Image processing method and system for liquid crystal panel display distortion correction

Abstract

This invention discloses an image processing method and system for correcting display distortion in liquid crystal panels, relating to the field of display distortion correction technology. The method includes the following steps: inputting a standard test image into the liquid crystal panel and simultaneously acquiring its actual physical display image; performing image registration and brightness response normalization processing on the standard test image and the actual physical display image to form image pair training data; constructing a display distortion correction model based on a U-Net network and training the model using the image pair training data to obtain a trained model; inputting the original input frame to be displayed into the trained model, the model outputting a pixel compensation matrix, and adding this matrix to the pixel value matrix of the original input frame to generate a liquid crystal panel correction signal, thereby achieving display distortion correction. By constructing a complete closed loop from image pair acquisition and model training to real-time correction, adaptive and pixel-level real-time correction of complex display distortions in liquid crystal panels is achieved.

Description

Technical Field

[0001] This invention relates to the field of display distortion correction engineering, specifically to an image processing method and system for correcting display distortion in a liquid crystal panel. Background Technology

[0002] In LCD panel manufacturing, process variations in areas such as backlight modules, liquid crystal molecule alignment, and thin-film transistor characteristics can lead to physical distortions in screen display, resulting in uneven brightness and color. This inherent defect directly causes visible cloudiness, bright / dark areas, or color shifts in the displayed image, severely impacting display uniformity and image quality. Therefore, correcting display distortion in LCD panels is a crucial step in improving the visual experience of end-user devices.

[0003] Currently, a typical calibration method in the industry is the compensation value calculation method based on multi-point sampling and spatial interpolation. This method first inputs a few standard test images of gray levels into the panel and uses a camera to capture the display screen. The algorithm selects sparsely distributed pixels on the captured images as sampling points, and constructs a simple brightness response model for each point based on the brightness measurement values ​​of these sampling points. Based on the brightness response model data of these sparse sampling points, the required brightness compensation value for all pixels on the screen at each gray level is calculated, and a static compensation lookup table is generated for application to the driving circuit.

[0004] However, the traditional methods based on sparse sampling and spatial interpolation rely heavily on the accuracy of the luminance response model constructed with a limited number of sampling points in representing the spatial distribution of complex distortion. They cannot effectively capture and reconstruct the mixed distortion features caused by process fluctuations that propagate in space and contain both low-frequency gradual and high-frequency abrupt components. This results in inaccurate distortion estimation for pixel areas that are not directly sampled, ultimately limiting the correction effect of this method on complex spatial frequency mixed distortion, and residual flaws may still exist in the image. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an image processing method and system for correcting display distortion in liquid crystal panels, thus solving the problems in the background technology.

[0006] To achieve the above objectives, the present invention provides an image processing method and system for correcting distortion in a liquid crystal panel display, comprising the following steps: Step S1: Input the standard test image into the LCD panel and simultaneously acquire the actual physical image of the LCD panel display; Step S2: Perform image registration and brightness response normalization on the standard test image and the actual display physical image to form image pair training data; Step S3: Construct a display distortion correction model based on the U-Net network, and train the display distortion correction model using the image training data to obtain a trained display distortion correction model; Step S4: Input the original input frame to be displayed into the trained display distortion correction model, output the pixel compensation matrix, add the pixel value matrix of the original input frame to the pixel compensation matrix to generate the liquid crystal panel correction signal, and realize the display distortion correction of the liquid crystal panel.

[0007] Preferably, inputting a standard test image to the LCD panel and simultaneously acquiring the actual physical display image of the LCD panel includes: Set up a production line testing environment, which includes an image signal generator and a high-precision industrial camera. Connect the image signal generator to the LCD panel under test and drive it to sequentially input a set of preset standard test images. Simultaneously, a precisely positioned high-precision industrial camera is controlled to synchronously acquire images of the LCD panel's display area, obtaining images consistent with each standard test image. Strictly corresponding actual display physical image .

[0008] Preferably, the standard test image and the actual display physical image are image registered and their brightness response normalized to form image pair training data, which includes: Image registration was performed using the ORB feature matching algorithm, which is effective on standard test images. Compared with the actual displayed physical image Keypoints are detected and descriptors are calculated for each image, thereby establishing the correspondence between feature points in the two images. Based on a set of correct matching point pairs, a 3×3 homography matrix describing the projection transformation between the two planar images is calculated. Using this matrix to Perform geometric transformations and resample using bilinear interpolation to obtain the result. Registered physical image with precise alignment at the pixel level. ; Next, brightness response normalization is performed to correct the nonlinearity of the photoelectric response system combining the high-precision industrial camera and the LCD panel, by registering the physical image. The pixel values ​​are mapped to the Consistent linear brightness space; By sequentially performing the image registration and brightness response normalization processes described above, the original image pairs... Transformed into high-quality image pairs that can be directly used for learning. The set of all such processed image pairs, which constitutes the final image pair training data, is denoted as . ,in Represents the total number of image pairs, and i is the index of the image pair order.

