Liquid crystal display screen backlight uniformity compensation method and system based on deep learning

Abstract

The invention provides a liquid crystal display screen backlight uniformity compensation method and system based on deep learning, and relates to the technical field of liquid crystal display, and the method comprises the steps: obtaining a display image, carrying out the blocking processing, extracting a channel dimension feature, a spatial dimension feature and a multi-scale local feature, calculating a backlight compensation value through combining a deep neural network, and obtaining the backlight uniformity of a liquid crystal display screen. The ambient light intensity and the working temperature are considered, and the historical compensation effect is optimized and adjusted in combination with Kalman filtering, so that the backlight uniformity of the liquid crystal display screen can be effectively improved, and the display quality and the visual experience are improved.

Description

Technical Field

[0001] This invention relates to the field of liquid crystal display technology, and in particular to a method and system for compensating backlight uniformity of a liquid crystal display screen based on deep learning. Background Technology

[0002] LCD screens are widely used in various display devices such as televisions, computers, and mobile phones. The display quality largely depends on the uniformity of its backlight. Due to factors such as manufacturing process, component aging, and environmental conditions, LCD screens often have uneven backlighting, resulting in uneven brightness distribution of the displayed image and affecting the user's viewing experience.

[0003] Traditional backlight uniformity compensation methods are mainly based on statistical modeling and simple mathematical algorithms. By mathematically modeling the brightness distribution of the display screen, a compensation matrix is ​​calculated and applied to the backlight control system. However, existing technologies still have limitations, such as difficulty in accurately compensating for different image content features, insufficient consideration of environmental factors, inability to maintain stable display quality, significant differences in user experience under different usage scenarios, and difficulty in self-adjusting and optimizing based on historical compensation effects. As a result, the compensation effect gradually declines after long-term use, affecting the lifespan of the display screen and the stability of display quality. Summary of the Invention

[0004] This invention provides a method and system for compensating backlight uniformity of a liquid crystal display screen based on deep learning, which can at least solve some of the problems existing in the prior art.

[0005] A first aspect of this invention provides a deep learning-based method for compensating backlight uniformity in a liquid crystal display screen, comprising:

[0006] The real-time display image of the LCD screen is acquired and converted into a grayscale image. The grayscale image is then divided into multiple image blocks, and the average grayscale value of each image block is calculated.

[0007] Extract the channel dimension features and spatial dimension features corresponding to each image block, perform pooling operations at different scales on each image block to obtain local features at multiple scales, and weight and fuse the channel dimension features, spatial dimension features and local features at multiple scales to obtain a comprehensive feature vector. Input the comprehensive feature vector into a deep neural network, and output the backlight compensation value corresponding to each image block by minimizing feature representation differences and contrast loss.

[0008] A backlight compensation matrix is ​​generated based on the backlight compensation value and mapped to the backlight unit of the liquid crystal display screen;

[0009] The ambient light intensity and operating temperature of the LCD screen are collected, and the ambient light compensation gain matrix and temperature compensation gain matrix are calculated. Based on the backlight compensation matrix and the ambient light compensation gain matrix, the ambient light correction matrix is ​​determined. The ambient light correction matrix and the temperature compensation gain matrix are multiplied to obtain the ambient light temperature correction matrix. The historical compensation effect of each element in the ambient light temperature correction matrix is ​​calculated by Kalman filtering, and the elements of the ambient light temperature correction matrix are weighted and adjusted to obtain the optimized backlight compensation matrix.

[0010] Backlight compensation for the liquid crystal display is performed based on the optimized backlight compensation matrix.

[0011] In one alternative implementation,

[0012] Acquiring the real-time display image of the LCD screen and converting it into a grayscale image, then dividing the grayscale image into multiple image blocks and calculating the grayscale mean value of each image block includes:

[0013] The system acquires real-time images displayed on the LCD screen and converts these images into grayscale images according to a preset grayscale conversion coefficient.

[0014] The spacing between the backlight units of the LCD screen is collected and the size of the image block is set. The grayscale image is divided according to the size of the image block. The grayscale values ​​of all pixels in each image block are accumulated. The accumulated value is divided by the total number of pixels in the image block to obtain the grayscale mean value of each image block.

[0015] In one alternative implementation,

[0016] The channel dimension features and spatial dimension features corresponding to each image patch are obtained. Pooling operations at different scales are performed on each image patch to obtain local features at multiple scales. The channel dimension features, spatial dimension features, and local features at multiple scales are weighted and fused to obtain a comprehensive feature vector, including:

[0017] Image blocks are acquired and the corresponding channel data of the image blocks are extracted. Correlation analysis is performed on the extracted channel data to obtain correlation data. The channel data is divided into multiple channel groups. The feature map pixel values ​​in the current channel group are summed and normalized to obtain the response value in the current channel group.

[0018] The response correlation matrix is ​​calculated based on the difference between the response value in each channel group and the average response value corresponding to the image block. The features in the channel group are reordered and weighted according to the response correlation matrix to generate channel dimension features.

