Flexible AMOLED image distortion compensation low-power-consumption compression driving method based on deep learning

Abstract

The invention discloses a flexible AMOLED image distortion compensation low-power-consumption compression driving method based on deep learning, and the method comprises the steps: obtaining to-be-displayed original image data, and dividing the to-be-displayed original image data into a plurality of image blocks; constructing a convolutional neural network model of a dynamic feature screening module, and dynamically identifying and screening high-frequency features and brightness features in the image blocks; performing feature extraction on each image block by using a convolutional neural network model; carrying out compressed encoding on the extracted features; performing reconstruction prediction on the compressed data by using the generative adversarial network; and the driving circuit receives the adjusted compression characteristic data, and realizes image display of the flexible AMOLED screen through decoding and a low-power-consumption driving algorithm based on pixel-level brightness mapping. The problems of image distortion control and power consumption regulation and control in the bending state of the flexible screen are solved, and the image display quality and the energy efficiency ratio of the flexible AMOLED are improved.

Description

Technical Field

[0001] The present invention relates to the field of flexible display image technology, and in particular to a low-power compression driving method for flexible AMOLED image distortion compensation based on deep learning. Background Art

[0002] With the rapid development of flexible display technology, flexible AMOLED screens have been widely used in smart phones, wearable devices, car displays and other fields due to their self-luminous, thin and bendable, and high contrast characteristics. However, the flexible substrate characteristics of flexible AMOLED make it easy to produce image distortion when bent, and the power consumption of self-luminous pixels is closely related to the brightness and luminous duration. At the same time, mobile devices have increasingly stringent requirements for real-time and low-power display of screens. Traditional image compression driving methods are difficult to take into account the deformation characteristics of flexible screens, the coordinated optimization of human visual perception and power consumption control. How to achieve low-power, real-time compression driving while ensuring image quality has become a key issue that needs to be urgently addressed in flexible AMOLED display technology.

[0003] Traditional image compression driving methods are mainly based on fixed blocking strategies and manually designed feature extraction algorithms. In the image blocking stage, uniform grid division is usually adopted, ignoring the spatial differences in pixel deformation when the flexible screen is bent, resulting in large differences in pixel characteristics within the block and low compression efficiency. In terms of feature extraction, traditional methods rely on fixed transformations such as discrete cosine transform (DCT) or wavelet transform, which makes it difficult to dynamically capture high-frequency features that are sensitive to the human eye and brightness-sensitive areas of the flexible screen. In the driving stage, traditional pulse width modulation (PWM) technology uses fixed current intensity and luminous duration, and does not consider the changes in pixel luminous efficiency of the flexible screen under different bending states, resulting in wasteful power consumption or insufficient brightness uniformity.

[0004] Furthermore, traditional solutions lack the ability to perceive and adapt to the dynamic characteristics of flexible screens. Without a mathematical model for screen curvature and pixel deformation, they are unable to quantify the impact of deformation on image features. Feature selection and encoding are not optimized based on human visual characteristics, resulting in loss of high-frequency detail or brightness distortion in compressed images. Furthermore, their high computational complexity makes real-time execution difficult on the low-power chips of mobile devices. These limitations result in traditional technologies performing poorly in complex display scenarios for flexible AMOLEDs (such as high curvature and rapid brightness changes), making them unable to meet the dual requirements of modern smart devices for both display quality and energy efficiency.

[0005] Existing technologies have proposed some improvement solutions for the compression drive problem of flexible displays. For example, some studies have introduced machine learning algorithms to optimize image segmentation and grouped pixels by brightness or color through clustering algorithms. However, the deformation characteristics caused by the bending of the flexible screen are not taken into account, and the segmentation results cannot adapt to changes in the screen curvature. In terms of feature compression, image compression methods based on deep learning (such as variational autoencoders (VAEs)) can learn the hierarchical features of images, but the model has a large number of parameters and is difficult to deploy in embedded systems. In addition, the brightness perception loss function is not designed for the self-luminous characteristics of flexible screens, and the brightness accuracy of the reconstructed image is insufficient.

[0006] In terms of optimizing driving power consumption, existing technologies achieve energy saving by dynamically adjusting pixel refresh rates or current intensity, but they rely solely on simple brightness threshold judgments and fail to establish a dynamic mapping relationship between pixel luminous efficiency and bending state and temperature, making it impossible to finely control the power consumption of each pixel. In addition, existing solutions lack closed-loop optimization of the entire compression drive process, and compression encoding parameters cannot be adjusted in real time according to the screen state, resulting in large fluctuations in display quality under different degrees of bending. Although existing technologies have improved compression efficiency or reduced power consumption to a certain extent, they have failed to systematically integrate the physical characteristics of flexible screens, human visual perception models, and lightweight neural network architectures, making it difficult to achieve an optimal balance between image quality, computing power consumption, and real-time performance. Summary of the Invention

[0007] Based on the above technical problems, this application discloses a low-power compression driving method for flexible AMOLED image distortion compensation based on deep learning, including:

[0008] Acquire original image data to be displayed, and divide the original image data into a plurality of image blocks;

[0009] Constructing a convolutional neural network model for a dynamic feature screening module. This module, based on the self-luminous properties of flexible AMOLED pixels, uses an adaptive threshold adjustment algorithm to dynamically identify and screen high-frequency features in image blocks that have a significant impact on human visual perception and brightness features that affect display brightness.