[0009] Preferably, the display distortion correction model constructed based on the U-Net network includes: A display distortion correction model M is constructed, which adopts a U-Net network structure that incorporates the Laplacian pyramid. Its final output is the target pixel compensation matrix. Normalized distorted image As input to the model, it is then constructed. Layered Laplace Pyramid ,in, Represents the finest scale; Represents the coarsest scale; intermediate layer The number of pyramid layers corresponds to different intermediate scales. This is a preset hyperparameter; Secondly, each level of the Pyramid of Laplace Each is fed into a shared-weight encoder subnetwork. Feature extraction is performed during this process, which can be described as follows:

[0010] in, Indicates from the pyramid number 1 Feature maps extracted from the layers; The learnable parameters representing the encoder subnetwork, and the shared weight design improves the parameter efficiency of the model; By concatenating all scale-aligned feature maps through channel stitching, a composite feature map is formed that integrates global contextual information and local detail information. ; Finally, the fusion features Input a decoder network decoder network Composed of multiple transposed convolutional layers and convolutional layers, it gradually restores the spatial resolution of the feature map through upsampling operations. The decoder ends with a layer that uses a linear activation function. A convolutional layer outputs a normalized, distorted image of the input. Pixel compensation matrices of identical size .

[0011] Preferably, the calculation process of the pixel compensation matrix includes: The formula for calculating the prime compensation matrix is ​​as follows:

[0012] in, This is the pixel compensation matrix predicted by the model; The learnable parameters of the decoder network, the pixel compensation matrix The middle is located in the coordinate Color channels element value , representing the input image The amount of compensation to be applied to the corresponding position and channel; a positive value indicates an increase in brightness or color intensity, while a negative value indicates a decrease.

[0013] Preferably, the display distortion correction model is trained using the image as training data to obtain a trained display distortion correction model, including: For image pairs training data Each sample Calculate its true compensation matrix label The calculation method is the difference between the corresponding pixel values ​​of the ideal standard image and the normalized distorted image; Determine a loss function L to quantify the compensation matrix predicted by the model. With the true compensation matrix The difference between them is minimized using an adaptive moment estimation optimizer to minimize the loss function. During training, the training data is analyzed from images. Randomly select small batches of data Input the model and calculate the predicted values ​​through forward propagation. The loss is then calculated relative to the model parameters using the backpropagation algorithm. gradient And update the parameters according to the following rules:

[0014] in, The learning rate controls the step size of each parameter update. Training terminates when the preset number of iterations is reached, or when the loss value on the independent validation dataset no longer decreases significantly. The model parameters that perform best on the validation set at this point are then saved. The complete computational graph corresponding to this set of parameters constitutes the trained display distortion correction model. .

[0015] Preferably, the determination of the loss function includes: The mean absolute error is used as the loss function, which imposes a uniform linear penalty on the prediction error:

[0016] in, and These represent the height and width of the image, respectively. and These represent the model prediction and the actual compensation matrices in coordinates, respectively. Color channels The value on; the summation symbol indicates that the absolute value errors of all spatial locations and color channels of the image are summed and averaged.

[0017] Preferably, the original input frame to be displayed is input into the trained display distortion correction model, and the output pixel compensation matrix includes: During the operation of the display device, its image processing front end receives the raw input frame to be displayed, denoted as... Then, the original input frame Input the trained display distortion correction model In the middle, the trained display distortion correction model Perform forward propagation inference on the original input frame and output a value that is identical to the original input frame. Pixel compensation matrices of identical size.

[0018] Preferably, the pixel value matrix and pixel compensation matrix of the original input frame are added together to generate a liquid crystal panel correction signal, thereby achieving liquid crystal panel display distortion correction, including: Obtain pixel compensation matrix Then, in the digital image processing domain, the original input frame is... Pixel value matrix and pixel compensation matrix Perform pixel-by-pixel, channel-by-channel addition, which generates an intermediate correction result matrix. ; For the intermediate correction result matrix A value range clipping operation is performed to limit all pixel values ​​to the valid range supported by the display device. Then, the signal is sent to the back-end processing unit of the display driver link, and converted into a correction drive signal that conforms to the electrical and timing interface specifications of the LCD panel via the timing controller and source driver standard hardware circuit.