[0019] The image patch is divided into multiple grid regions. The pixel values ​​in the grid regions are summed and normalized. The response values ​​in the grid regions and the difference between the response values ​​of adjacent grid regions are calculated. A response transfer matrix is ​​constructed and the response transfer direction and response transfer intensity are determined. Based on the response transfer direction and response transfer intensity, the response values ​​of the current grid region and the response values ​​of adjacent grid regions are fused to generate spatial dimension features. Pooling operations at different scales are performed on the image patch to obtain multi-scale local features.

[0020] Based on the pre-set initial fusion weights, the channel dimension features, spatial dimension features, and multi-scale local features are weighted and summed to obtain the fused features. The initial fusion weights are iteratively adjusted to minimize the difference between two adjacent fused features, resulting in optimized fusion weights. The channel dimension features, spatial dimension features, and multi-scale local features are then weighted and summed again to obtain the comprehensive feature vector.

[0021] In one alternative implementation,

[0022] The comprehensive feature vector is input into a deep neural network, and by minimizing the feature representation difference and contrast loss, the backlight compensation value corresponding to each image patch is output, including:

[0023] The comprehensive feature vector is input into a deep neural network. Initial features are obtained by feature extraction based on the convolutional and activation layers in the deep neural network. Two-dimensional discrete Fourier transform is performed on the initial features to obtain frequency domain features. Amplitude spectrum features and phase spectrum features are extracted from the frequency domain features.

[0024] Calculate the statistical characteristics of the initial features, calculate the spatial feature difference value of the initial features based on the statistical characteristics, calculate the Euclidean distance between the amplitude spectrum features and the comprehensive feature vector in the frequency domain and the cosine difference between the phase spectrum features, respectively, and combine the calculated Euclidean distance value and cosine difference value to obtain the frequency domain feature difference value.

[0025] In the initial features, feature pairs from the same image patch are selected to construct positive sample pairs, and feature pairs from different image patches are selected to construct negative sample pairs. The cosine similarity between the positive and negative sample pairs is calculated. The contrast loss value of the cosine similarity is calculated based on the pre-set normalized temperature parameter. The spatial domain feature difference value, frequency domain feature difference value, and contrast loss value are weighted and combined to obtain the joint optimization target value.

[0026] The initial features are input into the fully connected layer. The initial features are transformed by the weight matrix and bias vector in the fully connected layer. The weight matrix and bias vector are optimized according to the joint optimization objective value to obtain the initial value of backlight compensation. The absolute difference between the initial values ​​of backlight compensation at spatially adjacent positions is calculated to obtain the smoothing constraint value. The joint optimization objective value and the smoothing constraint value are weighted and minimized to obtain the backlight compensation value corresponding to the image patch.

[0027] In one alternative implementation,

[0028] The backlight compensation matrix generated based on the backlight compensation value and mapped to the backlight unit of the liquid crystal display includes:

[0029] The backlight compensation values ​​are arranged according to the spatial position relationship of the image blocks in the original image to generate an initial backlight compensation matrix corresponding to the number of blocks in the original image.

[0030] A bilinear interpolation operation is performed on the initial backlight compensation matrix to expand the dimension of the initial backlight compensation matrix to a backlight compensation matrix that matches the number of backlight units in the liquid crystal display screen. The compensation values ​​in the backlight compensation matrix are then mapped to the corresponding backlight units in the backlight unit array of the liquid crystal display screen in a one-to-one correspondence manner.

[0031] In one alternative implementation,

[0032] The ambient light intensity and operating temperature of the LCD screen are collected, and the ambient light compensation gain matrix and temperature compensation gain matrix are calculated. Based on the backlight compensation matrix and the ambient light compensation gain matrix, the ambient light correction matrix is ​​determined. The ambient light correction matrix and the temperature compensation gain matrix are multiplied by a dot product to obtain the ambient light temperature correction matrix, which includes:

[0033] The ambient light intensity and operating temperature of the liquid crystal display screen are collected. The operating temperature data is recorded in multiple consecutive sampling periods and the operating temperature difference between adjacent sampling periods is calculated. The operating temperature difference is matched with the preset liquid crystal molecule temperature response curve to determine the response hysteresis coefficient of the liquid crystal molecule and the temperature response hysteresis period is calculated based on the response hysteresis coefficient.

[0034] The operating temperature and the temperature response hysteresis period are substituted into a pre-set nonlinear hysteresis compensation function. The amplitude of the temperature response is compensated by the amplitude compensation term in the nonlinear hysteresis compensation function, and the time delay of the temperature response is compensated by the phase compensation term, so as to obtain the compensated actual operating temperature.

[0035] Calculate the deviation between the compensated actual operating temperature and the preset optimal operating temperature, substitute the deviation into the hysteresis compensation gain function, adjust the nonlinear attenuation parameter of the hysteresis compensation gain function according to the magnitude of the deviation, and calculate the temperature compensation gain matrix.

