Electronic ink screen local refresh method and system combining space-time prediction and waveform optimization

Abstract

The application discloses a local refresh method and system of an electronic ink screen based on space-time prediction and waveform optimization, and relates to the technical field of local refresh of an electronic ink screen, which comprises the following steps: collecting user behavior characteristics and screen state data and inputting the same into a user behavior prediction module, outputting a dirty area, a potential modification area and a prediction confidence thereof, forming a pre-rendering candidate area based on the same, and converting the pre-rendering candidate area into a standardized data structure after being combined with the prediction confidence and being issued; an intelligent waveform optimization module outputs target waveform parameters based on extracted key characteristics of the dirty area and the prediction confidence; a residual image elimination module generates a reverse compensation waveform based on the target waveform parameters and information of the dirty area, and the reverse compensation waveform is used for eliminating residual images; finally, a display controller is instructed to refresh a screen and update display content; the refresh method fully utilizes the characteristics of the electronic ink screen, i.e., no backlight and no frequency flashing, thereby providing a healthy programming learning environment for children, and reducing the risk of visual fatigue and vision damage caused by long-time use.

Description

Technical Field

[0001] This invention relates to the field of partial refresh technology for electronic ink screens, and more particularly to a method and system for partial refresh of electronic ink screens based on spatiotemporal prediction and waveform optimization. Background Technology

[0002] E-paper displays are reflective display devices based on electrophoretic display technology. Their core principle is to control the movement of charged pigment particles using an external electric field, thereby displaying images. Compared to traditional LCD screens, e-paper displays have the following significant advantages:

[0003] Low power consumption: It only consumes power when refreshing, and consumes almost no power when displaying statically;

[0004] Eye-protection features: No backlight, no flicker, reflects ambient light to form an image, providing a reading experience close to that of paper;

[0005] Outdoor visibility: It remains clearly visible even in strong light, without relying on backlight brightness adjustment.

[0006] These characteristics have led to the widespread use of e-ink screens in devices such as e-readers, electronic tags, and smartwatches. However, the dynamic refresh performance of e-ink screens remains insufficient in highly interactive scenarios, especially in programming education, where they lack adequate response speed and real-time feedback capabilities.

[0007] Programming education for teenagers and children refers to programming enlightenment and skills training activities for teenagers and children aged 6-18. It aims to cultivate logical thinking, creativity, and problem-solving abilities through graphical programming tools (such as Scratch) or text-based programming languages ​​(such as Python). This field presents the following core requirements for display devices:

[0008] High-frequency input and real-time feedback: Students need to frequently input code, debug programs, and view the results in real time during the programming process, requiring the display device to have low latency and high refresh rate;

[0009] Eye protection needs during prolonged use: Children and teenagers are in a critical stage of vision development, and prolonged use of traditional LCD screens can easily lead to problems such as eye fatigue and dry eye.

[0010] Clarity and readability: Code display requires high contrast and low ghosting to ensure that students can clearly read and edit the code.

[0011] While e-ink screens offer significant advantages in static display, low power consumption, and eye protection, they suffer from the following issues during dynamic refresh, particularly in programming education scenarios for children and teenagers:

[0012] (1) High response delay: The refresh rate of traditional e-ink screens is usually hundreds of milliseconds, which causes display delay when inputting code; when students input code, the screen updates are delayed, affecting the smoothness of operation and learning experience.

[0013] (2) Obvious ghosting: Due to the physical characteristics of e-ink screens, ghosting is easily generated when dynamically refreshing, especially in areas where code is frequently updated; ghosting reduces the readability of the code and increases the cognitive burden on students.

[0014] (3) Global refresh problem: Existing technologies mainly eliminate ghosting through global refresh, but global refresh causes screen flickering, affecting visual comfort. Global refresh also increases power consumption and reduces the device's battery life.

[0015] (4) Lack of dynamic adaptation capability: Existing partial refresh technology usually adopts a fixed refresh mode, which cannot dynamically adjust the refresh strategy according to user input behavior. In programming education scenarios, students' input behavior is highly uncertain and diverse, and the fixed refresh mode is difficult to meet the needs.

[0016] Existing e-ink screen optimization technologies mainly focus on simple partial refresh, global refresh optimization, and waveform optimization. However, these technologies fail to fundamentally solve the problems of response latency and ghosting, especially in educational scenarios where frequent code input and updates are required. Therefore, there is an urgent need for a new approach that combines user behavior prediction, intelligent waveform optimization, and efficient partial refresh technology to provide an e-ink screen display solution more suitable for programming education. Summary of the Invention

[0017] Based on the technical problems existing in the background technology, this invention proposes a spatiotemporal prediction and waveform optimization method and system for local refresh of e-ink screens, which fully utilizes the characteristics of e-ink screens being backlight-free and flicker-free, providing a healthy programming learning environment for children and teenagers.

[0018] The spatiotemporal prediction and waveform optimization method for local refresh of e-ink screens proposed in this invention includes:

[0019] Collect user behavior features and screen status data and input them into the user behavior prediction module to predict dirty regions that will undergo pixel changes in the next refresh cycle and potential modification regions that may be edited within a preset time window; after preprocessing the potential modification regions, obtain the pre-rendered candidate regions containing the dirty regions, perform pre-rendering, and encapsulate the results and the corresponding prediction confidence into a standardized data structure for distribution.

[0020] The intelligent waveform optimization module extracts key features of dirty regions and prediction confidence from a standardized data structure. When the prediction confidence is greater than or equal to a set confidence threshold, it refers to the relevant features of potential modification regions and dynamically selects waveform types from a preset waveform library according to a preset decision logic. Based on the dynamic adjustment rules, it optimizes the control parameters of the selected waveform type and outputs the target waveform parameters.

[0021] The ghosting elimination module generates a reverse compensation waveform based on the target waveform parameters and the information of the dirty region, combined with the extended boundary sub-region incorporated into the pre-rendered candidate region, to eliminate ghosting; the selected waveform type and the reverse compensation waveform are merged in chronological order to generate a driving instruction, which instructs the display controller to refresh the screen and update the display content.

[0022] Furthermore, the process of collecting user behavior features and screen state data and inputting them into the user behavior prediction module predicts dirty areas that will undergo pixel changes in the next refresh cycle and potential modification areas that may be edited within a preset time window, specifically:

[0023] Based on the Long Short-Term Memory network, short-term time-series features of user behavior are captured to predict dirty regions that will undergo pixel changes in the next refresh cycle.

[0024] The Transformer model is used to predict potential modification areas that may be edited within a preset time window.

[0025] Furthermore, in the user behavior prediction module, the pre-processing of the potential modification region to obtain the pre-rendered candidate region containing the dirty region specifically involves:

[0026] When the prediction confidence is greater than or equal to the set confidence threshold, potential modification sub-regions are selected from the potential modification regions according to priority. The priority is based on factors such as confidence, ghosting risk index and region area. The regions are merged in order according to the ranking results until the area budget is reached. The sub-regions are then combined with the dirty regions to form pre-rendered candidate regions.

[0027] When the prediction confidence is less than the set confidence threshold, the sub-regions in the potential modification region are not merged into the dirty region, and only the boundary expansion of the dirty region is performed as a pre-rendering candidate region.