[0019] An image processing system for correcting distortion in a liquid crystal panel display includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method.

[0020] This invention provides an image processing method for correcting distortion in liquid crystal panel displays, which involves machine learning and deep learning technologies, and has the following beneficial effects: (1) Through the unique encoder-decoder structure of the U-Net network combined with skip connections, it can simultaneously capture the global context information and local fine features of the image. When applied to the task of display distortion correction, this structure enables the model to effectively keep the details of image edges, textures and other features from being blurred when performing pixel-level compensation prediction, overcoming the shortcomings of traditional fully connected networks or simple convolutional networks in detail recovery.

[0021] (2) By incorporating the Laplacian pyramid framework into the U-Net network, the network is provided with inherent multi-scale analysis capabilities. This enables the display distortion correction model to distinguish and process distortion modes that exist at different scales caused by the manufacturing process of the liquid crystal panel. Through this multi-scale feature fusion, the display distortion correction model's ability to model complex distortions and its correction accuracy are further enhanced, especially in improving the correction effect on complex scenes where cross-scale distortions coexist.

[0022] (3) By directly learning complex nonlinear mapping relationships from a large number of standard-distortion image pairs, the model possesses strong adaptive and generalization capabilities, enabling it to automatically correct complex distortions in different types and batches of panels. Simultaneously, the trained model can complete full-image correction in a single forward propagation inference, meeting the high real-time processing requirements of display devices. Furthermore, this closed-loop framework has the potential for iterative optimization with new data, adapting to panel aging and process evolution, thereby fundamentally achieving intelligent, efficient, and precise long-term correction of LCD panel display distortions. Attached Figure Description

[0023] Figure 1 This is a flowchart of an image processing method for correcting display distortion in a liquid crystal panel, as proposed in this invention.

[0024] Figure 2 The image processing method for liquid crystal panel display distortion correction proposed in this invention obtains a hierarchical map of image pairs with training data.

[0025] Figure 3 This is a hierarchy diagram of the liquid crystal panel correction signal obtained in the image processing method for liquid crystal panel display distortion correction proposed in this invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Please see Figures 1-3This invention provides a technical solution: an image processing method for correcting distortion in liquid crystal panel displays. Specifically, it provides the following plant extraction method based on image processing; please refer to [link to relevant documentation]. Figure 1 The method includes the following steps: Step S1: Input the standard test image into the LCD panel and simultaneously acquire the actual physical image of the LCD panel.

[0028] First, a production line testing environment is set up, which includes an image signal generator and a high-precision industrial camera. The image signal generator is connected to the LCD panel under test, driving it to sequentially input a set of preset standard test images. This set of standard test images Specifically, it covers the following three categories: A full-screen uniform grayscale image, with grayscale values ​​covering the entire range from 0 to 255, including key grayscale levels such as 0, 64, 128, 192, and 255; a monochrome color image, containing at least three colors: pure red, pure green, and pure blue; and feature test images, such as checkerboard patterns or bitmaps, to provide feature basis for subsequent image registration.

[0029] Simultaneously, a precisely positioned high-precision industrial camera is controlled to synchronously acquire images of the LCD panel's display area under test conditions, thereby obtaining images consistent with each standard test image. Strictly corresponding actual display physical image : For darkroom environments, the ambient light intensity is strictly below 1 lux to eliminate interference from ambient light reflection; Regarding camera specifications, the sensor resolution of a high-precision industrial camera is no less than the physical resolution of the LCD panel being tested, the color bit depth is no less than 10 bits, and it is equipped with a low-distortion industrial lens. For environmental parameters, the test environment temperature was kept stable at 23°C±2°C, and the relative humidity was controlled at 50%±10%RH to ensure the stability of the LCD panel display.

[0030] Through the above synchronization operations, the final result is a standard test image that matches each input image. —Corresponding actual display physical image The resulting set of paired images These images form the original basis of the training data required for subsequent processing.

[0031] It should be noted that a full-screen uniform grayscale image refers to an image in which all pixels have completely consistent grayscale values, used to quantitatively evaluate the display uniformity of a panel under different brightness levels; a monochrome color image refers to an image in which only a single color channel is activated to its maximum value while other channels are set to zero, used to analyze color purity; a feature test image refers to an image containing clear geometric shapes and high-contrast edges, specifically providing feature points for image registration algorithms; ambient light intensity is a physical quantity that measures the intensity of ambient light; sensor resolution refers to the total number of pixels that a camera imaging device can capture; color bit depth refers to the number of brightness levels that each color channel of a camera can record; and a low-distortion industrial lens is an optical lens designed to minimize geometric distortion of an image.