[0036] The ambient light intensity is divided by the reference ambient light intensity to obtain the ambient light intensity ratio. The logarithm of the ambient light intensity ratio is calculated. The ambient light compensation gain matrix is ​​calculated based on the logarithm. The backlight compensation matrix is ​​multiplied by the ambient light compensation gain matrix to obtain the ambient light correction matrix.

[0037] The hysteresis characteristic compensation factor is calculated based on the temperature response hysteresis period. The ambient light correction matrix and the temperature compensation gain matrix are multiplied to obtain the first matrix. This first matrix is ​​then multiplied with the hysteresis characteristic compensation factor to obtain the ambient light temperature correction matrix.

[0038] In one alternative implementation,

[0039] The historical compensation effect of each element in the ambient light temperature correction matrix is ​​calculated using Kalman filtering, and the elements of the ambient light temperature correction matrix are weighted and adjusted to obtain the optimized backlight compensation matrix, which includes:

[0040] The ambient light temperature correction matrix is ​​received, and the ambient light temperature correction matrix is ​​divided into multiple sampling periods. In each sampling period, the historical compensation effect of each element in the ambient light temperature correction matrix is ​​obtained. The difference between the historical compensation effect of the current sampling period and the estimated compensation effect of the previous sampling period is calculated. The difference is multiplied by a preset Kalman gain to update the estimated compensation effect of the current sampling period and update the estimation error covariance, so as to obtain the historical compensation effect of each element in the ambient light temperature correction matrix.

[0041] The spatial correlation of each element in the ambient light temperature correction matrix is ​​calculated to obtain the kernel function value and construct the probability distribution of historical compensation effect. The mean vector and covariance matrix of the probability distribution are iteratively optimized by variational inference method to obtain the optimized covariance matrix. The square root of the diagonal element of the optimized covariance matrix is ​​calculated to obtain the uncertainty index. The uncertainty index is added to the preset sampling base to determine the sampling number. Each element in the ambient light temperature correction matrix is ​​sampled to obtain the sampled value and the confidence interval is calculated.

[0042] The uncertainty index of the sampled values ​​is calculated. The uncertainty index is substituted into a pre-set exponential function and normalized to obtain a weighted adjustment coefficient. The weighted adjustment coefficient is multiplied by the sampled values ​​and summed to obtain a weighted average value. The compensation direction is determined according to the sign of the confidence interval. The weighted average value is multiplied by a preset compensation intensity adjustment factor and then superimposed to obtain an optimized backlight compensation matrix.

[0043] A second aspect of the present invention provides a deep learning-based backlight uniformity compensation system for a liquid crystal display screen, comprising:

[0044] The first unit is used to acquire the real-time display image of the liquid crystal display screen and convert it into a grayscale image, divide the grayscale image into multiple image blocks and calculate the grayscale mean value of each image block;

[0045] The second unit is used to extract the channel dimension features and spatial dimension features corresponding to each image block, perform pooling operations at different scales on each image block to obtain local features at multiple scales, and weight and fuse the channel dimension features, the spatial dimension features and the local features at multiple scales to obtain a comprehensive feature vector. The comprehensive feature vector is input into a deep neural network, and the backlight compensation value corresponding to each image block is output by minimizing the feature representation difference and contrast loss.

[0046] The third unit is a backlight unit used to generate a backlight compensation matrix based on the backlight compensation value and map it to the liquid crystal display screen.

[0047] The fourth unit is used to collect the ambient light intensity and operating temperature of the LCD screen and calculate the ambient light compensation gain matrix and the temperature compensation gain matrix. Based on the backlight compensation matrix and the ambient light compensation gain matrix, the ambient light correction matrix is ​​determined. The ambient light correction matrix and the temperature compensation gain matrix are multiplied to obtain the ambient light temperature correction matrix. The historical compensation effect of each element in the ambient light temperature correction matrix is ​​calculated by Kalman filtering and the elements of the ambient light temperature correction matrix are weighted and adjusted to obtain the optimized backlight compensation matrix.

[0048] The fifth unit is used to perform backlight compensation on the liquid crystal display screen based on the optimized backlight compensation matrix.

[0049] A third aspect of the present invention provides an electronic device, comprising:

[0050] A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke instructions stored in the memory to perform the aforementioned method.

[0051] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0052] In this invention, deep learning methods are used to compensate for the backlight uniformity of a liquid crystal display screen. This method can automatically analyze the features of real-time displayed images, extract multi-dimensional and multi-scale image information, effectively identify and characterize backlight unevenness under different display scenarios, and improve the accuracy and adaptability of backlight uniformity compensation. By collecting ambient light intensity and operating temperature, the corresponding compensation gain matrix is ​​calculated, and a comprehensive correction matrix is ​​obtained through matrix operations. This allows the compensation effect to be dynamically adjusted according to changes in environmental conditions, significantly improving the robustness of the backlight compensation system in complex display environments. The Kalman filter algorithm is introduced to evaluate historical compensation effects and intelligently optimize and weight the ambient light temperature correction matrix, achieving adaptive optimization of backlight compensation. This avoids the problems of over-compensation or under-compensation, ensuring that the liquid crystal display screen always maintains optimal backlight uniformity under different usage scenarios, thus improving the user's visual experience. Attached Figure Description

[0053] Figure 1 This is a flowchart illustrating the backlight uniformity compensation method for liquid crystal displays based on deep learning, according to an embodiment of the present invention.