[0010] Using the convolutional neural network model to extract features from each image block, the convolutional neural network model uses a lightweight convolution structure that combines grouped convolution and depthwise separable convolution for feature extraction;

[0011] The extracted features are compressed and encoded. During the encoding process, an image block deformation compensation model is established based on the bendability of the flexible AMOLED, and the features of the image distortion area generated in the bent state are optimized and encoded.

[0012] Generate a prediction image by reconstructing the compressed data using a generative adversarial network. The predicted image is then compared with the original image data, and the compression coding parameters are dynamically adjusted based on the comparison results.

[0013] The driving circuit receives the adjusted compressed feature data and realizes image display on the flexible AMOLED screen through decoding and a low-power driving algorithm based on pixel-level brightness mapping. The low-power driving algorithm dynamically adjusts the pixel light emission duration and current intensity according to the brightness characteristics.

[0014] Preferably, when the original image data is divided into multiple image blocks, the division is performed according to the bending curvature distribution of the flexible AMOLED screen and the self-luminous characteristics of the pixel array. By constructing a pixel deformation model under the bent state of the screen and combining the self-luminous response curve of the pixel, the pixel cluster is used as the basic unit, and pixels with similar bending deformation trends and luminous characteristics are divided into the same image block, wherein the pixel cluster is composed of adjacent pixels whose deformation difference when bending is less than a preset threshold and whose self-luminous brightness adjustment curve similarity is greater than the set threshold.

[0015] Preferably, the pixel deformation model in the curved state of the screen is constructed as follows:

[0016] Assuming the initial plane of the flexible AMOLED screen in the Cartesian coordinate system is z = 0, when the screen is bent along the curvature radius R, the three-dimensional coordinate transformation formula of any pixel point P (x, y, 0) is:

[0017]

[0018] Where R is the dynamic curvature radius along the short side of the screen, and the formula is: E is determined by the Young's modulus, Poisson's ratio v, and external stress σ of the flexible substrate material. The rate of change of the spatial distance between a pixel point P and its adjacent pixel point Q (x+Δx, y+Δy, 0) after bending is calculated using the formula: Where Q' is the converted adjacent pixel point. When ΔS(P, Q) < ε, pixels P and Q are divided into the same pixel cluster, where ε is the material deformation tolerance coefficient.

[0019] Preferably, the acquisition of high-frequency features that have a greater impact on human visual perception in the dynamic feature screening module is specifically as follows: constructing a visual perception importance function, quantifying the degree of influence of feature frequency on human visual perception, and acquiring high-frequency features that are more critical to human visual perception in the identified image. The formula is: Where Φ(f) is the contrast sensitivity function of the human eye, f is the characteristic frequency, f0 is the visual sensitivity center frequency, σ s is the bandwidth parameter; multi-scale wavelet decomposition is performed on the image block to obtain the high-frequency sub-band coefficient W h, calculate the visual importance score of each coefficient, the formula is: Among them L i，j is the brightness value of the corresponding position, α e is the brightness influencing factor.

[0020] Preferably, the brightness feature affecting the display brightness is obtained in the dynamic feature screening module, specifically:

[0021] Construct the brightness sensitivity function, the formula is: Where L' is the CIELAB brightness component, L max is the maximum brightness value; calculate the brightness feature importance score, the formula is: in is the brightness change rate, is the brightness Laplace operator, β s is the edge enhancement factor.

[0022] Preferably, the convolution process adopts a lightweight convolution structure combining grouped convolution and depth-wise separable convolution, specifically:

[0023] Input feature map Divide into G groups according to the channel dimension, each group of feature maps H, W, and C are the height, width, and number of channels of the feature map, respectively. g Perform depth-wise separable convolution, the formula is: Y g =DWConv(X g , K g )*PWConv(X g , W g ), where DWConv is a depthwise convolution that extracts spatial features from each group, and PWConv is a point-by-point convolution that performs channel mixing and generates the output feature map Y through weighted aggregation. where α g is the dynamic grouping coefficient, and Shuffle is the channel shuffling operation.

[0024] Preferably, the image block deformation compensation model is established according to the bendability characteristics of the flexible AMOLED, specifically:

[0025] Get each image block and calculate the feature compensation factor in the image block. The formula is: Among them G i is the deformation gradient field of the screen in the curved state, G0 is the gradient field of the plane state, λ is the regularization parameter, M i is the attention mask of the brightness feature; during the encoding process, the feature tensor T iPerform dynamic compensation and obtain the compensated feature tensor T' i :The formula is: T' i =C i ⊙T i +(1-C i )⊙Shift(T i , d i ), where d i is the displacement vector field of pixel deformation, Shift is the feature offset operation; by obtaining the compensated feature tensor T' i Generate a compensated compressed code stream.