[0032] The image generated in this step This is a fundamental prerequisite for the effective execution of subsequent image registration, data normalization, and model training. The clearly defined test image types, relevant key performance indicators, and environmental parameters collectively ensure the quality and consistency of the original data.

[0033] Step S2: Perform image registration and brightness response normalization on the standard test image and the actual display physical image to form image pair training data.

[0034] This step follows the pairwise image set obtained in step S1. Its core objective is to eliminate external errors introduced during the data acquisition process through two key preprocessing steps, thereby constructing a pure, standardized image that represents the display distortion characteristics of the LCD panel itself, providing direct input for subsequent model training.

[0035] First, image registration is performed. To eliminate geometric misalignment caused by differences in the relative pose of the camera and the control panel, the ORB feature matching algorithm is used. This algorithm is applied to standard test images. Compared with the actual displayed physical image Keypoints are detected and descriptors are calculated for each image, thus establishing the correspondence between feature points in the two images. Based on a set of correct matching point pairs, a 3×3 homography matrix describing the projection transformation between the two planar images is calculated. Subsequently, this matrix was used to... Perform geometric transformations and resample using bilinear interpolation to obtain the result. Registered physical image with precise alignment at the pixel level. The process is described as follows:

[0036] in, Represented by homography matrix A combined operation of projection transformation and bilinear resampling on an image; This represents the registered physical image. The purpose of this formula is to achieve this through geometric transformation. Eliminate image distortion and displacement caused by camera perspective to ensure the physical image after registration. Compared with standard test images Strict spatial alignment lays the foundation for subsequent pixel-level comparison and correction.

[0037] Next, brightness response normalization is performed to correct the nonlinearity of the photoelectric response system combining the high-precision industrial camera and the LCD panel, by registering the physical image. The pixel values ​​are mapped to the A consistent linear brightness space. Specifically, this is achieved using standard test images. The known series of standard gray values ​​(such as 0, 64, 128, 192, 255) and their values ​​in... To fit the measured average brightness value within the uniform region, a polynomial mapping function is fitted. For example, a cubic function might look like this:

[0038] in, express The original brightness input value for the middle pixel; The coefficients are those determined by fitting using the least squares method; This represents the normalized standard brightness output value. Use this function... Applied to the registered physical image For each pixel, a normalized distorted image can be obtained after geometric registration and brightness response normalization. The purpose of this formula is to use a function... By establishing a mapping from the raw brightness values ​​captured by the camera to standard linear brightness values, and removing the nonlinear response characteristics of the acquisition device itself, the distortion of the image can be normalized. The pixel value changes accurately reflect the distortion of the LCD panel, rather than the influence of the camera.

[0039] By sequentially performing the image registration and brightness response normalization processes described above, the original image pairs... Transformed into high-quality image pairs that can be directly used for learning. The set of all such processed image pairs, which constitutes the final image pair training data, is denoted as . ,in Represents the total number of image pairs, and i is the index of the image pair order.

[0040] The purpose and effect of this step is to systematically isolate extrinsic geometric distortions and nonlinear brightness responses, ensuring that the image accurately reflects the training data. Each pair of data precisely and consistently encapsulates the inherent display distortion characteristics of the LCD panel. This provides a crucial and highly reliable data foundation for subsequently training a model that can directly learn and effectively correct within the field of view. The entire processing flow is logically rigorous, constituting a key transformation link from raw data to training data.

[0041] Step S3: Construct a display distortion correction model based on the U-Net network, and train the display distortion correction model on the training data of the image to obtain the trained display distortion correction model.

[0042] This step follows from the image pairs generated in step S2 for training data. Its core purpose is to construct a system capable of resolving normalized distorted images. In the middle, it accurately predicted how to restore it to the standard test image. The required pixel compensation is achieved through a deep learning model, which is then trained to master this mapping relationship. The entire process consists of two main stages: model building and model training.

[0043] First, a display distortion correction model M is constructed. This model employs a U-Net network structure that incorporates a Laplacian pyramid, designed to explicitly handle distortion features at different spatial frequencies. Its final output is the target pixel compensation matrix. The construction process proceeds in the following order: The image output from step 2, which has been registered and normalized to remove distortion. As input to the model, it is then constructed. Layered Laplace Pyramid .in, Represents the finest scale (original resolution), mainly containing high-frequency detail information of the image; Representing the coarsest scale, it mainly contains low-frequency contour information of the image; the intermediate layer These correspond to different intermediate scales. Number of pyramid levels This is a preset hyperparameter. This decomposition process separates the signal of the input image into different frequency bands, facilitating subsequent multi-scale feature analysis by the network.

