persistent dynamic state insufficiently, and the corresponding component in B should be appropriately increased. Regarding parameter Γ, if the dynamic region has obvious spatial expansion or connectivity propagation, and the model fails to reflect this trend in time, Γ should be increased. If the model excessively diffuses local isolated changes to the surrounding region, Γ should be decreased. To prevent excessively large update increments, the system manually sets maximum increment limits for the update values ​​of matrices A, B, and Γ. After consecutive misjudgments exceeding a preset number of frames, a conservative mode is triggered to limit the update increment. After each drive action, the system collects feedback quantities such as actual write latency, TE misses, residual block count after partial refresh, mode switching count, and bus utilization. Based on these feedback quantities, an error signal reflecting the deviation between the current region's hidden state determination and the actual execution effect is constructed. When a region is determined to be statically stable and therefore adopts a more aggressive background absorption or lighter refresh strategy, but the residual block count continues to increase after execution, it indicates that the hidden state estimation of that region is biased towards static stability. When a region is determined to be locally dynamic and a local refresh strategy is adopted, and the TE misses and ghosting risk significantly increase, it indicates that the state propagation or dynamic persistence modeling of that region is inaccurate. Based on these error signals, parameters can be updated online. In this embodiment, the system uses an exponential sliding estimation method for updating, specifically as follows: , , , Among them, A t+1 B represents the state continuation matrix after the update at time t+1. t+1 Γ represents the observation driving matrix updated at time t+1. t+1Let A' represent the neighborhood coupling parameters updated at time t+1. t B' represents the target state continuation matrix parameter calculated based on the current execution feedback error. t Γ' represents the target observation driving matrix parameters calculated based on the current execution feedback error. t η represents the target neighborhood coupling parameter calculated based on the current feedback error. A η represents the smooth update coefficient of the state continuation matrix. B η represents the smooth update coefficient of the observation driving matrix. Γ This represents the smooth update coefficient of the neighborhood coupling parameters. It should be noted that η... A η B η ΓThe error magnitude and reliability of the current execution feedback are adaptively determined. The system calculates the time continuation error, observation-driven error, and spatial coupling error respectively. Among them, the time continuation error is calculated by the system for the participating blocks that were judged as continuous dynamic or periodic dynamic in the previous moment. It is determined whether the short-term changes, background deviations, and persistence evidence in the subsequent three consecutive frames have synchronously decreased, and whether the corresponding dynamic state probability still maintains the dynamic state judgment result. If the aforementioned observation evidence has not met the dynamic judgment condition in at least two of the three consecutive frames, but the model still maintains the block as dynamic, it is judged as having excessively strong historical state inheritance. If the observation evidence has met the corresponding dynamic maintenance condition in at least two of the three consecutive frames, but the model has exited the dynamic state, it is judged as having excessively weak historical state inheritance. Finally, the ratio of the number of blocks that meet the judgment condition of excessively strong historical state inheritance to the number of participating judgment blocks is used to reflect the historical state inheritance strength represented by matrix A. The observation-driven error is determined by the system for the region block participating in the judgment of the current frame. If the current block only shows a large short-term change, such as a single frame jump, local noise disturbance, or short-term digital flip, but the background deviation, persistence, periodicity, and spatial diffusion are not considered to be in a continuous dynamic state, yet the model enters a continuous dynamic state or a periodic dynamic state, then the observation-driven error is determined to be too strong. If the evidence of short-term change, background deviation, or persistence of the current block has reached the dynamic judgment condition, but the model still remains in a static or transient state, then the observation-driven error is determined to be too weak. The error is obtained by the proportion of the number of blocks that meet the judgment condition of too strong observation-driven error to the number of blocks participating in the judgment, and is used to reflect the current observation-driven intensity represented by matrix B. Spatial coupling error is determined by the system based on the proportion of activity in the current block's neighborhood, the spatial distribution of residual blocks after local refresh, and whether the dynamically connected regions have formed a connected expansion result. If the current block only shows isolated point-like changes, the neighborhood dynamic distribution has not formed a connected expansion, and the dynamic region has not entered the outward expansion state relative to the previous frame, but the model pushes multiple surrounding blocks into a dynamic state simultaneously, then it is judged as spatial coupling too strong. If the residual block or dynamic block has already shown a connected expansion trend, but the model still does not reflect the corresponding propagation, then it is judged as spatial coupling too weak. Finally, the error is obtained by the proportion of the number of blocks that meet the criteria of excessively strong spatial coupling to the number of blocks participating in the judgment, which is used to reflect whether the neighborhood propagation effect represented by the parameter Γ is too strong or too weak. To facilitate the setting of the update step size, the system classifies the above three types of errors. Preferably, the classification is based on the proportion of the number of misjudged blocks to the number of blocks participating in the judgment: less than 10% is a mild error, 10% to 25% is a moderate error, and more than 25% is a severe error. If the error in the same direction appears in at least 2 out of 3 consecutive frames, it is judged as a recurrence. It should be noted that the above-mentioned grading ratio is only an example. During the operation, technicians can set it freely according to the actual situation. This embodiment does not impose any restrictions on this.