[0026] Preferably, the image block deformation compensation model is used to optimize and encode the image distortion region features generated in the bending state, specifically:

[0027] Calculate the distortion sensitivity coefficient of the image block in the curved state. The formula is: in is the deformation gradient of pixel j, σ r is the sensitivity parameter, is the brightness change rate, γ s is the brightness influencing factor, and an adaptive quantization step size is constructed in the feature quantization stage. The formula is: Q j =Q0·(1+α t β j ), where Q0 is the reference step size, α t is the adjustment coefficient, and residual prediction coding is used for distortion-sensitive areas. The formula is: R j =F j -Pred(F j-1 , D j ), where F j is the current feature, Pred is the prediction function of the neighborhood feature deformation vector, D j is the displacement vector of pixel point j; the features of the image distortion area generated in the bending state are obtained to generate an optimized code stream.

[0028] Preferably, the reconstructing and predicting the compressed data by generating an adversarial network is specifically as follows:

[0029] Build generator G a and the discriminator D a The adversarial network composed of the generator G a Mapping the compressed feature data to the predicted image space, the discriminator D aDistinguish the authenticity of the predicted image from the original image; during the training process, the generator learns to minimize the perceptual loss function between the predicted image and the original image. The loss function combines pixel-level loss, VGG feature space loss, and brightness perception loss of flexible AMOLED display characteristics; the discriminator gives high weight to the distortion-sensitive areas of the image block when the screen is bent by introducing a spatial attention mechanism; based on the output distribution of the discriminator during adversarial training, the quantization parameters, feature retention threshold, and entropy coding probability model in the compression coding are dynamically adjusted to make the compression parameters adaptive to the current image content and screen bending state.

[0030] Preferably, the low-power driving algorithm is used to dynamically adjust the pixel light emission duration and current intensity according to the brightness characteristics, specifically:

[0031] By establishing a pixel luminous characteristic database, the pixel luminous efficiency curves of flexible AMOLED under different bending states and temperature conditions are stored. The decoded feature data and the screened brightness features are used to divide the screen into multiple brightness control areas, and the average brightness value of each area is obtained. For each pixel, the average brightness value of the area where it is located, the luminous efficiency curve corresponding to the bending state and temperature in the database, and the distortion sensitivity coefficient β are calculated. j , calculate the pixel lighting duration, the formula is: Among them L t is the target brightness, η is the pixel luminous efficiency function, R is the screen curvature, T is the temperature, is the maximum driving current; according to the calculated luminous duration t on , combined with the grayscale value of the pixel, through pulse width modulation technology, the current intensity I is dynamically adjusted s To drive the pixel to emit light, the formula is: Where T c is the driving cycle.

[0032] Compared with the prior art, the technical solution of this application has the following technical effects:

[0033] The present invention constructs a pixel deformation model under the bent state of the screen and combines it with the self-luminous response curve to divide pixels with similar deformation trends and luminous characteristics into the same image block, thereby solving the problem of insufficient adaptability of traditional uniform blocking to the characteristics of flexible screens.

[0034] The present invention designs a convolutional neural network that includes a dynamic feature screening module. It uses the visual perception importance function and the brightness sensitivity function to accurately extract high-frequency features and key brightness features that the human eye is sensitive to, and reduces the amount of computation through grouped convolution and depthwise separable convolution to meet real-time requirements.

[0035] This invention establishes an image block deformation compensation model, performs residual predictive coding and adaptive quantization on the features of curved and distorted areas, and utilizes a generative adversarial network to construct a closed-loop optimization system for compression parameters, dynamically matching compression quality with screen status. Based on a database of pixel luminescence characteristics, this invention dynamically adjusts pixel luminescence duration and current intensity. Combining distortion sensitivity coefficients with brightness region division, this method achieves refined pixel-level power consumption control, significantly reducing energy consumption while maintaining display quality.

[0036] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application so that it can be implemented in accordance with the contents of the specification, and to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the following is a detailed description of the preferred embodiment of the present application in conjunction with the accompanying drawings.

[0037] Based on the detailed description of the specific embodiments of the present application in conjunction with the accompanying drawings below, those skilled in the art will become more aware of the above and other objects, advantages and features of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following will briefly introduce the drawings required for the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings without inventive work. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn according to the actual scale.

[0039] Figure 1 This is a flow chart of the low-power compression driving method for flexible AMOLED image distortion compensation based on deep learning of the present invention;

[0040] Figure 2 This is a diagram of a lightweight convolution structure combining grouped convolution and depthwise separable convolution of the present invention;

[0041] Figure 3 This is a comparison chart of the distortion rate between the distortion compensation low-power compression driving method of the present invention and the prior art;

[0042] Figure 4 This is a power consumption comparison diagram of the distortion compensation low-power compression driving method of the present invention and the prior art;

[0043] Figure 5 This is a comparison chart of power consumption and time duration between the distortion compensation low-power compression driving method of the present invention and the prior art. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all of the embodiments. In the following description, specific details such as specific configurations and components are provided only to help fully understand the embodiments of the present application. Therefore, it should be clear to those skilled in the art that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present application. In addition, for clarity and brevity, the description of known functions and structures has been omitted in the embodiments.

