Ink screen art painting generation method based on generative adversarial network

Abstract

The application relates to the technical field of image generation, in particular to an ink screen art painting generation method based on a generative adversarial network, which comprises the following steps: acquiring input instructions, color gamut parameters and a critical voltage area of a target ink screen; when a preference label exceeds the color gamut parameters, an initial latent variable is constructed; the initial latent variable is input into a fixed-weight generation network to output an initial feature map; the critical voltage area is converted into a penalty area, and a gray tensor of the initial feature map is extracted; when the gray tensor falls into the penalty area, a penalty error is calculated, and iterative updating is performed on the initial latent variable until the global gray tensor is separated from the penalty area, and a target feature map is output; when the gray tensor does not fall into the penalty area, the initial feature map is output as the target feature map; and a driving waveform is called to control the target ink screen to perform rendering. The application can solve the physical shock and residual image stacking caused by the critical voltage area interval of the ink screen, and realize that the gray scale distribution of the generated image avoids the voltage area interval.

Description

Technical Field

[0001] This application relates to the field of image generation technology, and in particular to a method for generating e-ink art paintings based on generative adversarial networks. Background Technology

[0002] Intelligent voice-interactive e-ink screens receive external commands through multimodal terminals and deliver the generated two-dimensional image data to the underlying hardware for display, often functioning as digital picture frames or display control terminals. The physical display of such devices relies on the displacement of microcapsule electrophoretic particles under the influence of a driving electric field. When handling generative art rendering tasks driven by multimodal commands, the underlying system must establish a mapping path between the grayscale features of the front-end image and the physical driving level of the back-end. Since artwork objectively contains complex and continuous grayscale gradient changes, the data stream must ensure that the grayscale parameters allocated at the front end match the analog voltage applied to the driving line before being sent to the display panel, in order to maintain the steady state of the physical flipping of the microcapsule pixel array.

[0003] Chinese invention patent CN121171180B discloses an image display system and method for an intelligent voice-interactive e-ink screen. This solution parses the semantic structure of descriptive information at the image generation end, combines the grayscale parameters of the e-ink screen to generate corresponding grayscale icons and text dot matrices, and packages the pixel data to generate target image data based on display refresh parameters. At the physical display end, the system relies on real-time collected ambient temperature to query the driving voltage value from a preset parameter table, converts it into an analog voltage and applies it to the driving line, and simultaneously uses a mechanism of collecting current feedback signals and comparing them with expected values ​​to dynamically adjust the driving voltage.

[0004] The aforementioned solution employs a serial architecture where the front end abstracts away underlying hardware constraints to generate conventional grayscale features, while the back end relies on temperature and current feedback for passive voltage adjustment. When the system operates under complex painting generation conditions involving large-area continuous grayscale gradients, the microcapsule electrophoretic particles in the e-ink screen encounter a critical unsteady voltage range during specific intermediate grayscale flipping, where stable polarity cannot be solidified. Limited by the fragmented data processing logic of this existing technology, pixel data with continuous gradient attributes is directly linearly mapped and sent out, causing a large number of pixel node driving instructions to inevitably fall within this critical blind zone of the electrophoretic particles. The microcapsule particles, forced by the unsteady electric field, cannot solidify their physical polarity, resulting in high-frequency ineffective oscillations at the screen's physical rendering level, accompanied by edge spikes and residual charge accumulation. Summary of the Invention

[0005] To address the physical oscillations and image stacking caused by the critical voltage range of e-ink screens, and to ensure that the grayscale distribution of the generated image avoids this voltage range, this application provides an e-ink screen art painting generation method based on generative adversarial networks.

[0006] The method for generating ink-screen art based on generative adversarial networks provided in this application adopts the following technical solution: The method for generating ink-screen art based on generative adversarial networks includes:

[0007] The input command is obtained, and the input command is decoupled into a semantic vector and a preference label. The color gamut parameters and critical voltage zone of the target e-ink screen are also obtained.