[0044] Secondly, each level of the Pyramid of Laplace Each is fed into a shared-weight encoder subnetwork. Feature extraction is performed within each encoder subnetwork. It consists of multiple convolutional layers, batch normalization layers, and activation functions stacked sequentially, gradually abstracting features at that scale through downsampling. This process can be described as follows:

[0045] in, Indicates from the pyramid number 1 Feature maps extracted from the layers; The learnable parameters of the encoder subnetwork are represented by shared weights, which improves the model's parameter efficiency. The purpose of this formula is to extract deep features from the input image representation at various scales, providing a foundation for subsequent fusion.

[0046] Next, feature maps extracted at different scales Perform fusion. First, combine the coarser particles ( Feature map The upsampling operation is used to adjust the feature map to the finest scale. The same spatial dimensions are used. Subsequently, a channel concatenation operation is performed to aggregate all scale-aligned feature maps, forming a composite feature map that integrates global contextual information and local detail information. .

[0047] Finally, the fusion features Input a decoder network decoder network It consists of multiple transposed convolutional layers and convolutional layers, gradually restoring the spatial resolution of the feature map through upsampling operations. The decoder ends with a linear activation function. A convolutional layer outputs a normalized, distorted image of the input. Pixel compensation matrices of identical size The formula for calculating the prime compensation matrix is ​​as follows:

[0048] in, To display the pixel compensation matrix predicted by the distortion correction model; The matrix represents the learnable parameters of the decoder network. The middle is located in the coordinate Color channels element value , representing the input image The formula specifies the amount of compensation to be applied to the corresponding location and channel. Positive values ​​indicate an increase in brightness or color intensity, while negative values ​​indicate a decrease. The purpose of this formula is to decode and map the fused multi-scale features into a final, pixel-level compensation output.

[0049] The purpose of this display distortion correction model construction process is to enable the network to simultaneously and effectively capture and process large-scale brightness unevenness (low-frequency distortion) and subtle pixel defects (high-frequency distortion) by introducing Laplacian pyramids for explicit multi-scale analysis, thereby predicting a more accurate spatial adaptive compensation matrix.

[0050] Use the images prepared in step 2 to train the data The constructed model is trained, and this stage aims to optimize the model parameters to improve the compensation matrix of its predictions. To approximate the ideal compensation value as closely as possible, the specific training process is as follows: First, for the image pairs training data Each sample Calculate its true compensation matrix label The calculation method is the difference between the corresponding pixel values ​​of the ideal standard image and the normalized distorted image:

[0051] in, That is, the true, ideal pixel compensation matrix, whose each element Indicates the location ,aisle The above should be Adjust to The exact amount of change required.

[0052] Secondly, a loss function L is determined to quantify the compensation matrix predicted by the model. With the true compensation matrix The difference between them. The mean absolute error is used as the loss function, which imposes a uniform linear penalty on the prediction error:

[0053] in, and These represent the height and width of the image, respectively. and These represent the model prediction and the actual compensation matrices in coordinates, respectively. Color channels The values ​​(usually corresponding to the R, G, and B channels) are used; the summation symbol indicates that the absolute errors of all spatial locations and color channels of the image are summed and averaged. The purpose of this loss function is to directly minimize the absolute deviation between the predicted compensation and the actual compensation, driving the model parameters to update in the correct direction.

[0054] Next, an adaptive moment estimation optimizer is used to minimize the loss function. During training, the training data is analyzed from images. Randomly select small batches of data Input the model. Calculate the predicted values ​​through forward propagation. The loss is then calculated relative to the model parameters using the backpropagation algorithm. gradient And update the parameters according to the following rules:

[0055] in, The learning rate controls the step size for each parameter update. This iterative process is repeated cyclically to adjust the model parameters. Continuous optimization.

[0056] Training terminates when the preset number of iterations is reached, or when the loss value on the independent validation dataset no longer decreases significantly. The optimal model parameters on the validation set at this point are saved. The complete computational graph corresponding to this set of parameters constitutes the trained display distortion correction model. This model is capable of handling distorted observation images with arbitrary input. Outputs a high-precision pixel compensation matrix. ,satisfy .