[0054] Figure 2 This is a flowchart illustrating the optimization of the ambient light temperature correction matrix for a deep learning-based liquid crystal display backlight uniformity compensation method according to an embodiment of the present invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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.

[0056] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0057] Figure 1 This is a flowchart illustrating the deep learning-based backlight uniformity compensation method for liquid crystal displays according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0058] The real-time display image of the LCD screen is acquired and converted into a grayscale image. The grayscale image is then divided into multiple image blocks, and the average grayscale value of each image block is calculated.

[0059] Extract the channel dimension features and spatial dimension features corresponding to each image block, perform pooling operations at different scales on each image block to obtain local features at multiple scales, and weight and fuse the channel dimension features, spatial dimension features and local features at multiple scales to obtain a comprehensive feature vector. Input the comprehensive feature vector into a deep neural network, and output the backlight compensation value corresponding to each image block by minimizing feature representation differences and contrast loss.

[0060] A backlight compensation matrix is ​​generated based on the backlight compensation value and mapped to the backlight unit of the liquid crystal display screen;

[0061] The ambient light intensity and operating temperature of the LCD screen are collected, and the ambient light compensation gain matrix and temperature compensation gain matrix are calculated. Based on the backlight compensation matrix and the ambient light compensation gain matrix, the ambient light correction matrix is ​​determined. The ambient light correction matrix and the temperature compensation gain matrix are multiplied to obtain the ambient light temperature correction matrix. The historical compensation effect of each element in the ambient light temperature correction matrix is ​​calculated by Kalman filtering, and the elements of the ambient light temperature correction matrix are weighted and adjusted to obtain the optimized backlight compensation matrix.

[0062] Backlight compensation for the liquid crystal display is performed based on the optimized backlight compensation matrix.

[0063] In one alternative implementation,

[0064] Acquiring the real-time display image of the LCD screen and converting it into a grayscale image, then dividing the grayscale image into multiple image blocks and calculating the grayscale mean value of each image block includes:

[0065] The system acquires real-time images displayed on the LCD screen and converts these images into grayscale images according to a preset grayscale conversion coefficient.

[0066] The spacing between the backlight units of the LCD screen is collected and the size of the image block is set. The grayscale image is divided according to the size of the image block. The grayscale values ​​of all pixels in each image block are accumulated. The accumulated value is divided by the total number of pixels in the image block to obtain the grayscale mean value of each image block.

[0067] A digital camera is used to capture real-time images of the LCD screen. The camera is positioned approximately 50 centimeters away from the screen, with the lens perpendicular to the screen surface to ensure accurate image representation. The camera resolution is set to 4032×3024 pixels, the exposure time to 1 / 60 second, and the ISO sensitivity to 400 to ensure clear capture of the LCD screen's content under normal indoor lighting conditions. After acquisition, the raw image is transmitted to a processing device via a data cable for further processing.

[0068] For the acquired real-time display image, the color image is converted to a grayscale image according to a preset grayscale conversion coefficient. For each pixel in the RGB color image, the grayscale value is calculated using a weighted average method: the pixel value of the red channel is multiplied by 0.299, the pixel value of the green channel is multiplied by 0.587, and the pixel value of the blue channel is multiplied by 0.114. These three weighted values ​​are then added together to obtain the grayscale value corresponding to that pixel. For example, for a pixel with an RGB value of (255, 200, 150), its corresponding grayscale value is calculated as: 0.299×255 + 0.587×200 + 0.114×150 = approximately 211.7, which is rounded down to 212.

[0069] After grayscale conversion, the appropriate size of the image blocks is determined, and the grayscale image is divided into multiple image blocks. The backlight unit spacing information of the LCD screen is acquired, and the spacing is measured to be 8.5 mm using a dedicated optical measuring instrument. Based on the measurement results, combined with the physical size of the display screen and the pixel resolution, the number of pixels corresponding to each backlight unit in the image is calculated. For example, for a 55-inch LCD screen with a resolution of 3840×2160, if the backlight unit spacing is 8.5 mm, each backlight unit corresponds to approximately 32×32 pixels in the acquired image. To ensure that an image block can completely cover at least one backlight unit, while also considering computational efficiency, the system sets the image block size to 64×64 pixels, a size that can contain approximately four complete backlight units.

[0070] After determining the image block size, the grayscale image is divided into multiple non-overlapping image blocks according to the set image block size of 64×64 pixels. For a grayscale image with a resolution of 4032×3024, it is divided into 63×47=2961 image blocks. There may be image blocks on the far right and at the bottom that are less than 64×64 pixels. These edge image blocks are specially processed, such as by padding with zeros or adjusting the block size to meet the processing requirements.