[0028] Furthermore, in the intelligent waveform optimization module, when the prediction confidence is greater than or equal to a set confidence threshold, the relevant features of the potential modification region are referenced, and the waveform type is dynamically selected from the preset waveform library according to a preset decision logic. The intelligent waveform optimization module selects the waveform type based on the following criteria:

[0029] When the prediction confidence is greater than or equal to the set confidence threshold, the waveform type is selected based on the information of the dirty area and the future risk of the potential modification area. If the ghosting risk index of the potential modification area at the target position is greater than or equal to the preset ghosting risk index threshold, a stable refresh curve is used in advance for the adjacent dirty areas or the phase reversal time is appropriately extended.

[0030] When the prediction confidence is less than the set confidence threshold, no refresh is triggered for the potential modification area; only parameter preset and caching are performed.

[0031] Furthermore, in the intelligent waveform optimization module, the waveform type is dynamically selected from the preset waveform library according to a preset decision logic, specifically as follows:

[0032] The key features of the dirty region include change frequency, character density, and content type;

[0033] When the frequency of change is greater than or equal to the preset frequency of change, select the fast waveform;

[0034] When the character density is greater than the preset character density, select the fast waveform;

[0035] When the ghosting risk index is greater than the preset ghosting risk index, select the high-quality waveform;

[0036] Conversely, select the mixed waveform.

[0037] Furthermore, the user behavior characteristics include key press type, cursor movement trajectory, and input interval time;

[0038] The screen status data includes the pixel distribution and code structure information of the current code editing area.

[0039] Furthermore, in the intelligent waveform optimization module, the waveform library includes:

[0040] Fast waveform, suitable for high-frequency input, allows for image retention, basic parameter range: voltage gradient 2.5–4.5V / ms, refresh time 6–12ms;

[0041] High-quality waveforms, suitable for static display, ignoring ghosting; basic parameter range: voltage gradient 0.6–1.5V / ms, refresh time 40–80ms.

[0042] Hybrid waveform, suitable for intermediate frequency input, balances ghosting and delay, with basic parameters ranging from 1.5 to 3.0 V / ms for voltage gradient and 15 to 30 ms for refresh time.

[0043] Furthermore, the preferred range of the basic parameters is:

[0044] Fast waveform: voltage gradient 3.0–4.0V / ms, refresh time 8–10ms;

[0045] High-quality waveform: voltage gradient 0.8–1.2V / ms, refresh time 50–60ms;

[0046] Mixed waveform: voltage gradient 2.0–3.0V / ms, refresh time 18–24ms.

[0047] Furthermore, in the intelligent waveform optimization module, the control parameters of the selected waveform type are optimized according to dynamic adjustment rules, and the target waveform parameters are output. The dynamic adjustment rules are as follows:

[0048] Fast waveform, if the area of ​​the dirty region is greater than 100 pixels 2 Increase the voltage gradient; if the residual image risk in the potential modification area is greater than or equal to the preset residual image risk index threshold, then extend the phase exit time in advance.

[0049] For high-quality waveforms, if the temperature is greater than 30°C, reduce the voltage gradient; if the ghosting risk index is greater than 0.15, extend the refresh time.

[0050] The voltage gradient is linearly adjusted based on the afterimage risk index for the mixed waveform; if the afterimage risk of the potential modification area is greater than or equal to the preset afterimage risk index threshold, the lower limit of the voltage gradient is increased.

[0051] Furthermore, the image retention module generates a reverse compensation waveform based on the target waveform parameters and the information of the dirty region to eliminate image retention, specifically:

[0052] Calculate the pixel difference between the dirty regions in the current frame and the previous frame. If the difference between the two is greater than the pixel difference threshold, it is marked as a potential ghosting region.

[0053] Image enhancement of potential ghosting regions is performed by contrast stretching and edge detection, generating a list of ghosting region coordinates;

[0054] The voltage polarity of each afterimage region is reversed, and the duration of the reversal is dynamically adjusted in combination with the afterimage intensity to return the generated reverse compensation waveform.

[0055] A spatiotemporal prediction and waveform optimization e-ink screen partial refresh system, including a user behavior prediction module, an intelligent waveform optimization module, and an afterimage elimination module;

[0056] The user behavior prediction module captures short-term time-series features of input user behavior based on a long short-term memory network, and predicts dirty regions that will undergo pixel changes in the next refresh cycle; it uses a Transformer model to predict potential modification regions that may be edited within a preset time window; after preprocessing the potential modification regions, it obtains pre-rendered candidate regions containing the dirty regions, performs pre-rendering, and encapsulates the results and corresponding prediction confidence scores into a standardized data structure, which is then sent to subsequent modules.

[0057] The intelligent waveform optimization module extracts key features of dirty regions and prediction confidence from a standardized data structure. When the prediction confidence is greater than or equal to a set confidence threshold, it refers to the relevant features of potential modification regions and dynamically selects waveform types from a preset waveform library according to a preset decision logic. Based on the dynamic adjustment rules, it optimizes the control parameters of the selected waveform type and outputs the target waveform parameters.

[0058] The ghosting elimination module generates a reverse compensation waveform based on the target waveform parameters and information about the dirty area to eliminate ghosting. The selected waveform type and the reverse compensation waveform are merged in chronological order to generate a drive instruction that instructs the display controller to refresh the screen and update the displayed content.

[0059] Furthermore, in the user behavior prediction module, the pre-processing of the potential modification region to obtain the pre-rendered candidate region containing the dirty region specifically involves:

[0060] When the prediction confidence is greater than or equal to the set confidence threshold, potential modification sub-regions are selected from the potential modification regions according to priority. The priority is based on a comprehensive ranking of confidence, ghosting risk index and region area. They are merged in order according to the ranking results until the area budget is reached. The sub-regions are then combined with the dirty regions to form pre-rendered candidate regions.

[0061] When the prediction confidence is less than the set confidence threshold, the sub-regions in the potential modification region are not merged into the dirty region, and only the boundary expansion of the dirty region is performed as a pre-rendering candidate region.

[0062] Furthermore, in the intelligent waveform optimization module, when the prediction confidence is greater than or equal to a set confidence threshold, the relevant features of the potential modification region are referenced, and the waveform type is dynamically selected from the preset waveform library according to a preset decision logic. The intelligent waveform optimization module selects the waveform type based on the following criteria:

[0063] When the prediction confidence is greater than or equal to the set confidence threshold, the waveform type is selected based on the information of the dirty area and the future risk of the potential modification area. If the ghosting risk index of the potential modification area at the target position is greater than or equal to the preset ghosting risk index threshold, a stable refresh curve is used in advance for the adjacent dirty areas or the phase reversal time is appropriately extended.

[0064] When the prediction confidence is less than the set confidence threshold, no refresh is triggered for the potential modification area; only parameter preset and caching are performed.

[0065] Furthermore, in the intelligent waveform optimization module, the waveform type is dynamically selected from the preset waveform library according to a preset decision logic, specifically as follows:

[0066] The key features of the dirty region include change frequency, character density, and content type;

[0067] When the frequency of change is greater than or equal to the preset frequency of change, select the fast waveform;

[0068] When the character density is greater than the preset character density, select the fast waveform;

[0069] When the ghosting risk index is greater than the preset ghosting risk index, select the high-quality waveform;

[0070] Conversely, select the mixed waveform.