[0008] An initial latent variable is constructed based on the semantic vector and the preference label. When the preference label is projected onto the mapped chromaticity coordinates generated by the standard color space and exceeds the physical rendering boundary corresponding to the color gamut parameter, a texture vector is extracted based on a preset mapping, and the texture vector is fused with the semantic vector to construct the initial latent variable.

[0009] The initial latent variables are input into a generator network with fixed weights, and the initial feature map is output.

[0010] The critical voltage region is transformed into a penalty region, and the grayscale tensor of the initial feature map is extracted.

[0011] When the grayscale tensor falls into the penalty region, the penalty error is calculated, and the penalty error is used to perform iterative updates only on the initial latent variable until the global grayscale tensor leaves the penalty region, and the target feature map is output; when the grayscale tensor does not fall into the penalty region, the initial feature map is output as the target feature map.

[0012] Based on the target feature map, the driving waveform is invoked to control the target e-ink screen to perform rendering.

[0013] Optionally, the step of obtaining the input instruction, decoupling the input instruction into a semantic vector and preference labels, and obtaining the color gamut parameters and critical voltage region of the target e-ink screen includes:

[0014] Receive hardware signaling from the target e-ink screen, and parse the hardware signaling to obtain the color gamut parameters and the critical voltage zone;

[0015] Extract the multimodal sequence from the input instruction, perform orthogonal separation on the multimodal sequence, and decouple the semantic vector representing the visual structure and the preference label representing the emotional attribute.

[0016] Optionally, the step of extracting texture vectors based on a preset mapping includes:

[0017] Read the preset compensation dictionary, which contains multiple discrete structural latent vectors;

[0018] The preference label is used as an index to address in the compensation dictionary and extract the matching structural latent vector;

[0019] The structural latent vector is output as the texture vector, and the portion of the mapped chromaticity coordinates that exceeds the physical rendering boundary is replaced by the spatial fundamental frequency distribution of the texture vector.

[0020] Optionally, the step of converting the critical voltage region into a penalty region includes:

[0021] Analyze the critical voltage region and extract the lower and upper threshold values;

[0022] Obtain the scaling factor, and calculate the bilateral difference values ​​of the lower threshold and the upper threshold based on the scaling factor;

[0023] The penalty region is generated by smoothly scaling the difference between the two-sided difference values ​​to generate a gradient that can be backpropagated.

[0024] Optionally, before the step of calculating the bilateral difference values ​​of the lower threshold and the upper threshold based on the scaling factor, the method further includes:

[0025] Obtain the ambient temperature of the target e-ink screen;

[0026] The drift bias is calculated based on the preset viscosity equation and the ambient temperature.

[0027] The drift bias is used to compensate for the lower threshold and the upper threshold to calibrate the calculation benchmark of the bilateral difference value.

[0028] Optionally, the step of performing iterative updates on the initial latent variables only using the penalized error further includes:

[0029] Count the number of times the iterative update is executed;

[0030] When the number of executions reaches the deadlock threshold, the grayscale tensor has not escaped the penalty region, and the initial latent variable contains the texture vector, a relaxation bias is generated based on the spatial fundamental frequency distribution of the texture vector.

[0031] The relaxation bias is added to the initial latent variable, and the iterative update continues.

[0032] Optionally, the step of inputting the initial latent variables into a generator network with fixed weights and outputting an initial feature map further includes:

[0033] Receive a loop command input for the target e-ink screen;

[0034] In response to the loop instruction, the driving feedback current and charge attenuation coefficient of the previous rendering frame are obtained, and the charge retention matrix generated on the target e-ink screen is calculated based on the driving feedback current and the charge attenuation coefficient.

[0035] The charge retention matrix is ​​inverted to generate a negative mask;

[0036] The negative mask is injected into the decoding layer built into the generator network. Based on the negative mask constraining the spatial topology, the initial feature map is controlled to avoid the residual region represented by the charge retention matrix.