[0024] Furthermore, the system does not directly equate the error level with the update step size. Instead, it first looks up the corresponding base step size based on the error level, and then performs scaling corrections based on the feedback reliability. Wherein, η A η B η Γ All parameters are pre-set to the same base value, typically the default update coefficient set during system initialization. This value can be manually preset, for example, 0.5, to ensure the system has a certain degree of self-adaptability in the initial stage without excessively rapid fluctuations. Preferably, these values ​​can be η. A η B η Γ Three basic step size ranges—low, medium, and high—are preset, and a mapping table is established between error levels and basic step sizes: minor errors are mapped to low-level basic step sizes, moderate errors to medium-level basic step sizes, and severe errors to high-level basic step sizes. During subsequent operation, the system does not directly replace the current η with the lookup table result. A η B η Γ Instead, it starts from the corresponding base value and adjusts according to the target step size mapped by the current error level. When the error reaches a high level but does not recur, it only enters the observation or low-level update state. Then, it judges whether the current feedback is reliable. The system performs anomaly detection on bus utilization, mode switching times, number of residual blocks after partial refresh, and TE misses. If the same type of error appears in the same direction in at least 2 out of 3 consecutive frames, it is judged as a recurrence. Abnormal bus utilization means that the current bus utilization is higher than the preset proportion of the sliding average of the last 8 frames for 3 consecutive frames, preferably 15% to 20% higher. Excessive mode switching means that the number of action type switching times in the last 5 frames reaches the preset upper limit, preferably no less than 3 times. Abnormal increase in residual blocks means that the number of residual blocks after partial refresh is higher than the preset proportion of the sliding average of the last 8 frames for 3 consecutive frames, preferably 20% higher. Abnormal TE means that the number of TE misses in the last 3 frames increases continuously. If any of the above abnormalities exist, it means that the current feedback is susceptible to bus congestion, synchronization mismatch, or execution link fluctuations. In this case, η A η B η Γ Place it in a low-level update interval to reduce the parameter update step size in this round. If none of the above anomalies occur, and the same type of error repeats within 3 consecutive frames, then set η according to the corresponding error level. A η B η Γ Preferably, a slight error corresponds to a low-level update, a moderate error corresponds to a medium-level update, and a severe error corresponds to a high-level update. Furthermore, in practical applications, η A η B η ΓThe value of η is also subject to upper and lower limits to prevent abnormal feedback from causing parameter jumps. For example, the value of η can be limited to between 0.05 and 0.30, with 0.05 to 0.10 for low settings, 0.10 to 0.20 for medium settings, and 0.20 to 0.30 for high settings, and the final result is also subject to limiting. Additionally, in cases of a surge in TE misses or severe ghosting across multiple consecutive frames, the system can force a switch to full-screen normal refresh mode as a fallback mechanism to ensure display stability.