[0071] For each pre-defined image block, the grayscale mean is calculated. This involves iterating through each pixel within the block, accumulating its grayscale value, and then dividing the accumulated value by the total number of pixels in the block to obtain the grayscale mean of that block. For example, for a 64×64 pixel image block with 4096 pixels, assuming the sum of the grayscale values ​​of all pixels in the block is 819200, the grayscale mean of the block is 819200 ÷ 4096 = 200. This grayscale mean represents the average brightness level of that area and is an important indicator for evaluating the uniformity of a display screen.

[0072] The sliding window technique is used to calculate the grayscale mean. First, the sum of pixels in each row of the grayscale image is calculated and stored as a row sum array. When processing each image block, the sum of pixels in the corresponding row is accessed to quickly calculate the sum of grayscale values ​​of all pixels in the image block, thereby avoiding repeated pixel access operations and significantly improving the processing speed.

[0073] In this embodiment, by acquiring the real-time display image of the liquid crystal display screen and converting it into a grayscale image according to a preset grayscale conversion coefficient, the data dimensions in the subsequent processing can be effectively simplified, highlighting brightness information. By setting the image block size in conjunction with the spacing of the backlight units and dividing the grayscale image according to the image blocks, the image segmentation can correspond to the actual backlight structure, ensuring that data processing matches physical characteristics. By accumulating the grayscale values ​​of all pixels in the image block and calculating the average value, the overall brightness level of the corresponding area of ​​each image block can be accurately reflected, avoiding interference caused by abnormalities in a single pixel. This achieves reasonable quantification and feature extraction of the brightness distribution of the liquid crystal display screen, providing a reliable data foundation for subsequent backlight uniformity analysis and compensation, thereby improving the uniformity and visual effect of the display screen.

[0074] In one alternative implementation,

[0075] The channel dimension features and spatial dimension features corresponding to each image patch are obtained. Pooling operations at different scales are performed on each image patch to obtain local features at multiple scales. The channel dimension features, spatial dimension features, and local features at multiple scales are weighted and fused to obtain a comprehensive feature vector, including:

[0076] Image blocks are acquired and the corresponding channel data of the image blocks are extracted. Correlation analysis is performed on the extracted channel data to obtain correlation data. The channel data is divided into multiple channel groups. The feature map pixel values ​​in the current channel group are summed and normalized to obtain the response value in the current channel group.

[0077] The response correlation matrix is ​​calculated based on the difference between the response value in each channel group and the average response value corresponding to the image block. The features in the channel group are reordered and weighted according to the response correlation matrix to generate channel dimension features.

[0078] The image patch is divided into multiple grid regions. The pixel values ​​in the grid regions are summed and normalized. The response values ​​in the grid regions and the difference between the response values ​​of adjacent grid regions are calculated. A response transfer matrix is ​​constructed and the response transfer direction and response transfer intensity are determined. Based on the response transfer direction and response transfer intensity, the response values ​​of the current grid region and the response values ​​of adjacent grid regions are fused to generate spatial dimension features. Pooling operations at different scales are performed on the image patch to obtain multi-scale local features.

[0079] Based on the pre-set initial fusion weights, the channel dimension features, spatial dimension features, and multi-scale local features are weighted and summed to obtain the fused features. The initial fusion weights are iteratively adjusted to minimize the difference between two adjacent fused features, resulting in optimized fusion weights. The channel dimension features, spatial dimension features, and multi-scale local features are then weighted and summed again to obtain the comprehensive feature vector.

[0080] For each image block, the corresponding channel data is obtained. Since LCD screens typically use RGB three-channel display, each image block contains data from three channels: red, green, and blue. Taking a 160×135 pixel image block as an example, the three channel data of this image block are extracted, with each channel forming a 160×135 matrix. Correlation analysis is performed on the extracted channel data to calculate the correlation coefficients between different channels. Taking the RGB channels as an example, the correlation coefficients between R and G, R and B, and G and B are calculated, forming a 3×3 correlation matrix. Assuming the calculated correlation coefficients are: R and G = 0.82, R and B = 0.75, and G and B = 0.89, it indicates that the correlation between the G and B channels is the highest.

[0081] Based on correlation data, the channel data is grouped. Channels with a correlation coefficient greater than a preset threshold of 0.8 are grouped together, while channels with a correlation coefficient less than the threshold are grouped separately. For example, the correlation coefficient between channel G and channel B is 0.89, which is greater than the threshold of 0.8, so channel G and channel B are grouped together, and channel R is grouped separately, forming two channel groups. The feature map pixel values ​​in the current channel group are summed. For the GB channel group, the G channel value and B channel value at each pixel position are added to obtain a new 160×135 matrix. The summation result is normalized to map all pixel values ​​to the range of 0-1, resulting in the response value matrix of the current channel group. The R channel group is also normalized, resulting in a response value matrix between 0 and 1.