[0071] Furthermore, in the intelligent waveform optimization module, the waveform library includes:

[0072] Fast waveform, suitable for high-frequency input, allows for image retention, basic parameter range: voltage gradient 2.5–4.5V / ms, refresh time 6–12ms;

[0073] High-quality waveforms, suitable for static display, ignoring ghosting; basic parameter range: voltage gradient 0.6–1.5V / ms, refresh time 40–80ms.

[0074] Hybrid waveform, suitable for intermediate frequency input, balances ghosting and delay, with basic parameters ranging from 1.5 to 3.0 V / ms for voltage gradient and 15 to 30 ms for refresh time.

[0075] The advantages of the spatiotemporal prediction and waveform optimization method and system for partial refresh of e-ink screens provided by this invention are as follows: By predicting user behavior characteristics and dynamically optimizing the refresh waveform, the response latency is reduced to below 15ms, significantly improving the fluency of code input. Students can see real-time feedback while inputting code, improving learning efficiency and experience. It fully leverages the backlight-free and flicker-free characteristics of e-ink screens, providing a healthy programming learning environment for children and teenagers. It reduces the risk of eye strain and vision damage caused by prolonged use. Simultaneously, through partial refresh and intelligent waveform optimization, power consumption is reduced, extending device battery life, making it suitable for long-term programming learning and conforming to green environmental protection principles. It provides children and teenagers with an eye-friendly and smooth programming learning tool, promoting the popularization and development of programming education, and has broad application prospects in scenarios such as distance education and programming competitions. Attached Figure Description

[0076] Figure 1 This is a schematic diagram of the process of the present invention;

[0077] Figure 2 A flowchart illustrating the user behavior prediction module;

[0078] Figure 3 A flowchart illustrating the intelligent waveform optimization module;

[0079] Figure 4 This is a flowchart illustrating the removal process of the afterimage removal module. Detailed Implementation

[0080] The technical solution of the present invention will now be described in detail through specific embodiments. Many specific details are set forth in the following description to provide a thorough understanding of the invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0081] like Figures 1 to 4 As shown, the spatiotemporal prediction and waveform optimization method for local refresh of e-ink screens proposed in this invention includes:

[0082] Collect user behavior features and screen status data and input them into the user behavior prediction module to predict dirty regions that will undergo pixel changes in the next refresh cycle and potential modification regions that may be edited within a preset time window; after preprocessing the potential modification regions, obtain the pre-rendered candidate regions containing the dirty regions, perform pre-rendering, and encapsulate the results and the corresponding prediction confidence into a standardized data structure for distribution.

[0083] The intelligent waveform optimization module extracts key features of dirty regions and prediction confidence from a standardized data structure. When the prediction confidence is greater than or equal to a set confidence threshold, it refers to the relevant features of potential modification regions and dynamically selects waveform types from a preset waveform library according to a preset decision logic. It optimizes the control parameters of the selected waveform type according to dynamic adjustment rules and outputs target waveform parameters, including voltage gradient and refresh time.

[0084] The ghosting elimination module generates a reverse compensation waveform based on the target waveform parameters and the information of the dirty region, combined with the extended boundary sub-region incorporated into the pre-rendered candidate region, to eliminate ghosting; the selected waveform type and the reverse compensation waveform are merged in chronological order to generate a driving instruction, which instructs the display controller to refresh the screen and update the display content.

[0085] This embodiment reduces the response latency to below 15ms by predicting user behavior characteristics and dynamically optimizing the refresh waveform, significantly improving the fluency of code input. Students can see real-time feedback while inputting code, improving learning efficiency and experience. It also fully leverages the backlight-free and flicker-free characteristics of e-ink screens to provide a healthy programming learning environment for children and teenagers, and reduces the risk of eye fatigue and vision damage caused by prolonged use.

[0086] Meanwhile, by using partial refresh and intelligent waveform optimization, power consumption is reduced and device battery life is extended, making it suitable for long-term programming learning and in line with the concept of green environmental protection. It provides children and teenagers with an eye-friendly and smooth programming learning tool, promotes the popularization and development of programming education, and has broad application prospects in scenarios such as distance education and programming competitions.

[0087] It should be noted that the definitions of the aforementioned dirty areas, potential modification areas, pre-rendered candidate areas, and ghosting risk index are as follows:

[0088] Dirty area The set of display areas that are predicted to undergo pixel changes in the next refresh cycle or a very short time window, satisfying the following conditions:

[0089] ;

[0090] in, The coordinates of the pixel location point, For pixel difference, This is the pixel difference threshold.

[0091] Potential modification areas Based on context and user trajectory prediction, this is a set of display areas that may be edited or redrawn within a preset time window.

[0092] Pre-rendered candidate regions : The target region used for background pre-rendering and resource prefetching, including dirty regions. When prediction confidence and area budget allow, some sub-regions of potentially modified regions are incorporated into the dirty region to obtain pre-rendering candidate regions, satisfying the following:

[0093] .

[0094] Ghosting Risk Index A comprehensive indicator used to characterize the risk of visually perceptible ghosting, which can be calculated based on features such as the mean pixel difference, edge intensity, and cross-frame persistence, specifically:

[0095] ;

[0096] in, The average pixel difference. For edge energy, For cross-frame persistence, To adjust the parameters, the preferred method is... The values ​​are 0.4, 0.3, and 0.3 respectively.

[0097] This embodiment mainly achieves its goal through a user behavior prediction module, an intelligent waveform optimization module, and a ghosting elimination module. The following is a detailed description of each module:

[0098] I. User Behavior Prediction Module;

[0099] Based on LSTM (Long Short-Term Memory) network and Transformer model,

[0100] like Figure 2 As shown, this module uses a Long Short-Term Memory (LSTM) network to capture short-term time-series features of user behavior and predicts dirty regions that will undergo pixel changes in the next refresh cycle. It also employs a Transformer model to predict potential editable regions that may be modified within a preset time window. Unlike traditional methods that rely on historical user input data, this module models real-time features such as commonly used library functions of programming languages, input frequency, cursor position, and code block structure to predict screen area changes in advance and pre-render them, thereby minimizing latency during dynamic refresh processes.

[0101] Technical implementation: Analyze user input (such as key type, cursor movement trajectory) and screen state (such as pixel distribution and code structure in the code editing area) to perform behavior prediction and dirty area prediction.

[0102] 1) Input pattern modeling: The LSTM model captures the temporal features of input behavior, analyzes input features such as cursor position and commonly used functions in real time, and accurately predicts the upcoming changes in the screen area.

[0103] 2) Contextual Information Modeling: Utilizing the Transformer model to model complex programming language syntax structures, identifying potential areas of input change, particularly predicting common operations such as library functions and variable assignments. This includes scenarios like jump modifications to code blocks or high-frequency dynamic updates.

[0104] The way LSTM and Transformer are combined in this embodiment:

[0105] Structure fusion: The LSTM model is first used to capture short-term time series features and predict the input behavior and dirty regions at the current moment, while the Transformer model is responsible for capturing global spatiotemporal information and more complex long-range dependencies. Especially in editing scenarios, it handles the behavior of multiple regions or jump modifications, predicts the potential modification regions that may be edited within a preset time window, and predicts the confidence level.