[0037] Optionally, the step of invoking the driving waveform based on the target feature map to control the target e-ink screen to perform rendering includes:

[0038] The target feature map is decoded into a target grayscale matrix corresponding to the microcapsule pixel array of the target e-ink screen;

[0039] Extract the extreme value distribution of each pixel node in the target grayscale matrix;

[0040] The driving waveform is generated by matching the steady-state level according to the extreme value distribution;

[0041] The driving waveform is applied to the target e-ink screen to drive the pixel nodes that avoid the residual area to physically flip.

[0042] Optionally, the steps of extracting the grayscale tensor of the initial feature map and calculating the penalty error when the grayscale tensor falls into the penalty region include:

[0043] The absolute grayscale value of each pixel node is extracted from the initial feature map to generate the grayscale tensor;

[0044] The grayscale tensor is input into a discriminator with a physical penalty term;

[0045] When the discriminator determines that the grayscale tensor falls into the penalty region, it calls the physical penalty term to calculate the bias and outputs it as the penalty error.

[0046] Optionally, the step of using the penalty error to iteratively update only the initial latent variables until the global grayscale tensor escapes the penalty region and outputs the target feature map includes:

[0047] Freeze the weights of the generator network and establish a gradient backpropagation path pointing to the initial latent variables;

[0048] The penalty error is propagated backward along the gradient backpropagation path to calculate the error gradient;

[0049] The initial latent variables are updated using the error gradient, and the updated initial latent variables are input into the generator network to regenerate the initial feature map.

[0050] The grayscale tensor of the initial feature map is re-extracted until it is determined that none of the global grayscale tensors fall into the penalty region, at which point the current initial feature map is output as the target feature map.

[0051] In summary, this application includes the following beneficial technical effects:

[0052] 1. By transforming the critical voltage region of the e-ink screen into a penalty region that can backpropagate gradients, and extracting the gray-level tensor from the initial feature map of the generated network output, when the gray-level tensor falls into the penalty region, the penalty error is calculated and the initial latent variable is iteratively updated until the global gray-level tensor leaves the penalty region, so that the gray-level distribution of the generated image avoids the non-steady voltage range in which electrophoretic particles cannot be solidified, and avoids high-frequency invalid oscillations and the stacking of residual charge.

[0053] 2. When the preference label exceeds the color gamut parameters of the target e-ink screen, the texture vector is extracted based on the preset mapping and fused with the semantic vector to construct the initial latent variable. The spatial fundamental frequency distribution of the texture vector is used to replace the out-of-bounds part in the preference label, so that the color emotion description that exceeds the range of the hardware color gamut can still be effectively mapped to the texture structure that the screen can display, thus expanding the freedom of expression of input commands.

[0054] 3. Responding to the loop instruction, extract the persistence matrix generated by the previous rendering frame on the target e-ink screen, perform inversion processing on the persistence matrix to generate a negative mask and inject it into the decoding layer of the generator network, constrain the spatial topology of the initial feature map to avoid the residual area, so that the rendering of subsequent frames during the loop display process is not affected by the previous charge residue, thus improving the image quality consistency of multi-frame continuous display. Attached Figure Description

[0055] Figure 1 A flowchart of an ink screen art painting generation method based on generative adversarial networks provided in an embodiment of this application;

[0056] Figure 2 An architecture diagram of an e-ink screen art generation system based on generative adversarial networks provided in an embodiment of this application;

[0057] Figure 3 This is a schematic diagram illustrating the construction principle of the penalty region provided in an embodiment of this application. Detailed Implementation

[0058] The following combination Figures 1-3 This application will be described in further detail.

[0059] This application discloses a method for generating e-ink art based on generative adversarial networks. This method relies on the interaction between a multimodal interactive terminal, an edge generation controller, and an e-ink physical display control terminal. The edge generation controller embeds a cross-modal feature decoupler, a compensatory dictionary, a generator, and a discriminator. The e-ink physical display control terminal includes an EPD driver IC, a microcapsule pixel array, and a bonded thermal component. Figure 1 As shown, the ink screen art generation method based on generative adversarial networks is controlled in terms of execution timing by a multi-conditional branching loop consisting of preference label out-of-bounds state, grayscale tensor penalty state, and deadlock threshold.