[0082] Calculate the difference between the response value of each channel group and the average response value of the image patch. Assuming the average response value of the GB channel group is 0.65, the average response value of the R channel group is 0.58, and the average response value of the entire image patch is 0.62, then the difference between the GB channel group and the R channel group is 0.03, and the difference is -0.04. Construct a response correlation matrix based on these differences. The matrix size is the number of channel groups multiplied by the number of channel groups; in this example, it is 2×2. Reorder the features within the channel groups according to the response correlation matrix, placing the channel groups with larger absolute differences at the top and assigning them higher weights. In this example, the absolute difference of the R channel group is 0.04, which is greater than the 0.03 of the GB channel group; therefore, the R channel group is placed first with a weight of 0.6, while the GB channel group has a weight of 0.4. Multiply the reordered channel group features by their weights and combine them to generate a channel-dimensional feature vector of size 1×512.

[0083] The image patch is divided into multiple grid regions, with a grid size of 20×20 pixels. For a 160×135 image patch, it can be divided into 8×7=56 grid regions. The pixel values ​​in each grid region are summed and normalized to obtain the response value of each grid region. The difference in response values ​​between adjacent grid regions is calculated to construct a response transfer matrix. The matrix size is the number of grid regions multiplied by the number of grid regions, i.e., 56×56. For any two adjacent grid regions i and j, if the response value of i is 0.72 and the response value of j is 0.65, then the response difference is 0.07.

[0084] The response propagation direction and intensity are determined based on the response propagation matrix. The response propagation direction points from the grid region with the highest response value to the grid region with the lowest response value, and the propagation intensity is proportional to the difference in response values. In the example above, the propagation direction points from grid region i to j, and the propagation intensity is 0.07. Based on the response propagation direction and intensity, the response values ​​of the current grid region and its adjacent grid regions are merged. If the response value of the current grid region is 0.65, the weighted average response value of the adjacent grid regions is 0.70, and the propagation intensity is 0.07, then the merged response value is 0.65 + 0.07 × 0.70 = 0.699. This process is performed on all grid regions to generate a spatial dimension feature vector of size 1 × 256.

[0085] Pooling operations at different scales are performed on image patches with kernel sizes of 2×2, 4×4, and 8×8, resulting in multiple downsampled feature maps. The 2×2 pooled feature map, with a size of 80×68, is then subjected to global average pooling to obtain a scalar value. Global average pooling is then performed on the 4×4 and 8×8 pooled feature maps respectively, each yielding a scalar value. These scalar values ​​are concatenated to form a multi-scale local feature vector of size 1×64.

[0086] The features are weighted and summed based on pre-set initial fusion weights. The initial weights are set as follows: channel dimension features 0.4, spatial dimension features 0.4, and multi-scale local features 0.2. The three features are then weighted and summed to obtain a fused feature vector of size 1×832. The weights are iteratively adjusted to minimize the difference between two adjacent fused feature vectors. The adjustment strategy uses gradient descent with a learning rate of 0.01 and 50 iterations. Assuming the optimized weights are: channel dimension features 0.35, spatial dimension features 0.45, and multi-scale local features 0.2, the three features are weighted and summed again using the optimized weights to obtain the final comprehensive feature vector of size 1×832.

[0087] In this embodiment, by performing correlation analysis on the channel data of the image block and dividing it into channel groups, the interference of redundant information can be effectively reduced, making the feature extraction process more focused on representative channel features. By constructing a response correlation matrix and reordering and weighting the channel group features, the channel information that has a greater impact on backlight uniformity can be highlighted, and the distinguishing ability of channel dimension features can be improved. Dividing the image block into multiple grid regions and calculating the response transfer matrix can capture the differences and trends of pixel values ​​in a local range. By iteratively optimizing the fusion weight, channel dimension features, spatial dimension features, and multi-scale local features are weighted and fused, which not only achieves effective complementarity of multi-source features, but also ensures the stability and accuracy of the comprehensive feature vector. It can comprehensively and accurately depict the brightness distribution characteristics of the liquid crystal display screen, providing high-quality input feature support for subsequent backlight uniformity compensation.

[0088] In one alternative implementation,

[0089] The comprehensive feature vector is input into a deep neural network, and by minimizing the feature representation difference and contrast loss, the backlight compensation value corresponding to each image patch is output, including:

[0090] The comprehensive feature vector is input into a deep neural network. Initial features are obtained by feature extraction based on the convolutional and activation layers in the deep neural network. Two-dimensional discrete Fourier transform is performed on the initial features to obtain frequency domain features. Amplitude spectrum features and phase spectrum features are extracted from the frequency domain features.

[0091] Calculate the statistical characteristics of the initial features, calculate the spatial feature difference value of the initial features based on the statistical characteristics, calculate the Euclidean distance between the amplitude spectrum features and the comprehensive feature vector in the frequency domain and the cosine difference between the phase spectrum features, respectively, and combine the calculated Euclidean distance value and cosine difference value to obtain the frequency domain feature difference value.

[0092] In the initial features, feature pairs from the same image patch are selected to construct positive sample pairs, and feature pairs from different image patches are selected to construct negative sample pairs. The cosine similarity between the positive and negative sample pairs is calculated. The contrast loss value of the cosine similarity is calculated based on the pre-set normalized temperature parameter. The spatial domain feature difference value, frequency domain feature difference value, and contrast loss value are weighted and combined to obtain the joint optimization target value.