[0106] Joint training: Through multi-task learning, LSTM and Transformer are jointly optimized. LSTM outputs local time series prediction results, Transformer performs global dependency modeling, and finally outputs dirty region prediction.

[0107] Predictive Co-training: LSTM and Transformer are trained in parallel, and the prediction ability of input patterns is optimized through fine-grained co-calibration. Combining the advantages of both, the prediction accuracy of input behavior and dirty regions is improved.

[0108] This embodiment innovatively combines LSTM and Transformer models to predict user behavior. By predicting user input behavior and dirty regions through temporal modeling, it pre-renders the content of dirty regions, thereby reducing dynamic refresh latency. This prediction mechanism can identify potential areas of user input and use the model's prediction of the cursor trajectory to determine which areas need local updates, breaking the passive refresh mode of traditional e-ink screens.

[0109] The user behavior prediction module in this embodiment outputs dirty regions, potential modification regions, and their prediction confidence scores. The preprocessing of the potential modification regions to obtain pre-rendered candidate regions containing the dirty regions specifically involves:

[0110] When the prediction confidence is greater than or equal to the set confidence threshold, several potential modification sub-regions are selected from the potential modification regions according to priority, and are incorporated into the potential modification regions without exceeding the area budget, to form pre-rendered candidate regions;

[0111] When the prediction confidence is less than the set confidence threshold, the dirty region is not incorporated into the potential modification region. Instead, the dirty region is expanded to cover the stroke expansion and anti-jitter requirements, thus forming a pre-rendered candidate region.

[0112] Finally, pre-rendering is performed on the candidate regions in the background buffer (specifically, pixels or image blocks are generated in advance in the background buffer). The results and the corresponding prediction confidence are then encapsulated into a standardized data structure and sent to subsequent modules to avoid invalid refreshes and resource waste.

[0113] Describe the user behavior prediction module in detail according to the specific scenario, including (a1) to (a6):

[0114] (a1) Scenario: Students input code on a programming learning machine, for example:

[0115] print("Hello, World!");

[0116] The collected data includes user behavior characteristics and screen status data.

[0117] Input behaviors include:

[0118] Key types: letters, numbers, delete key, newline key, etc.

[0119] Cursor movement trajectory: Records cursor coordinates Changes;

[0120] Input interval time: the time difference between two key presses (e.g., 0.2s).

[0121] Screen status data includes: pixel distribution in the current code editing area (such as character position and line number changes), and code structure information: function definitions, loop statements, comments, etc.

[0122] (a2) Behavioral modeling;

[0123] Use LSTM models to capture temporal features (such as the regularity of continuously input letters):

[0124] User input: The 10 most recent key presses (e.g., ('H', 'e', ​​'l', 'l', 'o')).

[0125] Output: Predict the cursor position for the next 3 frames (e.g., ...) ).

[0126] Use Transformer models to capture long-range dependencies (such as cursor jumps after a line break):

[0127] User input: code structure information (such as the syntax tree of print("Hello");

[0128] Output: Predicted potential modification regions (e.g., inside the string).

[0129] (a3) Predict dirty areas;

[0130] User input: Current cursor position: The last 10 key presses were: ('H', 'e', ​​'l', 'l', 'o').

[0131] The model predicts the cursor position for the next 3 frames: if the student inputs letters consecutively, the predicted cursor position is to the right. If the delete key is detected, the cursor is predicted to move to the left. .

[0132] Calculate the dirty region range based on the predicted cursor position: if the prediction moves to the right, the dirty region is... arrive A rectangular area.

[0133] (a5) Predict potential modification regions; based on the prediction confidence and area budget, gate and clip the potential modification regions to form pre-rendered candidate regions that include the dirty regions and are within the budget range.

[0134] (a6) Pre-rendering;

[0135] After pre-rendering the candidate regions, the data is combined with the prediction confidence and transformed into a standardized data structure (such as JSON format) and then distributed to the backend buffer.

[0136] Pre-render dirty regions in the screen's back buffer. Example: Predict that a student will input the character "H" and render it in advance. The position renders the pixel value of the character "H".

[0137] (a6) Data is transmitted to the intelligent waveform optimization module;

[0138] Transmission method: Data is transmitted in real time via the message queue ZeroMQ to ensure that each module can quickly receive the required data for subsequent processing.

[0139] Transmission logic:

[0140] 1) High-priority channel: When high-frequency input is detected (e.g., ≥5 keystrokes per second), data will be transmitted via the real-time message queue ZeroMQ to ensure that data can be delivered to downstream modules in real time with minimal latency. High-frequency input scenarios require special handling, prioritizing display smoothness and response speed. Therefore, data transmission will be processed through a high-priority channel to reduce any latency.

[0141] 2) Standardized format:

[0142] 2-1) To ensure seamless data transfer between different modules, all data will be encapsulated in JSON format. JSON is a cross-platform, cross-language standard format that ensures data interoperability between different systems and modules.

[0143] For example, the data content may include: pre-rendered candidate regions (e.g., the areas on the screen that need to be updated), input frequency (number of keystrokes per second), character density (the density of characters in the region), cursor position (the coordinates of the current cursor position), and prediction confidence.

[0144] 2-2) The transmitted data will undergo corresponding format conversion and processing according to the parsing requirements of different modules.

[0145] For example:

[0146] This embodiment pre-renders by predicting user input behavior and pre-rendering candidate regions, minimizing refresh latency. By combining LSTM and Transformer models and employing multi-task learning, it achieves accurate prediction of short-term behavior and long-term dependencies, thereby reducing latency during dynamic refresh.

[0147] II. Intelligent Waveform Optimization Module;

[0148] Based on deep learning algorithms, this module dynamically selects appropriate waveform optimization strategies to reduce latency, power consumption, and eliminate ghosting. Through reinforcement learning, it intelligently selects different waveform types (such as fast waveforms, high-quality waveforms, and mixed waveforms) based on information such as input frequency and character density to optimize the display effect.

[0149] Core functions:

[0150] 1) Waveform selection: Dynamically select fast waveform or high-quality waveform based on input characteristics (such as frequent input, static display, etc.).

[0151] 2) Waveform adjustment: Adjust waveform parameters for different input scenarios (such as high-frequency input, annotated text, etc.) to balance display quality and latency.

[0152] like Figure 3 The intelligent waveform optimization module, afterimage handling module, and display controller shown are described in detail in (b1) to (b3) as follows:

[0153] (b1) In the intelligent waveform optimization module, the waveform library is constructed as shown in Table 1:

[0154] Table 1

[0155]

[0156] Waveform characteristics:

[0157] Fast waveform: Suitable for scenarios where students quickly input code and frequently debug. It prioritizes reducing latency, and ghosting will be eliminated through subsequent compensation refresh.

[0158] High-quality waveforms: Suitable for displaying static content such as code comments and docstrings, completely eliminating ghosting;

[0159] Hybrid waveform: Suitable for scenarios where observation is paused after code modification, balancing refresh efficiency and display quality.

[0160] As preferred options for the basic parameter range, the following waveforms are used: fast waveform: voltage gradient 3.0–4.0V / ms, refresh time 8–10ms; high-quality waveform: voltage gradient 0.8–1.2V / ms, refresh time 50–60ms; mixed waveform: voltage gradient 2.0–3.0V / ms, refresh time 18–24ms.