[0045] It should be understood that references throughout this specification to "one embodiment" or "this embodiment" mean that a particular feature, structure, or characteristic associated with the embodiment is included in at least one embodiment of the present application. Therefore, the appearance of "one embodiment" or "this embodiment" throughout this specification does not necessarily refer to the same embodiment. Furthermore, these particular features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0046] In addition, the present application may repeat reference numerals and / or letters in different examples. This repetition is for the purpose of simplicity and clarity and does not in itself indicate the relationship between the various embodiments and / or settings discussed.

[0047] The term "and / or" in this article is only a description of the association relationship of associated objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist at the same time. The term " / and" in this article describes another type of association object relationship, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the previous and subsequent associated objects are in an "or" relationship.

[0048] The term "at least one" in this article is merely a description of the association relationship between associated objects, indicating that three relationships may exist. For example, at least one of A and B can mean: A exists alone, A and B exist at the same time, and B exists alone.

[0049] It should also be noted that, in this document, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "include," "comprises," or any other variations thereof are intended to cover non-exclusive inclusion.

[0050] Example 1

[0051] This embodiment mainly describes a low-power compression driving method for flexible AMOLED image distortion compensation based on deep learning, such as Figure 1 As shown, specifically including:

[0052] S1. Obtaining original image data to be displayed, and dividing the original image data into multiple image blocks;

[0053] S2. Constructing a convolutional neural network model for a dynamic feature screening module. The dynamic feature screening module, based on the self-luminous characteristics of flexible AMOLED pixels, uses an adaptive threshold adjustment algorithm to dynamically identify and screen out high-frequency features in image blocks that have a significant impact on human visual perception and brightness features that affect display brightness.

[0054] S3. Extract features from each image block using the convolutional neural network model, where the convolutional neural network model uses a lightweight convolution structure that combines grouped convolution and depthwise separable convolution for feature extraction;

[0055] S4. compressing and encoding the extracted features. During the encoding process, an image block deformation compensation model is established based on the bendability of the flexible AMOLED, and the features of the image distortion area generated in the bent state are optimized and encoded;

[0056] S5. Reconstruct and predict the compressed data using a generative adversarial network to generate a predicted image, compare the predicted image with the original image data, and dynamically adjust the compression coding parameters based on the comparison results;

[0057] S6. The driving circuit receives the adjusted compressed feature data and realizes image display on the flexible AMOLED screen through decoding and a low-power driving algorithm based on pixel-level brightness mapping. The low-power driving algorithm dynamically adjusts the pixel light emission duration and current intensity according to the brightness feature.

[0058] Furthermore, when dividing the original image data into multiple image blocks, the division is performed according to the bending curvature distribution of the flexible AMOLED screen and the self-luminous characteristics of the pixel array. By constructing a pixel deformation model under the bent state of the screen and combining the self-luminous response curve of the pixel, the pixel cluster is used as the basic unit, and pixels with similar bending deformation trends and luminous characteristics are divided into the same image block, wherein the pixel cluster is composed of adjacent pixels whose deformation difference when bending is less than a preset threshold and whose self-luminous brightness adjustment curve similarity is greater than the set threshold.

[0059] Furthermore, the pixel deformation model in the curved state of the screen is constructed as follows:

[0060] Assuming the initial plane of the flexible AMOLED screen in the Cartesian coordinate system is z = 0, when the screen is bent along the curvature radius R, the three-dimensional coordinate transformation formula of any pixel point P (x, y, 0) is:

[0061]

[0062] Where R is the dynamic curvature radius along the short side of the screen, and the formula is: E is determined by the Young's modulus, Poisson's ratio v, and external stress σ of the flexible substrate material. The rate of change of the spatial distance between a pixel point P and its adjacent pixel point Q (x+Δx, y+Δy, 0) after bending is calculated using the formula: Where Q' is the converted adjacent pixel point. When ΔS(P, Q) < ε, pixels P and Q are divided into the same pixel cluster, where ε is the material deformation tolerance coefficient.

[0063] Furthermore, the acquisition of high-frequency features that have a greater impact on human visual perception in the dynamic feature screening module is specifically as follows: constructing a visual perception importance function, quantifying the degree of influence of feature frequency on human visual perception, and obtaining high-frequency features that are more critical to human perception in the recognition image. The formula is: Where Φ(f) is the contrast sensitivity function of the human eye, f is the characteristic frequency, f0 is the visual sensitivity center frequency, σ s is the bandwidth parameter; multi-scale wavelet decomposition is performed on the image block to obtain the high-frequency sub-band coefficient W h , calculate the visual importance score of each coefficient, the formula is: Among them L i，j is the brightness value of the corresponding position, α e is the brightness influencing factor.

[0064] Furthermore, the brightness feature that affects the display brightness is obtained in the dynamic feature screening module, specifically:

[0065] Construct the brightness sensitivity function, the formula is: Where L' is the CIELAB brightness component, L max is the maximum brightness value; calculate the brightness feature importance score, the formula is: in is the brightness change rate, is the brightness Laplace operator, β s is the edge enhancement factor.