[0060] The multimodal interactive terminal captures the user-input description of the painting's theme, style, and color mood, encapsulates this description into a request frame, with the frame header identifying the instruction type, the frame body carrying the multimodal sequence, and the frame tail including a cyclic redundancy check (CRC) code. The multimodal interactive terminal sends the request frame to the edge generation controller via the SPI bus. The edge generation controller verifies the CRC code; if successful, it initiates a metadata request signal to the e-ink physical display control terminal. The EPD driver IC of the e-ink physical display control terminal returns color gamut parameters and critical voltage range in its status register, while the bonded thermistor returns a real-time ambient temperature scalar. The color gamut parameters characterize the upper and lower boundaries of the grayscale that the panel can stably display, and the critical voltage range characterizes the unstable voltage band where microcapsule particles cannot solidify. For a certain type of bistable electrophoretic panel, the measured values ​​in this range fall within two unstable voltage bands: [-3.5V, -1.2V] and [1.2V, 3.5V]. Due to the symmetrical drive polarity, after mapping the absolute value to a 256-level grayscale tensor, the corresponding grayscale values ​​are in the range of 85 to 170.

[0061] If no response is received within a preset 20-millisecond time limit for the metadata request signaling, the edge generation controller retransmits the metadata request signaling. If no response is received after three retransmissions, the cached metadata is read to maintain the process, avoiding handshake blocking that could cause the entire generation chain to stall. The edge generation controller calls the cross-modal feature decoupler to perform orthogonal separation on the multimodal sequence, decoupling a 512-dimensional semantic vector and a 64-dimensional preference label. The semantic vector carries visual structural information, and the preference label carries color emotion information. As a preferred implementation, the cross-modal feature decoupler performs layer normalization on the multimodal sequence before orthogonal separation. Layer normalization scales each modal component to the same numerical scale, suppressing the amplitude differences between the speech modality and the text modality from infiltrating the decoupling result. The SPI bus can be equivalently replaced by the I2C bus.

[0062] After receiving a 512-dimensional semantic vector and a 64-dimensional preference label, the edge generation controller projects the preference label onto the CIELAB standard color space to generate mapped chromaticity coordinates. Then, it compares the color components of these mapped chromaticity coordinates with the physical rendering boundary values ​​corresponding to the color gamut parameters dimension by dimension. When each dimension of the mapped chromaticity coordinates of the preference label falls within the physical rendering boundary, the edge generation controller uses the semantic vector to construct initial latent variables. When any dimension of the mapped chromaticity coordinates of the preference label exceeds the physical rendering boundary, the edge generation controller uses the preference label as an index to address the data in a compensation dictionary. The compensation dictionary stores multiple discrete structural latent vectors, each bound to a structural texture pattern. These structural texture patterns include cross shadows, halftone dots, and random dotting. After a successful address match, the matching structural latent vector is extracted as a texture vector. The spatial fundamental frequency distribution of the texture vector replaces the out-of-bounds color components in the preference label. The texture vector and the semantic vector are then concatenated and fused along the channel dimension to form the initial latent variables.

[0063] The generator, with its initial latent variables input and weights frozen, employs a style-generating network structure that supports latent variable injection. Its feedforward output carries an initial feature map with continuous gray-level gradients. As a preferred implementation, when an addressing fails to find a match in the compensation dictionary, the edge generation controller selects the structural latent vector with the closest Euclidean distance to the out-of-bounds color component as the texture vector, thus preventing addressing failures from causing fusion interruptions.