[0025] Furthermore, to achieve the aforementioned corrections, the system collects feedback quantities such as actual write latency, TE misses, residual block count after partial refresh, mode switching count, and bus utilization after each drive action. Based on these feedback quantities, it constructs an error signal reflecting the deviation between the current hidden state determination and the actual execution effect, which is then mapped to target correction quantities for the state continuation matrix, observation drive matrix, and neighborhood coupling parameters. Specifically, the system preferably pre-establishes a parameter correction rule table or correction template library between error direction, correction object, and correction magnitude. That is, it first classifies error types into time continuation errors, observation drive errors, and spatial coupling errors, and then classifies error directions into two categories: excessively strong and excessively weak. For each category combination, a corresponding correction object and correction method are preset. During runtime, based on the currently identified error type, error direction, and error level, the corresponding correction quantity is retrieved from the rule table or template library, and then the correction quantity is added to the matrix parameters generated in the previous time step to generate the target parameters. Thus, A' t B' t ,Γ' t All parameters are determined jointly by the current parameters and the error signal of this round, rather than being independently specified manually. Specifically, if a region is determined to be statically stable or transient and a lightweight refresh strategy is adopted, and the local refresh residual continues to increase or the mode jitter intensifies, it indicates that the inheritance of historical states is either too strong or too weak. Based on this, the target state continuation matrix parameter A' is generated. t If the actual write latency, TE misses, or bus utilization continue to increase, or if local changes are significant but there are still many residuals after execution, it indicates that the observed features are deviating from the driving force of the hidden state. Based on this, the target observation driving matrix parameter B' is generated. t If the residual blocks exhibit spatial connectivity expansion but the model fails to reflect this propagation trend, or if local isolated changes are excessively diffused to the surrounding region, it indicates that the spatial coupling modeling is inaccurate. Based on this, the target neighborhood coupling parameter Γ' is generated. tThe correction magnitudes of the aforementioned target parameters correspond to the error levels. Preferably, a slight error corresponds to a 2% parameter correction, a moderate error to a 5% correction, and a severe error to an 8% correction. These 2%, 5%, and 8% correction magnitudes are applied to the current parameter as relative correction magnitudes, meaning the corresponding correction amounts do not exceed 2%, 5%, and 8% of the current corresponding parameter or matrix item, respectively. The specific direction of increase or decrease is determined by the error direction. Subsequently, the system uses the aforementioned exponential moving average estimation formula to adjust A'. t B' t ,Γ' t Perform smooth updates separately.

[0026] The system also adjusts the weights of various costs in the action cost function. The system adjusts the weights β1, β2, β3, β4, and β5 based on feedback and the severity of the actual problem. The weight of the corresponding problem is increased when the problem is more severe, and decreased when it is not prominent over a long period. For example, if the system finds that the actual write latency is consistently high after executing the current candidate action, it indicates that the current decision does not pay enough attention to latency factors. In this case, β1 should be increased appropriately, and subsequent action selection will favor low-latency solutions. If the bus utilization is consistently high, β2 should be increased appropriately, and subsequent action selection will favor reducing bus usage. If the number of residual blocks increases, the number of missed TEs increases, or the risk of display tearing or ghosting increases after a partial refresh, β4 should be increased appropriately to increase the system's penalty for display quality risks. If mode switching is too frequent and the driving mode jitters between multiple frames, β5 should be increased appropriately to suppress frequent switching. The energy consumption cost weight corresponding to β3 can be corrected based on the refresh area, refresh frequency, bus active time, or panel power consumption estimation results. This embodiment uses an exponential sliding estimation method to update each weight coefficient, and the specific updates are as follows: k=1, 2, 3, 4, 5 Where, β k (t) represents the weight coefficient of the k-th cost term at time t, β k (t+1) represents the weight coefficient of the k-th cost term after the update at the next time step, β' k (t) represents the target weight calculated based on the current execution feedback result, η β β represents the smoothing update coefficient. k min This represents the lower bound of the weight, β. k max This indicates the upper limit of the weight, and `clip` represents the clipping function. During the weight update process, the system sets β. k The minimum and maximum values ​​of (t) are defined to prevent short-term abnormal feedback from causing the weights to be too large or too small. Furthermore, when TE is missed or ghosting occurs in several consecutive frames, a weight backoff or conservative adjustment strategy can be triggered to ensure the stability of action selection. In addition, β'k (t) is not fixed in advance, but is determined based on the relationship between the corresponding feedback quantity and the anomaly level at the current moment. This relationship can be established in advance in two ways: first, through an experience mapping table obtained from prototype debugging, aging tests, and historical operating data statistics; and second, through manually formulated segmented mapping rules based on engineering experience. In actual implementation, the mapping table lookup method is preferred, while manually formulated segmented mapping rules are used when historical samples are insufficient. Specifically, the system first establishes normal reference ranges and anomaly level classification rules for each type of feedback quantity. The normal reference range can use a fixed engineering threshold or a dynamic baseline formed by sliding statistics from the most recent 5-8 frames. Then, the system compares the current feedback value with the corresponding normal reference range, calculates the anomaly level, and outputs the target weight β' according to the preset mapping rules. k (t), where the anomaly level is preferably divided into four levels: normal, mildly abnormal, moderately abnormal, and severely abnormal, and mapped to four target weight levels respectively. The system pre-writes the anomaly interval and target weight corresponding to each type of feedback quantity into the mapping table. During runtime, it first determines the anomaly interval to which the current feedback value belongs based on the current feedback value, and then looks up β' in the table. k (t). For example, for bus utilization, a mapping table can be set up so that the normal interval of β'1(t) corresponds to normal, the warning interval corresponds to mild anomaly, the congestion interval corresponds to moderate anomaly, and the high congestion interval corresponds to severe anomaly; a joint judgment table can be set up for the number of missed TEs and the number of residual blocks, and the target level of β'4(t) can be directly output based on the combination of the two. To construct the target weight β' k (t), the system maps the feedback quantity after execution accordingly, mapping the actual write latency to the correction basis of β'1(t), the bus utilization to the correction basis of β'2(t), the power consumption estimate or refresh resource consumption to the correction basis of β'3(t), the TE missed number and the number of residual blocks after partial refresh to the correction basis of β'4(t), and the mode switching number to the correction basis of β'5(t). After completing the weight update, S7 performs corresponding adjustments for each candidate action U. k The overall cost will be recalculated using the updated β1, β2, β3, β4, and β5, thus affecting the selection of subsequent actions. If the main problems of the system recently are missed TEs and increased local residuals, the updated cost function will improve the efficiency of M. k The penalty makes the system more inclined to select candidate actions with lower explicit risk; if the main problem is excessive bus pressure, the updated cost function will increase the penalty for B. k The penalty will cause the system to prioritize refresh methods with lower bandwidth consumption; if the main problem is mode jitter, it will increase the penalty for S. k Punishment is used to enhance the continuity and stability of control strategies.