[0161] The optimized scheme within this basic parameter range enables various waveforms to achieve better display effects in their respective application scenarios: fast waveforms reduce energy consumption while ensuring low latency, high-quality waveforms eliminate ghosting while maintaining contrast, and hybrid waveforms achieve a balance between refresh efficiency and ghosting control. This optimized scheme further enhances the feasibility and stability of the method.

[0162] (b2) Waveform selection;

[0163] Intelligent waveform optimization module: Receives key features of dirty regions (including coordinates, character density, change frequency, etc.) and prediction confidence extracted from standardized data structures, selects waveform type according to preset decision logic, and generates target waveform parameters (including voltage gradient, refresh time, effective area, etc.) according to dynamic adjustment rules.

[0164] Specifically:

[0165] 1) Parsing the incoming input data:

[0166] Key features of dirty regions extracted from standardized data structures and prediction confidence:

[0167] Change frequency: The number of inputs per second;

[0168] Character density (char_density): The character density of a region, used to evaluate the complexity of the input region;

[0169] Content type (content_type): Distinguishes between content types such as code input, comments, and debugging.

[0170] 2) The intelligent waveform optimization module selects the waveform type based on prediction confidence and key features of the dirty region;

[0171] Regarding prediction confidence:

[0172] When the prediction confidence is greater than or equal to the set confidence threshold, the waveform type is selected based on the information of the dirty area and the future risk of the potential modification area. If the ghosting risk index of the potential modification area at the target position is greater than or equal to the preset ghosting risk index threshold, a stable refresh curve or an appropriate extension of the phase-out time is used for the adjacent dirty areas in advance.

[0173] When the prediction confidence is less than the set confidence threshold, no refresh is triggered for the potential modification area; only parameter preset and caching are performed.

[0174] When the area budget is insufficient, fast refresh waveforms are prioritized for dirty areas, while hybrid or high-quality waveforms are used for delayed processing of selected potential modification sub-areas.

[0175] Targeting key features of the dirty area, such as Figure 3 As shown, the decision tree decision logic is as follows:

[0176] When the frequency of change is greater than or equal to the preset frequency of change (e.g., 5 times / second), select the fast waveform;

[0177] When the character density is greater than the preset character density (e.g., 0.8), select the fast waveform;

[0178] When the ghosting risk index is greater than the preset ghosting risk index (e.g., 0.1), select a high-quality waveform;

[0179] Conversely, select the mixed waveform.

[0180] Under the waveform selection logic described in (b2), the following detailed rules (c1) to (c2) are further given for different input scenarios:

[0181] (c1) High-frequency input priority:

[0182] This corresponds to the specification of high-frequency input conditions in the decision tree logic when selecting a fast waveform, provided that the change frequency is greater than or equal to the preset change frequency (e.g., 5 times / second).

[0183] Condition: The student is rapidly inputting code (e.g., inputting a loop statement for i in range(10):), pressing the key ≥ 5 times per second;

[0184] Decision: Select a fast waveform (refresh time 8ms, voltage gradient 3V / ms);

[0185] Technical basis: High-frequency inputs require priority to ensure response speed, and afterimages are eliminated through subsequent compensation and refresh.

[0186] (c2) Static content optimization:

[0187] In scenarios where the character density is high but the frequency of change is low, the static display conditions in the decision tree logic are specified when selecting a high-quality waveform.

[0188] Conditions: Dirty regions are code comments (such as # for calculating the sum of squares), with high character density but low change frequency (<1 time / second);

[0189] Decision: Select a high-quality waveform (refresh time 50ms, voltage gradient 1V / ms).

[0190] Technical basis: Static content must completely eliminate ghosting and tolerate higher latency.

[0191] (b3) Waveform generation and parameter optimization;

[0192] In the intelligent waveform optimization module, the control parameters (such as voltage gradient and refresh time) of the selected waveform type are optimized according to dynamic adjustment rules, and the target waveform parameters are output. The specific dynamic adjustment rules are shown in Table 2.

[0193] Table 2

[0194]

[0195] Decision logic rule: If the frequency of change is ≥ 5 times / second, select the fast waveform.

[0196] Dynamic parameter adjustment: Due to the dirty area area of ​​400 pixels 2 >100 pixels 2 The voltage gradient increases from 3V / ms to 4V / ms; specifically, the voltage gradient can be adjusted according to the following formula:

[0197] ;

[0198] in, and These represent the adjusted voltage gradient and the voltage gradient before adjustment, respectively, and the area of ​​the region corresponds to the dirty region area.

[0199] After selecting the waveform type in (b2), the system will output the target waveform parameters according to the dynamic adjustment rules in (b3).

[0200] The intelligent waveform optimization module in this embodiment dynamically selects the optimal waveform using a deep learning algorithm to ensure low power consumption, low latency, and low ghosting in the display. Waveform optimization combines information such as input features, frequency, and character density to balance display quality, power consumption, and latency.

[0201] Meanwhile, the decision tree algorithm intelligently selects waveform types based on features such as input frequency, character density, and ghosting risk index, achieving a balance between low latency and high display quality.

[0202] III. Afterimage Removal Module;

[0203] (d1) Ghosting detection: By comparing the data of the previous and next frames, the ghosting area is identified and a compensation waveform is generated to eliminate the ghosting in real time.

[0204] 1) Comparison of data between previous and previous frames: By calculating the pixel differences between the current frame and the previous frame, potential ghosting areas can be identified.

[0205] 2) Visual feature enhancement: Enhance the visibility of afterimage areas through edge detection and other methods.

[0206] (d2) Image retention compensation: Based on the reverse electric field waveform, the image retention area is cleared in real time to ensure that the screen display is clean and clear.

[0207] Reverse waveform generation: Generates waveforms based on the intensity of the afterimage region (e.g., applying a reverse voltage to the afterimage region to eliminate residual pigment) and applies it to the display screen.

[0208] For fast waveforms, the ghosting elimination module will remove ghosting by compensating the waveform during the refresh process.

[0209] The goal of this step is to ensure that, after waveform selection, the display quality of the e-ink screen is as clear as possible, without obvious afterimages or visual defects.

[0210] The implementation steps of the afterimage removal module (taking the student's modified code print("Hello")→print("Hi") as an example) are as follows (e1) to (e3):

[0211] (e1) Calculate the pixel difference between the dirty regions in the current frame and the previous frame; if the difference between the two is greater than the pixel difference threshold, it is marked as a potential ghosting region:

[0212] Data source:

[0213] Current frame buffer: Stores the content currently displayed on the screen, such as a modified print("Hi");

[0214] History frame buffer: Stores the content displayed in the previous frame, such as print("Hello").

[0215] Comparison logic:

[0216] Calculate the pixel difference between the dirty regions (regions that have changed) in the current frame and the previous frame. Use the character "o" at position... For example, to detect changes in pixel values ​​at a given location; the formula is:

[0217] ;

[0218] in, The difference value of pixels in the dirty region. and The current frame ( (moment) and the previous frame ( (Time) at coordinates The pixel value of the location;

[0219] when Then it is marked as a potential afterimage region, where This is the pixel difference threshold.