[0093] The initial features are input into the fully connected layer. The initial features are transformed by the weight matrix and bias vector in the fully connected layer. The weight matrix and bias vector are optimized according to the joint optimization objective value to obtain the initial value of backlight compensation. The absolute difference between the initial values ​​of backlight compensation at spatially adjacent positions is calculated to obtain the smoothing constraint value. The joint optimization objective value and the smoothing constraint value are weighted and minimized to obtain the backlight compensation value corresponding to the image patch.

[0094] A comprehensive feature vector is obtained and input into a deep neural network for further processing and feature extraction. The deep neural network consists of multiple convolutional layers and activation layers, including five convolutional layers with kernel sizes of 3×3, 3×3, 5×5, 5×5, and 7×7, and convolutional channels of 64, 128, 256, 512, and 512, respectively. Each convolutional layer is followed by a LeakyReLU activation function with a negative slope parameter set to 0.2 to improve the network's non-linear expressive power. The network input is the aforementioned comprehensive feature vector with a dimension of 1×832, which is reshaped into a 32×26 two-dimensional feature map to meet the input requirements of convolutional operations.

[0095] After processing through convolutional and activation layers in a deep neural network, a feature tensor of size 16×13×512 is obtained, called the initial features. A two-dimensional discrete Fourier transform is performed on the initial features to convert the spatial domain information to the frequency domain. The Fourier transform processes each channel separately, maintaining the original number of channels, resulting in frequency domain features with the same dimension of 16×13×512. The frequency domain features are in complex form, containing real and imaginary parts. Amplitude spectrum features are extracted from the frequency domain features, calculated as the modulus of the complex number (the square root of the sum of the squares of the real and imaginary parts). Phase spectrum features are extracted, calculated as the argument of the complex number (the arctangent of the imaginary part divided by the real part). The extraction results in two feature tensors of size 16×13×512, representing the amplitude and phase characteristics of the frequency domain information, respectively.

[0096] The initial statistical features are calculated, including mean, variance, skewness, and kurtosis. For the mean, the average of all elements in the spatial dimension for each channel is calculated, resulting in a 1×512 vector. The variance is calculated by averaging the squared differences between each channel element and its mean, also resulting in a 1×512 vector. Skewness measures the asymmetry of the distribution, and kurtosis measures the sharpness of the distribution; both are calculated using the third and fourth central moments, each yielding a 1×512 vector. Based on the aforementioned statistical features, the spatial feature difference value of the initial features is calculated. The mean vector, variance vector, skewness vector, and kurtosis vector are concatenated into a 1×2048 vector, which is then mapped to a scalar value through a fully connected layer as the spatial feature difference value. This difference value reflects the degree of variation of the features in the spatial domain and is of great significance for evaluating backlight uniformity.

[0097] Calculate the Euclidean distance between the amplitude spectra of the amplitude spectral features and the composite feature vector in the frequency domain. Perform a two-dimensional discrete Fourier transform on the composite feature vector to extract its amplitude spectrum, and calculate the Euclidean distance between it and the previously obtained amplitude spectral features. Flatten the two amplitude spectra into one-dimensional vectors and calculate the Euclidean distance between them. Assuming the calculated distance value is 0.85, calculate the cosine difference between the phase spectral features and the cosine similarity of the two flattened phase spectral vectors. Subtract the similarity from 1 to obtain the difference value. Assuming the calculated cosine difference value is 0.32, weight the Euclidean distance value of 0.85 and the cosine difference value of 0.32 with weights of 0.6 and 0.4 respectively, resulting in a frequency domain feature difference value of 0.85 × 0.6 + 0.32 × 0.4 = 0.638.

[0098] Positive sample pairs are constructed by selecting feature pairs from the same image patch in the initial features. Within the same channel of the initial features, feature vectors from two adjacent regions are randomly selected as a pair. For example, feature values ​​at positions (5, 7) and (5, 8) in channel 256 are selected to form a pair. Similarly, negative sample pairs are constructed by selecting feature pairs from different image patches. This is done by extracting feature vector pairs from the same channel positions of different image patches. For example, feature values ​​at positions (5, 7) in channel 256 of the first image patch and at positions (5, 7) in channel 256 of the second image patch are extracted to form a pair. For the constructed sample pairs, cosine similarity is calculated. Assume that the cosine similarity of a positive sample pair is 0.92 and the cosine similarity of a negative sample pair is 0.15.

[0099] The contrast loss value of cosine similarity is calculated based on a pre-set normalized temperature parameter. The normalized temperature parameter is set to 0.07, which is used to adjust the smoothness of the similarity distribution. The contrast loss is calculated by comparing the similarity of positive sample pairs with the similarity of all negative sample pairs. For each positive sample pair, the exponential value of its similarity after temperature parameter adjustment is calculated, and then divided by the sum of the exponential values ​​of the similarity of that positive sample pair and all negative sample pairs after temperature parameter adjustment. The negative logarithm of this ratio is taken as the contrast loss value. Assume the calculated contrast loss value is 0.25.