[0066] Further, if Figure 2 As shown in the figure, the convolution process adopts a lightweight convolution structure that combines grouped convolution and depth-separable convolution, specifically:

[0067] Input feature map Divide into G groups according to the channel dimension, each group of feature maps H, W, and C are the height, width, and number of channels of the feature map, respectively. g Perform depth-wise separable convolution, the formula is: Y g =DWConv(X g , K g )*PWConv(X g , W g ), where DWConv is a depthwise convolution that extracts spatial features from each group, and PWConv is a point-by-point convolution that performs channel mixing and generates the output feature map Y through weighted aggregation. where α g is the dynamic grouping coefficient, and Shuffle is the channel shuffling operation.

[0068] Furthermore, according to the bendability of the flexible AMOLED, an image block deformation compensation model is established, specifically:

[0069] Get each image block and calculate the feature compensation factor in the image block. The formula is: Among them G i is the deformation gradient field of the screen in the curved state, G0 is the gradient field of the plane state, λ is the regularization parameter, M i is the attention mask of the brightness feature; during the encoding process, the feature tensor Ti is dynamically compensated to obtain the compensated feature tensor T' i :The formula is: T' i =C i ⊙T i +(1-C i )⊙Shift(T i , d i ), where d i is the displacement vector field of pixel deformation, Shift is the feature offset operation; by obtaining the compensated feature tensor T' i Generate a compensated compressed code stream.

[0070] Furthermore, the image block deformation compensation model is used to optimize and encode the image distortion region features generated in the bending state, specifically:

[0071] Calculate the distortion sensitivity coefficient of the image block in the curved state. The formula is: in is the deformation gradient of pixel j, σ r is the sensitivity parameter, is the brightness change rate, γ s is the brightness influencing factor, and an adaptive quantization step size is constructed in the feature quantization stage. The formula is: Q j=Q0·(1+α t β j ), where Q0 is the reference step size, α t is the adjustment coefficient, and residual prediction coding is used for distortion-sensitive areas. The formula is: R j =F j -Pred(F j-1 , D j ), where F j is the current feature, Pred is the prediction function of the neighborhood feature deformation vector, D j is the displacement vector of pixel point j; the features of the image distortion area generated in the bending state are obtained to generate an optimized code stream.

[0072] Furthermore, the reconstructed prediction of the compressed data by generating an adversarial network is specifically as follows:

[0073] Build generator G a and the discriminator D a The adversarial network composed of the generator G a Mapping the compressed feature data to the predicted image space, the discriminator D a Distinguish the authenticity of the predicted image from the original image; during the training process, the generator learns to minimize the perceptual loss function between the predicted image and the original image. The loss function combines pixel-level loss, VGG feature space loss, and brightness perception loss of flexible AMOLED display characteristics; the discriminator gives high weight to the distortion-sensitive areas of the image block when the screen is bent by introducing a spatial attention mechanism; based on the output distribution of the discriminator during adversarial training, the quantization parameters, feature retention threshold, and entropy coding probability model in the compression coding are dynamically adjusted to make the compression parameters adaptive to the current image content and screen bending state.

[0074] Furthermore, the low-power driving algorithm dynamically adjusts the pixel light emission duration and current intensity according to the brightness characteristics, specifically:

[0075] By establishing a pixel luminous characteristic database, the pixel luminous efficiency curves of flexible AMOLED under different bending states and temperature conditions are stored. The decoded feature data and the screened brightness features are used to divide the screen into multiple brightness control areas, and the average brightness value of each area is obtained. For each pixel, the average brightness value of the area where it is located, the luminous efficiency curve corresponding to the bending state and temperature in the database, and the distortion sensitivity coefficient β are calculated. j , calculate the pixel lighting duration, the formula is: Among them L t is the target brightness, η is the pixel luminous efficiency function, R is the screen curvature, T is the temperature, is the maximum driving current; according to the calculated luminous duration ton , combined with the grayscale value of the pixel, through pulse width modulation technology, the current intensity I is dynamically adjusted s To drive the pixel to emit light, the formula is: Where T c is the driving cycle.

[0076] This embodiment describes in detail the flexible AMOLED image compression driving method proposed in this application. Through an adaptive blocking strategy combined with the screen curvature characteristics and pixel luminescence patterns, feature extraction is made more suitable for flexible display scenarios. A dynamic feature screening mechanism accurately identifies key visual information. A lightweight convolutional structure effectively reduces the computational load. An image distortion compensation model specifically optimizes feature encoding in the curved state, significantly improving image reconstruction accuracy. A generative adversarial network closes the loop to adjust compression parameters to ensure consistent display quality under different curvatures. A pixel-level low-power driving algorithm dynamically controls power consumption based on brightness characteristics and luminous efficiency, significantly reducing energy consumption while ensuring display effects. This provides an efficient solution for the application of flexible AMOLED in high-dynamic, low-power scenarios.