[0057] This step yields a core correction model capable of directly predicting precise pixel-level compensation amounts from distorted images. This provides a reliable algorithmic basis for achieving real-time online display distortion correction in step S4.

[0058] Step S4: Input the original input frame to be displayed into the trained display distortion correction model, output the pixel compensation matrix, add the pixel value matrix of the original input frame to the pixel compensation matrix to generate the liquid crystal panel correction signal, and realize the display distortion correction of the liquid crystal panel.

[0059] This step follows the training of the display distortion correction model obtained in step 3. Its core objective is to deploy this model within the image processing chain of an actual display device, enabling online, real-time distortion correction for any image content to be displayed. This process transforms the model's predicted output into actual adjustments to the driving signals, ultimately eliminating the inherent defects of the LCD panel at the physical display level. The entire process is executed in the following logical sequence.

[0060] First, model inference is performed to obtain the real-time pixel compensation matrix. During the operation of the display device, its image processing front end receives the raw input frame to be displayed, denoted as... ,Should This represents the raw image data that has not been corrected by this method. Then, the raw input frame... Input the trained display distortion correction model In the middle, the trained display distortion correction model Perform forward propagation inference on the original input frame and output a value that is identical to the original input frame. Pixel compensation matrices of identical size, denoted as :

[0061] in, This represents the original input frame, which is a frame with a width of Height is The number of channels is A digital image matrix (usually 3, representing RGB); This represents the display distortion correction model obtained from step 3. This represents the pixel compensation matrix predicted by the model, where each element... The value represents the value of At pixel position and color channels The amount of compensation to be applied is specified; positive values ​​indicate that brightness or color needs to be enhanced, while negative values ​​indicate that it needs to be reduced. The purpose of this formula is to use a trained model to infer the real-time input of each frame and quickly obtain a pixel-level accurate compensation amount covering the entire image.

[0062] Obtain pixel compensation matrix Then, in the digital image processing domain, the original input frame is... Pixel value matrix and pixel compensation matrix Perform pixel-by-pixel, channel-by-channel addition. This operation generates an intermediate correction result matrix. :

[0063] in, The original input frame is in position ,aisle Pixel values; In position ,aisle The corresponding compensation value for the pixel value; This is the calculated corrected digital pixel value. The purpose of this formula is to apply the predicted compensation amount to the original signal to generate a preliminary correction result.

[0064] After completing the addition operation, the intermediate correction result matrix needs to be adjusted. Perform a range cropping operation to restrict all pixel values ​​to a valid range supported by the display device (e.g., for an 8-bit depth image, restrict to a range of values). This is to prevent signal overflow (overexposure) or invalid values ​​(underexposure) caused by excessive compensation, and to ensure the physical validity of the data.

[0065] Finally, the final LCD panel calibration signal is generated and output. The intermediate calibration result matrix after domain cropping is shown. This matrix contains the final digital brightness and color information after pixel correction. It is then sent to the back-end processing unit of the display drive chain, where standard hardware circuits such as timing controllers and source drivers convert it into a correction drive signal that conforms to the electrical and timing interface specifications of the LCD panel. This signal incorporates compensation information calculated to counteract inherent panel distortion, ultimately driving the LCD panel to display a corrected image with significantly improved uniformity.

[0066] The purpose and effect of this step is to complete the closed loop from "offline model training" to "online real-time correction." This is achieved by using the trained display distortion correction model... By integrating into the display pipeline and performing a "reasoning-compensation-superposition-output" process on each frame of input signal in real time, dynamic, fully automatic, and pixel-level precision correction of display distortion on LCD panels is achieved, thereby achieving the ultimate goal of improving display uniformity and visual fidelity.

[0067] This invention proposes an image processing method and system for correcting distortion in liquid crystal panel displays. The core of this method lies in constructing a data-driven end-to-end process. By training a dedicated neural network model, it can automatically learn the accurate mapping relationship from a distorted display to an ideal display, and finally achieve efficient correction at the hardware level by generating a real-time pixel compensation matrix.

[0068] The entire technical process begins with the construction of high-quality training data. In a standardized production line testing environment, the system sequentially inputs a series of preset standard test images into the LCD panel to be calibrated. These images comprehensively cover the full grayscale range from pure black to pure white, red, green, and blue monochrome, as well as special patterns with clear geometric features. Simultaneously, a high-precision industrial camera synchronously captures the actual display image of the panel under strictly controlled lighting, temperature, and humidity conditions. Subsequently, image registration algorithms are used to eliminate geometric pose deviations between the camera and the panel, and luminance response normalization is used to remove the nonlinear effects of the camera's own photoelectric conversion, ultimately resulting in a set of strictly aligned and photometrically normalized image pairs. These image pairs accurately record the inherent distortion information of the panel, forming a reliable data foundation for subsequent model training.