[0027] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions within the technical scope disclosed in this application should be included 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 multi-display mode driving control method for COG liquid crystal modules, characterized in that, Includes the following steps: S1: Obtain the current display screen and divide the current display screen into multiple display sub-region blocks, and construct a block description vector for each of the display sub-region blocks; S2: For each of the aforementioned display sub-region blocks, establish a short-time reference block description and a background reference block description, and determine the first change amount and the second change amount based on the block description vector at the current time. S3: Extract time-series change features based on the first change and the second change; S4: Based on the first change amount, the second change amount, the temporal change characteristics, and the state association information of the neighboring display sub-region blocks, construct a region-level hidden state recursive model to obtain the region hidden state probability corresponding to each of the display sub-region blocks. S5: Determine the background update coefficients based on the hidden state probabilities of each region, and update the corresponding background reference block description; S6: Extract the image-level aggregation based on the hidden state probability of the region, and determine the image state corresponding to the currently displayed image; S7: Based on the screen state, select the target driving action from multiple candidate driving actions and generate driving parameters to control the display module to execute the display driving; S8: After executing the display driver, collect the execution feedback quantity, and update the parameters in the region-level hidden state recursive model and the cost function weights corresponding to the candidate driving actions based on the execution feedback quantity.

2. The multi-display mode driving control method for COG liquid crystal modules as described in claim 1, characterized in that: The building block description vector includes: Extract at least two of the following features from each of the display sub-region blocks: brightness statistical features, dispersion features, structural energy features, and content signature features; and construct the block description vector of the corresponding display sub-region block based on the extracted features. The block description vector is constructed using a streaming method that is extracted synchronously with the generation process of the current display screen.

3. The multi-display mode driving control method for COG liquid crystal modules as described in claim 1, characterized in that: The establishment of the short-term reference block description and the background reference block description includes: The block description vector of the corresponding display sub-region block in the previous moment or the preset short-time window is used as the short-time reference block description; The reference block description of the corresponding display sub-region block on a longer time scale is used as the background reference block description; The first change amount is obtained by calculating the description distance between the block description vector at the current time and the short-time reference block description using the description distance function, and the second change amount is obtained by calculating the description distance between the block description vector at the current time and the background reference block description.