[0077] Based on Example 1, this embodiment describes in detail how a generative adversarial network reconstructs and predicts compressed data, specifically:

[0078] Generate adversarial networks to reconstruct and predict compressed data, build an adversarial learning framework including generators and discriminators, and realize intelligent reconstruction of compressed features and dynamic adjustment of encoding parameters through joint optimization of multi-dimensional loss functions. a Using the compressed feature data as input, the low-dimensional features are mapped to a predicted image space with the same size as the original image through multi-layer convolution and deconvolution operations. Its goal is to minimize the perceptual difference between the predicted image and the original image I. Discriminator D a As the "judge" of the image's authenticity, it extracts the image's spatial features and structural information through a convolutional neural network and outputs the probability value of the predicted image being a real image. During adversarial training, the generator continuously optimizes network parameters through gradient backpropagation to deceive the discriminator into misjudging the predicted image as the original image. The discriminator simultaneously updates its parameters to improve its ability to distinguish between real and fake images. The two achieve a dynamic balance through the minimax game mechanism.

[0079] The perceptual loss function of the reconstructed image integrates multi-dimensional evaluation indicators and is formulated as follows: Among them, pixel-level loss Minimize the difference in pixel values ​​between the predicted image and the original image, and solve the mosaic effect common in traditional compression algorithms; VGG feature space loss The pre-trained VGG network extracts high-level semantic features, constrains the texture structure and semantic consistency of the predicted image, and avoids the loss of high-frequency details; brightness perception loss logL(I i )|Based on the self-luminous characteristics of flexible AMOLED, a weighted penalty is applied to the brightness distortion in the curved state to ensure the brightness accuracy in high curvature areas; Through the adversarial game mechanism, the generator is forced to learn the underlying laws of data distribution and improve the visual authenticity of the reconstructed image; p ,λ f ,λ l ,λ g is the corresponding weight coefficient, The generator predicts the value of the image at pixel i, I i is the value of the original image at pixel i, N is the total number of pixels in the image, φ l (·) is the feature extraction function of the lth layer of the VGG network, L(·) is the conversion function, and D a (I) is the discriminator’s predicted probability of authenticity for the real image I, Generate images for the discriminator pair The true prediction probability, p data (I) is the data distribution of real images, The distribution of images generated by the generator;

[0080] Dynamic optimization is achieved through the discriminator's spatial attention mechanism and closed-loop feedback of the compression parameters. When calculating the probability of image authenticity, the discriminator introduces an attention weight matrix associated with the distortion sensitivity coefficient, assigning higher discrimination weights to high-distortion areas when the screen is bent (such as edge areas with a small curvature radius), forcing the generator to prioritize the reconstruction accuracy of these areas. By monitoring the discriminator's output distribution (such as the mean and variance of the probability difference between the real image and the predicted image), the quantization parameters and feature retention thresholds in the compression coding are dynamically adjusted. For example, when the discriminator's probability of authenticity of the predicted image is lower than the threshold, the quantization step size is reduced and the feature retention ratio is increased, thereby dynamically balancing the compression code stream and reconstruction quality, forming a closed-loop process of "compression-reconstruction-feedback-optimization", which enables the entire driving scheme to adapt to the real-time bending state and image content complexity of the flexible AMOLED.

[0081] This embodiment describes in detail the construction of a compression and reconstruction closed-loop system through a generative adversarial network, the use of a multi-dimensional loss function to improve the visual consistency and brightness accuracy of the predicted image and the original image, the combination of a spatial attention mechanism to enhance the reconstruction accuracy of distorted areas, and the dynamic optimization of compression parameters based on discriminator feedback, significantly improving the image reconstruction quality and compression efficiency of flexible AMOLEDs in different bending states.

[0082] Based on Example 1, this embodiment describes in detail the specific implementation technical effects of this application on the prior art, specifically:

[0083] In terms of image distortion control, the existing technologies mostly use compensation algorithms based on geometric transformation (such as bilinear interpolation and thin plate spline interpolation). However, when the flexible AMOLED screen is bent with a radius of curvature less than 50mm, the image edge distortion rate generally exceeds 18%, and the peak pixel displacement error is more than 12 pixels. This application uses the synergy of the pixel deformation model and the image block deformation compensation algorithm to reduce the edge distortion rate to 7.2% under the same curvature conditions (curvature radius of 30mm), and the pixel displacement error is controlled within 4 pixels, which is more than 50% higher than the compensation algorithm based on geometric transformation. At the same time, Figure 3 As shown in Figure 2, in 200 bending tests with different curvatures (dynamic variation of curvature radius from 20 to 100 mm), the average distortion rate of the present application solution was stable at 6.8% to 11.5%, while the average distortion rate of the prior art fluctuated between 15.3% and 22.7%, demonstrating stronger curvature adaptability and distortion control capabilities.