[0069] The next stage involves the construction and training of the core correction model. The solution innovatively designs an improved U-Net network architecture that incorporates the Laplacian pyramid as the correction model. The input image is first decomposed into different scales of the Laplacian pyramid, allowing the network to explicitly and in parallel process distortion features of different spatial frequencies. For example, large-scale, gradual brightness unevenness belongs to low-frequency information, while small spots or line defects belong to high-frequency information. The U-Net network's encoder-decoder structure, combined with skip connections, excels at pixel-level dense prediction, enabling precise recovery of image details. The model's learning objective is to directly predict a pixel compensation matrix, where each value represents the specific amount of brightness or color adjustment needed for the corresponding pixel in the input image. The training process uses preprocessed image pairs as data, and the model parameters are continuously adjusted through optimization algorithms to minimize the difference between the predicted compensation matrix and the actual compensation values ​​until the model converges.

[0070] Finally, the trained model is deployed and applied. The optimized model is integrated into the real-time image processing pipeline of the display device. For each frame of the original image signal to be displayed, the system inputs it into the model. The model quickly outputs a pixel compensation matrix with the exact same size as the input image through a single forward propagation calculation. Subsequently, in the digital domain, each pixel value of the original input image is added to the corresponding compensation value in this compensation matrix. After necessary numerical range clipping to prevent overflow, the final calibrated display drive signal is generated. This signal is sent to the panel's drive circuitry, thereby enabling the image displayed on the screen to compensate for inherent distortion defects, achieving significantly improved uniformity and visual effects.

[0071] The present invention also protects an image processing system for correcting distortion in a liquid crystal panel display, comprising a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above method.

[0072] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, the phrase "comprising an element defined as..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0073] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their likenesses.

Claims

1. An image processing method for correcting distortion in a liquid crystal panel display, characterized in that, Includes the following steps: Step S1: Input the standard test image into the LCD panel and simultaneously acquire the actual physical image of the LCD panel display; Step S2: Perform image registration and brightness response normalization on the standard test image and the actual display physical image to form image pair training data; Step S3: Construct a display distortion correction model based on the U-Net network, and train the display distortion correction model using the image training data to obtain a trained display distortion correction model; Step S4: Input the original input frame to be displayed into the trained display distortion correction model, output the pixel compensation matrix, add the pixel value matrix of the original input frame to the pixel compensation matrix to generate the liquid crystal panel correction signal, and realize the display distortion correction of the liquid crystal panel.

2. The image processing method for correcting display distortion in a liquid crystal panel according to claim 1, characterized in that, Input a standard test image into the LCD panel and simultaneously acquire the actual physical image of the LCD panel's display, including: Set up a production line testing environment, which includes an image signal generator and a high-precision industrial camera. Connect the image signal generator to the LCD panel under test and drive it to sequentially input a set of preset standard test images. Simultaneously, a precisely positioned high-precision industrial camera is controlled to synchronously acquire images of the LCD panel's display area, obtaining images consistent with each standard test image. Strictly corresponding actual display physical image .

3. The image processing method for correcting display distortion in a liquid crystal panel according to claim 2, characterized in that, Image registration and brightness response normalization are performed on the standard test image and the actual display physical image to form image pair training data, including: Image registration was performed using the ORB feature matching algorithm, which is effective on standard test images. Compared with the actual displayed physical image Keypoints are detected and descriptors are calculated for each image, thereby establishing the correspondence between feature points in the two images. Based on a set of correct matching point pairs, a 3×3 homography matrix describing the projection transformation between the two planar images is calculated. Using this matrix to Perform geometric transformations and resample using bilinear interpolation to obtain the result. Registered physical image with precise alignment at the pixel level. ; Next, brightness response normalization is performed to correct the nonlinearity of the photoelectric response system combining the high-precision industrial camera and the LCD panel, by registering the physical image. The pixel values ​​are mapped to the Consistent linear brightness space; By sequentially performing the image registration and brightness response normalization processes described above, the original image pairs... Transformed into high-quality image pairs that can be directly used for learning. The set of all such processed image pairs, which constitutes the final image pair training data, is denoted as . ,in Represents the total number of image pairs, and i is the index of the image pair order.