[0064] After receiving the initial feature map, the discriminator analyzes the critical voltage range and extracts the lower and upper threshold values. Before performing the mapping determination, the discriminator obtains the real-time ambient temperature scalar reported by the bonded thermistor and calculates the drift bias based on a preset viscosity equation. The specific calculation relationship is as follows:

[0065]

[0066] in, For drift bias, For real-time ambient temperature scalar, Standard reference temperature and The viscosity coefficient of the electrophoretic solution is pre-calibrated. The calculation benchmark of the bilateral difference value is calibrated by using the drift bias compensation lower and upper threshold values. This viscosity equation characterizes the physical relationship that the migration rate of electrophoretic particles in the suspension medium increases with increasing temperature and the critical voltage decreases with increasing temperature.

[0067] The discriminator obtains the scaling factor and constructs Sigmoid functions for the compensated lower and upper thresholds based on the scaling factor, calculating the bilateral difference values. Let the absolute grayscale bias of the pixel node be... The lower threshold after compensation is The upper limit threshold is The scaling factor is Then the lower limit bilateral difference value Bilateral difference with upper limit They are respectively:

[0068]

[0069]

[0070] Combination Figure 3 It can be seen that in the mapped coordinate system with grayscale and voltage values ​​on the horizontal axis and difference values ​​on the vertical axis, the bilateral difference curve of the lower threshold exhibits an exponential approximation earlier than the bilateral difference curve of the upper threshold. This smooths and scales the difference between the two sides, thereby generating a continuously differentiable shadow region between the two curves whose projection width completely covers the critical voltage region. This, in turn, generates a penalty region function that can backpropagate gradients. :

[0071]

[0072] The penalty region is calculated as the difference between the lower and upper limits of the Sigmoid function. The scaling factor ranges from 5 to 50. If the scaling factor is too small, the penalty region edges are too smooth and the gradient discrimination is insufficient. If the scaling factor is too large, the penalty region edges approach a step, and the gradient backpropagation fluctuates violently. A scaling factor of 20 balances edge steepness and backpropagation smoothness. The discriminator extracts the absolute grayscale bias value of each pixel node relative to the zero-level reference on the initial feature map and directly outputs its spatial distribution tensor as a grayscale tensor. As a preferred implementation, the adhesive thermal component can be equivalently replaced by a temperature sampling register embedded inside the EPD driver IC.

[0073] The discriminator inputs the grayscale tensor into the physical penalty term, which outputs a bias based on the relative position of the grayscale tensor to the penalty region. A non-zero physical bias penalty error is output when any grayscale tensor falls into the penalty region. Before iteration, the edge generation controller marks the parameters of each generator layer as untrainable, retaining only the gradient differentiability of the initial latent variable nodes. A gradient backpropagation path pointing to the initial latent variables is established, and the physical bias penalty error is propagated back along this path to calculate the error gradient. This error gradient is used to update only the initial latent variables, which are then re-inputted into the generator to regenerate the initial feature map. The discriminator re-extracts the grayscale tensor to compare with the penalty region. This iterative loop re-enters until the global grayscale tensor escapes the penalty region, at which point the current initial feature map is output as the target feature map. If the grayscale tensor never falls into the penalty region, the initial feature map is directly output as the target feature map.

[0074] The edge generation controller counts the number of iterations. When the number of iterations reaches the deadlock threshold, the grayscale tensor still hasn't escaped the penalty region, and the initial latent variables contain texture vectors, the edge generation controller generates a relaxation bias based on the spatial fundamental frequency distribution of the texture vectors. This relaxation bias is then superimposed on the initial latent variables to break the saddle point balance and continue iteration. The deadlock threshold ranges from 50 to 100 iterations. A value that is too low will falsely trigger the relaxation bias before normal convergence, while a value that is too high will prolong the stagnation time. The deadlock threshold is set at 80 iterations. As a preferred implementation, the amplitude of the relaxation bias increases with the accumulation of deadlock iterations. The increasing amplitude falls back to the initial amplitude after breaking the saddle point, preventing a constant amplitude bias from failing to escape deep saddle points.