[0220] (e2) Image enhancement is performed on potential ghosting regions through contrast stretching and edge detection, and a list of ghosting region coordinates is generated;

[0221] When pixel differences are small, but afterimages are still visually perceptible (such as light-colored character residue), image enhancement techniques can be applied to improve the visibility of the afterimage area and further improve the accuracy of afterimage detection.

[0222] Implementation: Image enhancement techniques, including contrast stretching and edge detection (Canny edge detection in OpenCV), are used to enhance the marked region. Edge detection further enhances the afterimage region, thereby improving its visibility and marking the area.

[0223] Output: Generates a list of coordinates for the afterimage region, for example... .

[0224] Optimization merging: Adjacent afterimage pixels are merged into a single rectangular region to reduce the computational load in subsequent processing.

[0225] (e3) Reverse the voltage polarity of each afterimage region, and dynamically adjust the duration of the reversal based on the afterimage intensity, so as to return the generated reverse compensation waveform;

[0226] Ghosting elimination: software and hardware work together;

[0227] The core idea is to generate a reverse compensation waveform through software algorithms and then invoke the hardware refresh mechanism of the e-ink screen to eliminate ghosting. This solution, through the coordinated work of software and hardware, precisely removes ghosting areas, ensuring the clarity and high refresh rate of each frame.

[0228] like Figure 4 As shown, the technical solution of the image retention module includes steps 1 to 3:

[0229] Step 1: Reverse waveform generation (software);

[0230] Reverse voltage polarity reversal: The software generates a voltage polarity waveform that is opposite to that of the previous frame based on the intensity of the afterimage area to clear the afterimage.

[0231] Duration adjustment: The duration of the waveform is dynamically adjusted based on the detected afterimage intensity (determined by the change in the afterimage area). Stronger afterimages require a longer reverse electric field.

[0232] Compensation waveform generation: By evaluating the intensity of the afterimage, the voltage and application time are dynamically adjusted to generate a compensation waveform that adapts to the afterimage region.

[0233] It should be noted that reverse voltage polarity inversion involves reversing the voltage in each ghosting region to completely eliminate ghosting. By reversing the voltage polarity in the ghosting region, the removal of previously residual pixel information can be maximized.

[0234] Step 2: Waveform merging and drive instruction generation;

[0235] Merge logic:

[0236] The selected waveform type (i.e., the main refresh waveform (such as a fast waveform or a hybrid waveform)) is merged with the reverse compensation waveform in chronological order to ensure that the compensation waveform is executed after the main refresh waveform, thereby maximizing the ghosting removal effect.

[0237] Output driver instructions: Generate the final driver instructions, instructing the display controller to refresh the screen and update the displayed content.

[0238] Step 3: Hardware driver refresh (hardware);

[0239] Hardware interface call: Send the merged waveform data to the display controller through the standard driver interface (SPI and I2C) of the e-ink screen;

[0240] Example command: send_waveform_to_controller(ghost_region, waveform_data).

[0241] Physical removal of afterimages:

[0242] The display controller applies a reverse electric field to the afterimage area via electrodes, dispersing residual pigment particles. Precise voltage control ensures complete removal of the afterimage. Effect: For example, the residual "o" character is removed, resulting in a clear "Hi".

[0243] This embodiment precisely addresses ghosting of varying intensities by dynamically adjusting the voltage polarity and duration of the reverse waveform. Compared to traditional reverse waveform generation methods, this approach better adapts to various types of ghosting, ensuring display quality. Precise timing control is implemented during waveform merging to ensure that the reverse compensation waveform always executes immediately after the selected waveform type (i.e., the main refresh waveform), thereby maximizing the ghosting removal effect. This effectively avoids potential timing discrepancies during waveform merging, ensuring consistent display performance.

[0244] Highly efficient ghosting removal is achieved through the collaborative work of software and hardware. The software is responsible for detecting the ghosting area and generating a compensation waveform, while the hardware precisely drives the electrodes to remove the ghosting based on the generated waveform. This software-hardware cooperation ensures display accuracy and a high refresh rate, making it particularly suitable for high-frequency interactive applications—such as programming education.

[0245] In high-frequency input scenarios, the system maintains extremely low latency by dynamically adjusting the voltage gradient, ensuring real-time feedback to user input while maintaining stable display quality. This method enhances the dynamic display performance of e-ink screens, making them suitable for rapidly updating and highly interactive environments.

[0246] IV. Waveform data is transmitted to the display controller;

[0247] Display controller: Receives the target waveform parameters output by the intelligent waveform optimization module and the reverse compensation waveform output by the ghosting elimination module. It merges the selected waveform type and the reverse compensation waveform in chronological order to generate a driving command, which drives the electronic paper display panel to apply an electric field to the target area to perform local refresh and compensation refresh, eliminate ghosting, and ensure that the final displayed content meets expectations.

[0248] Combining points one through four, this embodiment demonstrates that the hardware (e-ink display) and software (waveform optimization, ghosting elimination algorithm) work closely together to ensure that each frame can be updated quickly and accurately, reducing ghosting and improving the user experience.

[0249] The intelligent waveform optimization module in this embodiment selects the optimal waveform in real time based on the dynamic characteristics of user input, thereby balancing display latency, power consumption, and display quality. This optimization strategy ensures the efficient operation of the e-ink screen by intelligently adjusting parameters such as input frequency, character density, and ghosting risk index.

[0250] In high-frequency input scenarios, the system prioritizes fast waveforms (low latency), while in static or low-frequency input scenarios, it selects high-quality waveforms (low ghosting). This effectively improves screen response speed and eliminates ghosting issues during frequent updates.

[0251] This embodiment achieves proactive prediction through spatiotemporal prediction technology, enabling the e-ink screen to predict potential update areas before user input and pre-render them in the background. This technology changes the traditional passive update mechanism, making the local refresh process smoother and more instantaneous, thus improving the user experience.

[0252] In addition, by predicting dirty areas in real time and pre-rendering them, refresh latency is reduced, enhancing the smooth experience in dynamic scenarios such as interactive programming.

[0253] Example 1: Application of programming in scenarios for children and teenagers;

[0254] Scenario: Students frequently modify loop variables, including (f1) to (f5), while debugging code;

[0255] (f1) Initial state: The code shows for i in range(5);

[0256] (f2) Modification operation: The student changes 5 to 10, and the screen updates to for i in range(10);

[0257] (f3) Ghosting detection: The software detected that there was a ghosting at the position of the number 5 (the difference value exceeded the threshold).

[0258] Potential modification areas:

[0259] The system analyzes the context using the Transformer model and predicts that the cyclic interval may still be modified again (e.g., changing 10 to 20), marking the "numerical parameter area" as a potential modification area.

[0260] Pre-rendered candidate regions:

[0261] Since the prediction confidence level is higher than the preset confidence threshold, the system merges the potential modification region (digital parameter region) into the dirty region to form a pre-rendering candidate region and performs background pre-rendering.

[0262] Intelligent waveform optimization module selection:

[0263] Input features: The key features of the dirty region are approximately 16×16 pixels in size and change at a frequency of 2 times per second;

[0264] The prediction confidence level is greater than or equal to the set confidence threshold;

[0265] Decision: Select a hybrid waveform with a refresh time of 20ms and a voltage gradient of 2.5V / ms to achieve a balance between ghosting and delay.