[0084] In terms of low power consumption, existing flexible AMOLED driving technologies usually adopt a fixed pulse width modulation (PWM) strategy, where the pixel light emission current intensity and duration are adjusted only according to the global brightness, such as Figure 4 As shown in the figure, in a typical display scenario (50% global brightness, medium curvature), the overall power consumption of the screen is about 450mW, and the pixel power consumption fluctuation coefficient in the high-brightness area (such as the white block) is 0.32 (indicating that the power consumption difference between pixels is significant). This application reduces the overall power consumption to 306mW under the same test conditions through pixel-level brightness mapping and luminous efficiency dynamic adjustment algorithm, which is a 32% reduction in power consumption. At the same time, the pixel power consumption fluctuation coefficient in the high-brightness area is compressed to 0.11, and the power consumption uniformity is improved by 65.6%. Figure 5 As shown, in a comparative experiment of continuously displaying dynamic images (such as 60fps video), the average power consumption of the solution of the present application is 289mW, while the existing technology is 421mW, the battery life is extended by about 45%, and there is no local pixel brightness attenuation caused by uneven power consumption, which verifies its significant advantages in low power consumption and display stability.

[0085] This embodiment describes in detail that the technology of the present application demonstrates excellent performance in the detection of termite holes in dams by virtue of multi-module collaboration and advanced algorithms. The detection efficiency, positioning accuracy, and missed detection rate far exceed those of existing technologies. The data processing is efficient, the anti-interference ability is strong, and it still maintains high stability in complex environments, which is comprehensively superior to traditional detection methods.

[0086] The above are only preferred embodiments of the present invention, which do not limit the scope of protection of the present invention. For those skilled in the art, the present invention can be modified and varied in various ways. Any changes, modifications, replacements, integrations and parameter changes to these embodiments through conventional substitutions or that can achieve the same functions without departing from the principles and spirit of the present invention fall within the scope of protection of the present invention.

Claims

1. A low-power compression driving method for flexible AMOLED image distortion compensation based on deep learning, characterized in that: The following steps are involved: Acquire original image data to be displayed, and divide the original image data into a plurality of image blocks; Constructing a convolutional neural network model for a dynamic feature screening module. This module, based on the self-luminous properties of flexible AMOLED pixels, uses an adaptive threshold adjustment algorithm to dynamically identify and screen high-frequency features in image blocks that have a significant impact on human visual perception and brightness features that affect display brightness. Using the convolutional neural network model to extract features from each image block, the convolutional neural network model uses a lightweight convolution structure that combines grouped convolution and depthwise separable convolution for feature extraction; The extracted features are compressed and encoded. During the encoding process, an image block deformation compensation model is established based on the bendability of the flexible AMOLED, and the features of the image distortion area generated in the bent state are optimized and encoded. Generate a prediction image by reconstructing the compressed data using a generative adversarial network. The predicted image is then compared with the original image data, and the compression coding parameters are dynamically adjusted based on the comparison results. The driving circuit receives the adjusted compressed feature data and realizes image display on the flexible AMOLED screen through decoding and a low-power driving algorithm based on pixel-level brightness mapping. The low-power driving algorithm dynamically adjusts the pixel light emission duration and current intensity according to the brightness characteristics.

2. The method for low-power compression driving of flexible AMOLED image distortion compensation based on deep learning according to claim 1, characterized in that: When dividing the original image data into multiple image blocks, the division is performed based on the bending curvature distribution of the flexible AMOLED screen and the self-luminous characteristics of the pixel array. By constructing a pixel deformation model when the screen is bent and combining the self-luminous response curve of the pixels, pixels with similar bending deformation trends and luminous characteristics are divided into the same image block using pixel clusters as basic units. The pixel clusters are composed of adjacent pixels whose deformation difference when bent is less than a preset threshold and whose self-luminous brightness adjustment curve similarity is greater than a set threshold.

3. The method for low-power compression driving of flexible AMOLED image distortion compensation based on deep learning according to claim 2, characterized in that: The pixel deformation model under the curved screen state is constructed as follows: Assuming the initial plane of the flexible AMOLED screen in the Cartesian coordinate system is z = 0, when the screen is bent along the curvature radius R, the three-dimensional coordinate transformation formula of any pixel point P (x, y, 0) is: Where R is the dynamic curvature radius along the short side of the screen, and the formula is: E is determined by the Young's modulus, Poisson's ratio ν, and external stress σ of the flexible substrate material. The rate of change of the spatial distance between the pixel point P and its adjacent pixel point q (x+Δx, y+Δy, 0) after bending is calculated using the formula: Where Q' is the converted adjacent pixel point. When ΔS(P,Q)<ε, pixels P and Q are divided into the same pixel cluster, where ε is the material deformation tolerance coefficient.

4. The method for low-power compression driving of flexible AMOLED image distortion compensation based on deep learning according to claim 1, characterized in that: The acquisition of high-frequency features that have a greater impact on human visual perception in the dynamic feature screening module is specifically as follows: constructing a visual perception importance function, quantifying the degree of influence of feature frequency on human visual perception, and obtaining high-frequency features that are more critical to human visual perception in the recognition image. The formula is: Where Φ(f) is the contrast sensitivity function of the human eye, f is the characteristic frequency, f0 is the visual sensitivity center frequency, σ s is the bandwidth parameter; multi-scale wavelet decomposition is performed on the image block to obtain the high-frequency sub-band coefficient W h , calculate the visual importance score of each coefficient, the formula is: Among them L i,j is the brightness value of the corresponding position, α e is the brightness influencing factor.