[0075] The edge generation controller sends the converged target feature map to the physical display control terminal of the e-ink screen via the SPI bus. The EPD driver IC decodes the target feature map into a target grayscale matrix corresponding to the microcapsule pixel array, extracts the extreme value distribution of each pixel node in the target grayscale matrix, and generates a driving waveform based on the stable level matched according to the extreme value distribution. The stable level is taken from the high-level and low-level regions outside the critical voltage range. The driving waveform is applied to the microcapsule pixel array to drive the pixel nodes to physically flip and complete the rendering. After the flip is completed, the EPD driver IC collects the current feedback of the row and column driving lines. When the current feedback amplitude exceeds the expected band, a compensation pulse is fed back.

[0076] When the multimodal interactive terminal triggers a loop display, the edge generation controller receives the loop instruction, obtains the driving feedback current and charge decay coefficient of the previous rendered frame, and calculates the charge retention matrix generated on the e-ink screen based on the driving feedback current and charge decay coefficient. It then performs inversion processing on the charge retention matrix to generate a negative mask, injects the negative mask into the topology of the decoding layer built into the generator, and controls the initial feature map of subsequent frames to avoid the residual area represented by the charge retention matrix. In a preferred implementation, the charge retention matrix is ​​calculated, updated, and cached in real time by the EPD driver IC after each frame rendering, based on the current feedback backflow amount and the exponential decay function over time. Each subsequent frame of the loop display reads the latest charge retention matrix to generate a negative mask.

[0077] like Figure 2 As shown in the embodiments, this application also discloses an e-ink screen art generation system based on generative adversarial networks, comprising a multimodal interactive terminal, an edge generation controller, and an e-ink screen physical display control terminal. The multimodal interactive terminal is coupled to the edge generation controller via an SPI bus, and transmits request frames through this bus; the edge generation controller is coupled to the e-ink screen physical display control terminal via an SPI bus to interact with hardware signaling. The edge generation controller integrates four core logic processing modules: a cross-modal feature decoupling unit, a compensatory dictionary, a generator, and a discriminator.

[0078] The multimodal interactive terminal is equipped with a multimodal parsing component and an interactive interface. The multimodal parsing component is configured to capture the user's input of the painting theme, style and color mood description and encapsulate it into a request frame carrying a cyclic redundancy check code. The interactive interface is configured to carry the trigger entry for saving, sharing and cyclic display.

[0079] The edge generation controller embeds a cross-modal feature decoupler, a compensation dictionary, a generator, and a discriminator. In the overall architecture of the generative adversarial network in this application, the generator is used for feedforward inference to output a feature map carrying continuous gray-level gradients; while the discriminator is only called during the optimization iteration phase, calculating a penalty error through a physical penalty term to constrain the update space of the generator network's latent variables. The cross-modal feature decoupler is configured to perform orthogonal separation on the multimodal sequences in the request frame, outputting semantic vectors representing visual structure and preference labels representing emotional attributes, and projecting the preference labels onto a standard color space to generate mapped chromaticity coordinates. The compensation dictionary is configured to store discrete structural latent vectors of multiple bound structural texture patterns, and is addressed by the preference label as an index when the mapped chromaticity coordinates exceed the physical rendering boundary corresponding to the color gamut parameters. The generator weights are in a frozen state, the input is coupled to the initial latent variable node, and it is configured to feedforward based on the initial latent variables to output an initial feature map carrying continuous gray-level gradients. The generator's built-in decoding layer is configured to receive a negative mask to constrain the spatial topology. The discriminator is equipped with a physical penalty term and is configured to map the critical voltage range into a penalty region through a two-sided Sigmoid differential, extract the absolute grayscale bias value of each pixel node relative to the zero-level reference on the initial feature map, directly output its spatial distribution tensor as a grayscale tensor, and output the physical bias penalty error when the grayscale tensor falls into the penalty region.