[0266] (f4) Image Retention: Generates a reverse compensation waveform to drive the e-ink screen to refresh locally in the digital area.

[0267] (f5) Final display: Number 10 is clearly displayed, and number 5 has no afterimage.

[0268] Example 2: Students use an e-ink screen programming learning machine to write Python code, changing print("Hello") to print("Hi"), as shown in steps (g1) to (g3):

[0269] Modification process: print("Hello") → print("Hi");

[0270] (g1) Spatiotemporal prediction: User behavior capture and dirty region prediction, including (h1) to (h2):

[0271] (h1) Input event capture;

[0272] Hardware trigger: The student presses Backspace to delete the character "o";

[0273] Data collection:

[0274] Key type: Delete key;

[0275] Cursor position: Move left from the right of "o" to after "l";

[0276] Code context: The current code action is print("Hello").

[0277] (h2) Behavioral modeling and prediction;

[0278] LSTM inference: capturing short-term temporal features and predicting dirty regions after deletion operations. = The pixel rectangle range of "o" (140, 200, 160, 220).

[0279] Transformer inference: Recognizing the context as string editing and predicting potential modification areas. = "Hell" may add a new character position (empty space on the right side of the entire string).

[0280] Pre-rendered candidate region generation :

[0281] The prediction confidence level is 0.87, which is greater than or equal to the preset confidence threshold (0.8). Therefore, from the potential modification region... Incorporating a portion of the sub-regions into the dirty region, and generating pre-rendered candidate regions accordingly. ;

[0282] form ={ +The location near where "i" might be entered;

[0283] Backend Render the pixel buffer in advance to reduce subsequent refresh latency.

[0284] (g2) Intelligent waveform optimization module: waveform selection and parameter optimization, including (i1) to (i2):

[0285] (i1) Waveform selection decision;

[0286] Input data:

[0287] Dirty region characteristics: area 20×20 pixels, change frequency ≈ 3 times / second;

[0288] Potential modification area: Predicted new character "i", ghosting risk index = 0.12;

[0289] The prediction confidence level is 0.87, which is greater than or equal to the set confidence threshold.

[0290] Decision-making logic:

[0291] High-frequency editing tends to favor fast waveforms;

[0292] However, since the afterimage risk index = 0.12 ≥ 0.1 → improves stability, a hybrid waveform is selected.

[0293] (i2) Waveform generation and dynamic parameter adjustment;

[0294] Waveform type: Hybrid waveform (suitable for intermediate frequency input, taking into account delay and ghosting).

[0295] Initial parameters: voltage gradient 2.5V / ms, refresh time 20ms.

[0296] Dynamic adjustment: due to dirty areas The area is 400 pixels 2 >100 pixels 2 →Voltage increased to 3.0V / ms.

[0297] Final target waveform parameters: {voltage gradient = 3.0V / ms, refresh time = 20ms, effective area =} }

[0298] Distribution method: The target waveform parameters are transmitted to the afterimage elimination module and the display controller via an asynchronous message queue.

[0299] (g3) Afterimage removal and display output, including (j1) to (j3);

[0300] (j1) Ghosting detection;

[0301] enter:

[0302] Current frame = the deleted print("Hell") + the newly added "i";

[0303] History frame = print("Hello").

[0304] Detection logic: Calculate the pixel difference value of dirty areas. .

[0305] Result: The grayscale difference at the "o" position was detected to be >15, indicating significant edge energy → marked as a ghosting region.

[0306] (j2) Afterimage compensation;

[0307] Compensation waveform generation:

[0308] Polarity: opposite;

[0309] Voltage amplitude: Va = -3V (if the ghosting risk index > 0.2, it can be increased to -3.5V);

[0310] Duration: 10ms (can be extended to 12ms depending on the afterimage risk index);

[0311] Region: The rectangle (140, 200, 160, 220) that overlaps with the afterimage of "o".

[0312] (j3) Display driver;

[0313] Main refresh: Generates pre-rendered candidate regions Apply a mixed waveform to display "Hi";

[0314] Compensation refresh: Apply a reverse electric field to the "o" afterimage area to completely remove the residue;

[0315] Physical effect: "i" is displayed normally within 20ms, and the afterimage of "o" is completely eliminated.

[0316] Control command: Merge the main waveform and the reverse compensation waveform, and send the command to the display controller via the SPI interface.

[0317] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An electronic ink screen local refresh method based on space-time prediction and waveform optimization, characterized in that, The method comprises the following steps: Collecting user behavior features and screen state data and inputting them into a user behavior prediction module to predict dirty regions that will have pixel changes in the next refresh cycle and potential modification regions that may be edited within a preset time window; preprocessing the potential modification regions to obtain a pre-rendering candidate region containing the dirty regions, performing pre-rendering, and encapsulating the results and the corresponding prediction confidence into a standardized data structure for delivery; The intelligent waveform optimization module selects a waveform type from a preset waveform library based on the key features of the dirty regions and the prediction confidence extracted from the standardized data structure, and refers to the related features of the potential modification region when the prediction confidence is greater than or equal to a set confidence threshold, and follows a preset decision logic; the control parameters of the selected waveform type are optimized according to the dynamic adjustment rules, and the target waveform parameters are output; The image sticking elimination module generates a reverse compensation waveform based on the target waveform parameters and the information of the dirty regions, and combines the extended boundary sub-region in the pre-rendering candidate region to eliminate image sticking; the selected waveform type and the reverse compensation waveform are merged in chronological order to generate a driving instruction to instruct the display controller to refresh the screen and update the display content.

2. The refresh method of claim 1, wherein, The method of collecting user behavior features and screen state data and inputting them into a user behavior prediction module to predict dirty regions that will have pixel changes in the next refresh cycle and potential modification regions that may be edited within a preset time window comprises the following steps: Using a long short-term memory network to capture the short-term time sequence features of user behavior features to predict dirty regions that will have pixel changes in the next refresh cycle; Using a Transformer model to predict potential modification regions that may be edited within a preset time window.

3. The refresh method of claim 1, wherein, In the user behavior prediction module, the preprocessing of the potential modification regions to obtain a pre-rendering candidate region containing the dirty regions comprises the following steps: When the prediction confidence is greater than or equal to the set confidence threshold, potential modification sub-regions are selected from the potential modification region according to priority, the priority is comprehensively sorted according to factors such as confidence, image sticking risk index, and region area, and the sorting results are sequentially incorporated until the area budget is reached, and the union of the dirty regions forms a pre-rendering candidate region; When the prediction confidence is less than the set confidence threshold, the sub-regions in the potential modification region are not incorporated into the dirty region, and only boundary expansion of the dirty region is performed as a pre-rendering candidate region.

4. The refresh method of claim 1, wherein, In the intelligent waveform optimization module, when the prediction confidence is greater than or equal to the set confidence threshold, the related features of the potential modification region are referred to, and a waveform type is dynamically selected from a preset waveform library according to a preset decision logic, wherein the intelligent waveform optimization module selects a waveform type according to the following rules: When the prediction confidence is greater than or equal to the set confidence threshold, the selection of the waveform type is based on the information of the dirty regions and the future risk of the potential modification region, and if the image sticking risk index of the potential modification region at the target position is greater than or equal to a preset image sticking risk index threshold, a stable refresh curve is used for the adjacent dirty region in advance or the phase-out time is appropriately prolonged. When the prediction confidence is less than the set confidence threshold, no refresh is triggered for the potential modification region, only parameter preset and cache are performed.