5. The method for low-power compression driving of flexible AMOLED image distortion compensation based on deep learning according to claim 1 or 4, characterized in that: The brightness feature acquisition that affects the display brightness in the dynamic feature screening module is specifically as follows: Construct the brightness sensitivity function, the formula is: Where L' is the CIELAB brightness component, L max is the maximum brightness value; Calculate the brightness feature importance score using the formula: in is the brightness change rate, is the brightness Laplace operator, β s is the edge enhancement factor.

6. The method for low-power compression driving of flexible AMOLED image distortion compensation based on deep learning according to claim 1, characterized in that: The convolution process adopts a lightweight convolution structure that combines grouped convolution and depth-wise separable convolution, specifically: Input feature map Divide into G groups according to the channel dimension, each group of feature maps H, W, and C are the height, width, and number of channels of the feature map, respectively. g Perform depth-wise separable convolution, the formula is: Y g =DWConv(X g ,K g )*PWConv(X g ,W g ), where DWConv is a depthwise convolution that extracts spatial features from each group, and PWConv is a point-by-point convolution that performs channel mixing and generates the output feature map Y through weighted aggregation. where α g is the dynamic grouping coefficient, and Shuffle is the channel shuffling operation.

7. The method for low-power compression driving of flexible AMOLED image distortion compensation based on deep learning according to claim 1, characterized in that: The image block deformation compensation model is established based on the bendability of the flexible AMOLED, specifically: Get each image block and calculate the feature compensation factor in the image block. The formula is: Among them G i is the deformation gradient field of the screen in the curved state, G0 is the gradient field of the plane state, λ is the regularization parameter, M i is the attention mask of the brightness feature; during the encoding process, the feature tensor Ti is dynamically compensated to obtain the compensated feature tensor T' i :The formula is: T' i =C i ⊙T i +(1-C i )⊙Shift(T i ,d i ), where d i is the displacement vector field of pixel deformation, Shift is the feature offset operation; by obtaining the compensated feature tensor T' i Generate a compensated compressed code stream.

8. The method for low-power compression driving of flexible AMOLED image distortion compensation based on deep learning according to claim 1 or 7, characterized in that: The image block deformation compensation model is used to optimize and encode the image distortion region features generated in the bending state, specifically: Calculate the distortion sensitivity coefficient of the image block in the curved state. The formula is: in is the deformation gradient of pixel j, σ r is the sensitivity parameter, is the brightness change rate, γ s is the brightness influencing factor, and an adaptive quantization step size is constructed in the feature quantization stage. The formula is: Q j =Q0·(1+α t β j ), where Q0 is the reference step size, α t is the adjustment coefficient, and residual prediction coding is used for distortion-sensitive areas. The formula is: R j =F j -Pred(F j-1 ,D j ), where F j is the current feature, Pred is the prediction function of the neighborhood feature deformation vector, D j is the displacement vector of pixel point j; the features of the image distortion area generated in the bending state are obtained to generate an optimized code stream.

9. The method for low-power compression driving of flexible AMOLED image distortion compensation based on deep learning according to claim 1, characterized in that: The reconstruction prediction of compressed data by generating an adversarial network is specifically as follows: Build generator G a and the discriminator D a The adversarial network composed of the generator G a Mapping the compressed feature data to the predicted image space, the discriminator D a Distinguish the authenticity of the predicted image from the original image; during the training process, the generator learns to minimize the perceptual loss function between the predicted image and the original image. The loss function combines pixel-level loss, VGG feature space loss, and brightness perception loss of flexible AMOLED display characteristics; the discriminator gives high weight to the distortion-sensitive areas of the image block when the screen is bent by introducing a spatial attention mechanism; based on the output distribution of the discriminator during adversarial training, the quantization parameters, feature retention threshold, and entropy coding probability model in the compression coding are dynamically adjusted to make the compression parameters adaptive to the current image content and screen bending state.

10. The method for low-power compression driving of flexible AMOLED image distortion compensation based on deep learning according to claim 1, characterized in that: The low-power driving algorithm dynamically adjusts the pixel light emission duration and current intensity according to the brightness characteristics, specifically: By establishing a pixel luminous characteristic database, the pixel luminous efficiency curves of flexible AMOLED under different bending states and temperature conditions are stored. The decoded feature data and the screened brightness features are used to divide the screen into multiple brightness control areas, and the average brightness value of each area is obtained. For each pixel, the average brightness value of the area where it is located, the luminous efficiency curve corresponding to the bending state and temperature in the database, and the distortion sensitivity coefficient β are calculated. j , calculate the pixel lighting duration, the formula is: Among them L t is the target brightness, η is the pixel luminous efficiency function, R is the screen curvature, T is the temperature, is the maximum driving current; according to the calculated luminous duration t on , combined with the grayscale value of the pixel, through pulse width modulation technology, the current intensity I is dynamically adjusted s To drive the pixel to emit light, the formula is: Where T c is the driving cycle.

Images