4. The multi-display mode driving control method for COG liquid crystal modules as described in claim 3, characterized in that: The extraction of temporal variation features includes: The activity markers of each of the display sub-region blocks are determined based on the first change and the second change. Within a preset time window, persistent features are extracted based on the activity markers or change sequences, periodic features are extracted based on the change sequences, and spatial diffusion features are extracted based on the activity distribution of the currently displayed sub-region block and neighboring displayed sub-region blocks. The temporal variation characteristics include at least persistence characteristics, periodic characteristics, and spatial diffusion characteristics.

5. The multi-display mode driving control method for COG liquid crystal modules as described in claim 1, characterized in that: The construction of the region-level hidden state recursive model to obtain the region hidden state probability corresponding to each of the displayed sub-region blocks includes: For each of the aforementioned display sub-region blocks, a region-level hidden state vector is established. Based on the region-level hidden state vector of the previous time step, the first change amount, the second change amount, the temporal change characteristics, and the state association information of the neighboring display sub-region blocks, the region-level hidden state vector of the current time step is recursively updated. The recursively updated region-level hidden state vector is normalized to obtain the region hidden state probabilities of the corresponding display sub-region block belonging to the static steady state, transient change state, continuous dynamic state, and periodic dynamic state.

6. The multi-display mode driving control method for a COG liquid crystal module as described in claim 1, characterized in that: The determination of background update coefficients based on the hidden state probabilities of each region includes: The background update coefficient is calculated based on the hidden state probability of each of the display sub-region blocks, and the background update coefficient is subject to boundary constraints. Based on the background update coefficient, the block description vector at the current time, and the background reference block description at the current time, the background reference block description of the corresponding display sub-region block is recursively updated.

7. The multi-display mode driving control method for COG liquid crystal modules as described in claim 1, characterized in that: The S6 includes: Extract at least two of the following parameters from the area ratio of various hidden states in the current display screen, the number of dynamically connected regions, the area ratio of the largest dynamically connected region, the change in the centroid of the dynamic region, and the state distribution entropy: use them as the screen-level aggregation quantity; and construct a screen-level state vector based on the screen-level aggregation quantity. The screen state corresponding to the currently displayed screen is determined based on the screen-level state vector; The screen states include at least two of the following: fully static screen state, local transient change screen state, local continuous dynamic screen state, local periodic dynamic screen state, regional expansion transition screen state, and global switching screen state.

8. The multi-display mode driving control method for a COG liquid crystal module as described in claim 1, characterized in that: The S7 includes: A candidate driving action set is pre-established, and the action cost is calculated for each candidate driving action in the candidate driving action set. The target driving action is selected based on the action cost corresponding to each of the candidate driving actions, and the driving parameters are generated based on the target driving action. The candidate driving actions include at least two of the following: full-screen normal refresh action, synchronous local window refresh action, background preservation and dynamic area high-frequency refresh action, layered periodic refresh action, and full-screen transition enhancement refresh action. The driving parameters include at least two of the following: local window coordinates, refresh cycle, synchronization start position, bus burst length, command batch size, and display mode parameters. The action cost is composed of multiple weighted cost values, which include at least two of the following: refresh completion delay cost, bus occupancy cost, energy consumption cost, display risk cost, and mode switching cost.

9. The multi-display mode driving control method for a COG liquid crystal module as described in claim 1, characterized in that: The S8 includes: Collect the execution feedback quantity corresponding to the execution result of the target-driven action, and construct error information based on the execution feedback quantity; The state continuation parameters, observation driving parameters, and neighborhood coupling parameters in the region-level hidden state recursive model are updated based on the error information, and the cost function weights corresponding to the candidate driving actions are updated based on the error information. The execution feedback quantity includes at least two of the following: actual write latency, number of missed synchronization signals, number of residual blocks after partial refresh, number of mode switching, and bus utilization. The update adopts one or more of the following methods: recursive least squares update, exponential sliding estimation update, or a combination of both.

10. A multi-display mode driving control method for a COG liquid crystal module as described in claim 1 or 9, characterized in that, S8 further includes: Update boundaries are set for the parameters and cost function weights in the region-level hidden state recursive model. When the execution feedback quantity meets the preset abnormal conditions, the update magnitude of the corresponding parameter or weight is limited. When the execution feedback quantity continuously meets the preset trigger conditions, the display driver corresponding to the preset conservative driving action is switched.

Images