[0080] The edge generation controller is further configured to freeze the generator weights and establish a gradient backpropagation path pointing to the initial latent variable node. The physical bias penalty error is backpropagated along this path to update the initial latent variable. When the number of iterations reaches the deadlock threshold, the grayscale tensor has not left the penalty region, and the initial latent variable contains a texture vector, a relaxation bias is generated and superimposed based on the spatial fundamental frequency distribution of the texture vector. When receiving a loop instruction for the target e-ink screen input, the driving feedback current and charge attenuation coefficient of the previous rendering frame are obtained, and the charge retention matrix is ​​calculated based on the driving feedback current and charge attenuation coefficient to generate a negative mask.

[0081] The physical display control unit of the e-ink screen includes an EPD driver IC, a microcapsule pixel array, and a bonded thermistor. The EPD driver IC's status register is configured to return color gamut parameters and critical voltage ranges, and is configured to generate a driving waveform based on the extreme value distribution of each pixel node in the target feature map, matching the steady-state level and applying it to the microcapsule pixel array. The microcapsule pixel array is coupled to the row and column driving lines of the EPD driver IC and is configured to receive the driving waveform and perform physical flipping. The bonded thermistor is bonded to the microcapsule pixel array and is configured to acquire a real-time ambient temperature scalar and report it to the edge generation controller.

[0082] The edge generation controller is carried by an edge computing chip containing computing power units and a non-volatile storage medium coupled to the chip. The compensation dictionary resides in the non-volatile storage medium. The operations of the cross-modal feature decoupling unit, generator, and discriminator are executed by the computing power unit. The physical display control terminal of the e-ink screen can be equivalently replaced by a cholesteric liquid crystal display control terminal. The microscopic details of latent variable iteration, penalty region construction, and temperature calibration have been described in detail in the aforementioned method embodiments and will not be repeated here.

[0083] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for generating e-ink screen artworks based on generative adversarial networks, characterized in that, include: The input command is obtained, and the input command is decoupled into a semantic vector and a preference label. The color gamut parameters and critical voltage zone of the target e-ink screen are also obtained. An initial latent variable is constructed based on the semantic vector and the preference label. When the preference label is projected onto the mapped chromaticity coordinates generated by the standard color space and exceeds the physical rendering boundary corresponding to the color gamut parameter, a texture vector is extracted based on a preset mapping, and the texture vector is fused with the semantic vector to construct the initial latent variable. The initial latent variables are input into a generator network with fixed weights, and the initial feature map is output. The critical voltage region is transformed into a penalty region, and the grayscale tensor of the initial feature map is extracted. When the grayscale tensor falls into the penalty region, the penalty error is calculated, and the penalty error is used to perform iterative updates only on the initial latent variable until the global grayscale tensor leaves the penalty region, and the target feature map is output; when the grayscale tensor does not fall into the penalty region, the initial feature map is output as the target feature map. Based on the target feature map, the driving waveform is invoked to control the target e-ink screen to perform rendering.

2. The method for generating e-ink screen artwork based on generative adversarial networks according to claim 1, characterized in that, The steps of acquiring input commands, decoupling the input commands into semantic vectors and preference labels, and acquiring the color gamut parameters and critical voltage region of the target e-ink screen include: Receive hardware signaling from the target e-ink screen, and parse the hardware signaling to obtain the color gamut parameters and the critical voltage zone; Extract the multimodal sequence from the input instruction, perform orthogonal separation on the multimodal sequence, and decouple the semantic vector representing the visual structure and the preference label representing the emotional attribute.

3. The method for generating e-ink screen artwork based on generative adversarial networks according to claim 2, characterized in that, The step of extracting texture vectors based on a preset mapping includes: Read the preset compensation dictionary, which contains multiple discrete structural latent vectors; The preference label is used as an index to address in the compensation dictionary and extract the matching structural latent vector; The structural latent vector is output as the texture vector, and the portion of the mapped chromaticity coordinates that exceeds the physical rendering boundary is replaced by the spatial fundamental frequency distribution of the texture vector.