5. The refresh method of claim 4, wherein, In the intelligent waveform optimization module, a waveform type is dynamically selected in a preset waveform library according to a preset decision logic, and specifically: The dirty region key features include a change frequency, a character density and a content type. When the change frequency is greater than or equal to a preset change frequency, a fast waveform is selected. When the character density is greater than a preset character density, a fast waveform is selected. When the residual image risk index is greater than a preset residual image risk index, a high-quality waveform is selected. Conversely, a mixed waveform is selected.

6. The refresh method of claim 1, wherein, The user behavior features include a key type, a cursor moving track and an input interval time. The screen state data includes pixel distribution and code structure information of a current code editing region.

7. The refresh method of claim 1, wherein, In the intelligent waveform optimization module, the waveform library includes: The fast waveform is suitable for high-frequency input, allows residual image to exist, and the basic parameter range is: voltage gradient 2.5-4.5 V / ms, refresh time 6-12 ms. The high-quality waveform is suitable for static display, ignores residual image, and the basic parameter range is: voltage gradient 0.6-1.5 V / ms, refresh time 40-80 ms. The mixed waveform is suitable for medium-frequency input, balances residual image and delay, and the basic parameter range is: voltage gradient 1.5-3.0 V / ms, refresh time 15-30 ms.

8. The refresh method of claim 7, wherein, The basic parameter range is: Fast waveform: voltage gradient 3.0-4.0 V / ms, refresh time 8-10 ms. High-quality waveform: voltage gradient 0.8-1.2 V / ms, refresh time 50-60 ms. Mixed waveform: voltage gradient 2.0-3.0 V / ms, refresh time 18-24 ms.

9. The refresh method of claim 1, wherein, In the intelligent waveform optimization module, the control parameters of the selected waveform type are optimized according to a dynamic adjustment rule, and target waveform parameters are output, wherein the dynamic adjustment rule is as follows: a fast waveform if the dirty region area is greater than 100 pixels 2 , increasing the voltage gradient; if the risk of ghosting of the potential modification region is greater than or equal to a preset risk of ghosting index threshold, then the phase-out time is extended in advance; For the high-quality waveform, if the temperature is greater than 30°C, the voltage gradient is reduced; if the residual image risk index is greater than 0.15, the refresh time is extended. For the mixed waveform, the voltage gradient is linearly adjusted according to the residual image risk index; if the residual image risk of the potential modification region is greater than or equal to a preset residual image risk index threshold, the lower limit of the voltage gradient is increased.

10. The refresh method of claim 1, wherein, The residual image elimination module generates a reverse compensation waveform according to the target waveform parameters and the information of the dirty region, for eliminating residual image, and specifically: The pixel difference between the current frame and the previous frame is calculated, and when the difference value between the two is greater than a pixel difference threshold, the potential residual image region is marked; The potential residual image region is image-enhanced through contrast stretching and edge detection to generate a residual image region coordinate list; The voltage polarity of each residual image region is reversed, and the duration of the reversal is dynamically adjusted according to the residual image intensity, and the generated reverse compensation waveform is returned.

11. An electronic ink screen partial refresh system with spatio-temporal prediction and waveform optimization, characterized in that, The user behavior prediction module, the intelligent waveform optimization module and the residual image elimination module are included. The user behavior prediction module captures the time sequence characteristics of the short-term user behavior characteristics of the input based on a long short-term memory network, predicts a dirty area in which a pixel change will occur in a next refresh cycle, and predicts a potential modification area that is likely to be edited in a preset time window using a Transformer model; the potential modification area is preprocessed to obtain a pre-rendering candidate area containing the dirty area, pre-rendering is performed, and the result is encapsulated into a standardized data structure together with a corresponding prediction confidence, and is delivered to a subsequent module; The intelligent waveform optimization module selects a waveform type in a preset waveform library according to a preset decision logic based on the key features of the dirty area and the prediction confidence extracted from the standardized data structure, and refers to the related features of the potential modification area when the prediction confidence is greater than or equal to a set confidence threshold; and optimizes the control parameters of the selected waveform type according to a dynamic adjustment rule, and outputs target waveform parameters; The residual image elimination module generates a reverse compensation waveform according to the target waveform parameters and the information of the dirty area, and uses the reverse compensation waveform to eliminate residual images; and generates a driving instruction by merging the selected waveform type and the reverse compensation waveform in time sequence, and instructs the display controller to refresh the screen and update the display content.

12. The refresh system of claim 11, wherein, In the user behavior prediction module, the preprocessing of the potential modification area to obtain the pre-rendering candidate area containing the dirty area is specifically as follows: When the prediction confidence is greater than or equal to the set confidence threshold, a potential modification sub-area is selected from the potential modification area according to a priority, the priority is comprehensively sorted according to the confidence, a residual image risk index and an area size, and the sorted results are sequentially integrated until the area budget is reached, and a union set of the dirty area and the potential modification sub-area forms a pre-rendering candidate area; When the prediction confidence is less than the set confidence threshold, the sub-area in the potential modification area is not integrated into the dirty area, and only a boundary expansion of the dirty area is performed as a pre-rendering candidate area.

13. The refresh system of claim 11, wherein, In the intelligent waveform optimization module, when the prediction confidence is greater than or equal to the set confidence threshold, the related features of the potential modification area are referred to, and a waveform type is dynamically selected in a preset waveform library according to a preset decision logic, wherein the intelligent waveform optimization module selects the waveform type according to the following: When the prediction confidence is greater than or equal to the set confidence threshold, the selection of the waveform type is based on the information of the dirty area and the future risk of the potential modification area, and if the residual image risk index of the potential modification area at a target position is greater than or equal to a preset residual image risk index threshold, a stable refresh curve is used for adjacent dirty areas in advance or the phase-out time is appropriately prolonged; When the prediction confidence is less than the set confidence threshold, no refresh is triggered for the potential modification area, and only parameter presetting and caching are performed.

14. The refresh system of claim 11, wherein, In the intelligent waveform optimization module, the waveform type is dynamically selected in the preset waveform library according to the preset decision logic, and specifically: The key features of the dirty area include a change frequency, a character density and a content type; When the change frequency is greater than or equal to a preset change frequency, a fast waveform is selected; When the character density is greater than a preset character density, a fast waveform is selected; When the residual image risk index is greater than a preset residual image risk index, a high-quality waveform is selected; Conversely, a mixed waveform is selected.

15. The refresh system of claim 11, wherein, In the intelligent waveform optimization module, the waveform library includes: Fast waveforms, suitable for high frequency input, allowing for persistence, base parameter range: voltage gradient 2.5-4.5V / ms, refresh time 6-12ms; High quality waveforms, suitable for static display, ignoring persistence, base parameter range: voltage gradient 0.6-1.5V / ms, refresh time 40-80ms; Mixed waveforms, suitable for medium frequency input, balancing persistence and delay, base parameter range: voltage gradient 1.5-3.0V / ms, refresh time 15-30ms.

Images