4. The image processing method for correcting display distortion in a liquid crystal panel according to claim 3, characterized in that, A display distortion correction model based on the U-Net network is constructed, including: A display distortion correction model M is constructed, which adopts a U-Net network structure that incorporates the Laplacian pyramid. Its final output is the target pixel compensation matrix. Normalized distorted image As input to the model, it is then constructed. Layered Laplace Pyramid ,in, Represents the finest scale; Represents the coarsest scale; intermediate layer The number of pyramid layers corresponds to different intermediate scales. This is a preset hyperparameter; Secondly, each level of the Pyramid of Laplace Each is fed into a shared-weight encoder subnetwork. Feature extraction is performed during this process, which can be described as follows: ； in, Indicates from the pyramid number 1 Feature maps extracted from the layers; The learnable parameters representing the encoder subnetwork, and the shared weight design improves the parameter efficiency of the model; By concatenating all scale-aligned feature maps through channel stitching, a composite feature map is formed that integrates global contextual information and local detail information. ; Finally, the fusion features Input a decoder network decoder network Composed of multiple transposed convolutional layers and convolutional layers, it gradually restores the spatial resolution of the feature map through upsampling operations. The decoder ends with a layer that uses a linear activation function. A convolutional layer outputs a normalized, distorted image of the input. Pixel compensation matrices of identical size .

5. The image processing method for correcting display distortion in a liquid crystal panel according to claim 4, characterized in that, The calculation process of the pixel compensation matrix includes: The formula for calculating the prime compensation matrix is ​​as follows: ； in, This is the pixel compensation matrix predicted by the model; The learnable parameters of the decoder network, the pixel compensation matrix The middle is located in the coordinate Color channels element value , representing the input image The amount of compensation to be applied to the corresponding position and channel; a positive value indicates an increase in brightness or color intensity, while a negative value indicates a decrease.

6. The image processing method for correcting display distortion in a liquid crystal panel according to claim 5, characterized in that, The display distortion correction model is trained using the image and training data to obtain a trained display distortion correction model, including: For image pairs training data Each sample Calculate its true compensation matrix label The calculation method is the difference between the corresponding pixel values ​​of the ideal standard image and the normalized distorted image; Determine a loss function L to quantify the compensation matrix predicted by the model. With the true compensation matrix The difference between them is minimized using an adaptive moment estimation optimizer to minimize the loss function. During training, the training data is analyzed from images. Randomly select small batches of data Input the model and calculate the predicted values ​​through forward propagation. The loss is then calculated relative to the model parameters using the backpropagation algorithm. gradient And update the parameters according to the following rules: ； in, The learning rate controls the step size of each parameter update. Training terminates when the preset number of iterations is reached, or when the loss value on the independent validation dataset no longer decreases significantly. The model parameters that perform best on the validation set at this point are then saved. The complete computational graph corresponding to this set of parameters constitutes the trained display distortion correction model. .

7. The image processing method for correcting display distortion in a liquid crystal panel according to claim 6, characterized in that, The determination of the loss function includes: The mean absolute error is used as the loss function, which imposes a uniform linear penalty on the prediction error: ； in, and These represent the height and width of the image, respectively. and These represent the model prediction and the actual compensation matrices in coordinates, respectively. Color channels The value on; the summation symbol indicates that the absolute value errors of all spatial locations and color channels of the image are summed and averaged.

8. The image processing method for correcting display distortion in a liquid crystal panel according to claim 7, characterized in that, The original input frame to be displayed is input into the trained display distortion correction model, and the output is the pixel compensation matrix, including: During the operation of the display device, its image processing front end receives the raw input frame to be displayed, denoted as... Then, the original input frame Input the trained display distortion correction model In the middle, the trained display distortion correction model Perform forward propagation inference on the original input frame and output a value that is identical to the original input frame. Pixel compensation matrices of identical size.

9. The image processing method for correcting display distortion in a liquid crystal panel according to claim 8, characterized in that, The pixel value matrix and pixel compensation matrix of the original input frame are added together to generate a liquid crystal panel correction signal, thereby achieving liquid crystal panel display distortion correction, including: Obtain pixel compensation matrix Then, in the digital image processing domain, the original input frame is... Pixel value matrix and pixel compensation matrix Perform pixel-by-pixel, channel-by-channel addition, which generates an intermediate correction result matrix. ; For the intermediate correction result matrix A value range clipping operation is performed to limit all pixel values ​​to the valid range supported by the display device. Then, the signal is sent to the back-end processing unit of the display driver link, and converted into a correction drive signal that conforms to the electrical and timing interface specifications of the LCD panel via the timing controller and source driver standard hardware circuit.

10. An image processing system for correcting distortion in a liquid crystal panel display, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1-9.

Images