4. The method for generating e-ink screen artwork based on generative adversarial networks according to claim 1, characterized in that, The step of converting the critical voltage region into a penalty region includes: Analyze the critical voltage region and extract the lower and upper threshold values; Obtain the scaling factor, and calculate the bilateral difference values ​​of the lower threshold and the upper threshold based on the scaling factor; The penalty region is generated by smoothly scaling the difference between the two bilateral difference values ​​to generate a gradient that can be backpropagated.

5. The method for generating e-ink screen artwork based on generative adversarial networks according to claim 4, characterized in that, Before the step of calculating the bilateral difference values ​​of the lower threshold and the upper threshold based on the scaling factor, the method further includes: Obtain the ambient temperature of the target e-ink screen; The drift bias is calculated based on the preset viscosity equation and the ambient temperature. The drift bias is used to compensate for the lower threshold and the upper threshold to calibrate the calculation benchmark of the bilateral difference value.

6. The method for generating e-ink screen artwork based on generative adversarial networks according to claim 3, characterized in that, The step of using the penalty error to perform iterative updates only on the initial latent variables further includes: Count the number of times the iterative update is executed; When the number of executions reaches the deadlock threshold, the grayscale tensor has not escaped the penalty region, and the initial latent variable contains the texture vector, a relaxation bias is generated based on the spatial fundamental frequency distribution of the texture vector. The relaxation bias is added to the initial latent variable, and the iterative update continues.

7. The method for generating e-ink screen artwork based on generative adversarial networks according to claim 1, characterized in that, The step of inputting the initial latent variables into a generator network with fixed weights and outputting an initial feature map further includes: Receive a loop command input to the target e-ink screen; In response to the loop instruction, the driving feedback current and charge attenuation coefficient of the previous rendering frame are obtained, and the charge retention matrix generated on the target e-ink screen is calculated based on the driving feedback current and the charge attenuation coefficient. The charge retention matrix is ​​inverted to generate a negative mask; The negative mask is injected into the decoding layer built into the generator network. Based on the negative mask constraining the spatial topology, the initial feature map is controlled to avoid the residual region represented by the charge retention matrix.

8. The method for generating e-ink screen artwork based on generative adversarial networks according to claim 7, characterized in that, The step of controlling the target e-ink screen to perform rendering based on the target feature map by calling the driving waveform includes: The target feature map is decoded into a target grayscale matrix corresponding to the microcapsule pixel array of the target e-ink screen; Extract the extreme value distribution of each pixel node in the target grayscale matrix; The driving waveform is generated by matching the steady-state level according to the extreme value distribution; The driving waveform is applied to the target e-ink screen to drive the pixel nodes that avoid the residual area to physically flip.

9. The method for generating e-ink screen artwork based on generative adversarial networks according to claim 1, characterized in that, The steps of extracting the grayscale tensor of the initial feature map and calculating the penalty error when the grayscale tensor falls into the penalty region include: The absolute grayscale value of each pixel node is extracted from the initial feature map to generate the grayscale tensor; The grayscale tensor is input into a discriminator with a physical penalty term; When the discriminator determines that the grayscale tensor falls into the penalty region, it calls the physical penalty term to calculate the bias and outputs it as the penalty error.

10. The method for generating e-ink screen artwork based on generative adversarial networks according to claim 9, characterized in that, The step of using the penalty error to iteratively update only the initial latent variables until the global grayscale tensor escapes the penalty region and outputs the target feature map includes: Freeze the weights of the generator network and establish a gradient backpropagation path pointing to the initial latent variables; The penalty error is propagated backward along the gradient backpropagation path to calculate the error gradient; The initial latent variables are updated using the error gradient, and the updated initial latent variables are input into the generator network to regenerate the initial feature map. The grayscale tensor of the initial feature map is extracted again until it is determined that none of the global grayscale tensors fall into the penalty region. Then, the current initial feature map is output as the target feature map.

Images