[0100] The spatial feature difference value, frequency domain feature difference value, and contrast loss value are weighted and combined to obtain the joint optimization objective value. The weights of the three are set to 0.3, 0.3, and 0.4, respectively. Therefore, the joint optimization objective value is: spatial feature difference value × 0.3 + frequency domain feature difference value × 0.3 + contrast loss value × 0.4 = 0.3 × 0.5 + 0.3 × 0.638 + 0.4 × 0.25 = 0.4064.

[0101] The initial features are input into a fully connected layer, which consists of three layers with 256, 128, and 384 backlight units, respectively. The fully connected layer transforms the initial features using a weight matrix and a bias vector. The weight matrix is ​​initialized using the He initialization method, and the bias vector is initialized to zero. The initial backlight compensation value is calculated through forward propagation; this is a 384-dimensional vector corresponding to each backlight unit of the display screen. The absolute difference between the initial backlight compensation values ​​of spatially adjacent positions is calculated to obtain the smoothing constraint value. For example, if the compensation value of the i-th backlight unit is 0.72, and the compensation value of its adjacent j-th unit is 0.68, the absolute difference is 0.04. This difference is calculated for all adjacent unit pairs and summed to obtain the total smoothing constraint value, assumed to be 15.6.

[0102] The joint optimization objective value and smoothing constraint value are weighted and combined with weights of 0.7 and 0.3 respectively, resulting in 0.7 × 0.4064 + 0.3 × 15.6 = 5.075. This value is minimized using gradient descent to update the network parameters, including the weights and biases of convolutional and fully connected layers. The optimization process uses the Adam optimizer with a learning rate of 0.001 and a decay rate of 0.95 / epoch, and is trained for 2000 iterations. The backlight compensation values ​​corresponding to the image patches are then obtained.

[0103] In this embodiment, by inputting the comprehensive feature vector into a deep neural network and using convolutional and activation layers for feature extraction, the initial feature representation of image blocks in the spatial domain can be effectively obtained. By introducing a two-dimensional discrete Fourier transform, the features are mapped to the frequency domain, and amplitude and phase spectrum features are extracted. This simultaneously captures the overall energy distribution characteristics and local structural information of the image brightness distribution, enabling a multi-angle expression of backlight non-uniformity features. By calculating the Euclidean distance and cosine difference between the frequency domain amplitude and phase spectra, the consistency differences in the spatial and frequency domains can be accurately characterized. By constructing positive and negative sample pairs and calculating cosine similarity, and then combining it with the contrast loss function for optimization, the feature discrimination ability can be improved, ensuring the aggregation of features of the same image block and the separability of features of different image blocks. By weighted combination of spatial domain feature differences, frequency domain feature differences, and contrast loss values, a joint optimization objective is formed, which can comprehensively balance the feature expression effects of multiple dimensions and avoid the deviation caused by single index optimization. This not only effectively improves backlight uniformity but also significantly improves the overall quality and visual comfort of the display screen.

[0104] In one alternative implementation,

[0105] The backlight compensation matrix generated based on the backlight compensation value and mapped to the backlight unit of the liquid crystal display includes:

[0106] The backlight compensation values ​​are arranged according to the spatial position relationship of the image blocks in the original image to generate an initial backlight compensation matrix corresponding to the number of blocks in the original image.

[0107] A bilinear interpolation operation is performed on the initial backlight compensation matrix to expand the dimension of the initial backlight compensation matrix to a backlight compensation matrix that matches the number of backlight units in the liquid crystal display screen. The compensation values ​​in the backlight compensation matrix are then mapped to the corresponding backlight units in the backlight unit array of the liquid crystal display screen in a one-to-one correspondence manner.

[0108] The backlight compensation values ​​are arranged according to the spatial relationship of the image blocks in the original image to generate an initial backlight compensation matrix. Each backlight compensation value is a one-dimensional vector with the same size as the number of image blocks. Assuming the original image is divided into 24 × 16 = 384 image blocks, the backlight compensation value vector contains 384 elements. Based on the row and column positions of the image blocks in the original image, these 384 compensation values ​​are rearranged into a 24 × 16 matrix, which is the initial backlight compensation matrix. For example, the backlight compensation value corresponding to the image block in the i-th row and j-th column is placed in the i-th row and j-th column of the initial backlight compensation matrix. This arrangement aims to maintain the spatial continuity of the compensation values, facilitating subsequent interpolation operations.

[0109] For a typical 4K resolution (3840×2160) LCD screen, if the image patch size is set to 160×135 pixels, an initial backlight compensation matrix of 24×16 will be formed. The value of each element in the matrix ranges from 0 to 1, representing the adjustment coefficient for the backlight brightness at the corresponding position. A larger value indicates a greater increase in brightness, and a smaller value indicates a greater decrease in brightness. For example, a value of 0.85 at position (5, 8) in the initial backlight compensation matrix means that the backlight unit at that position needs to adjust its brightness to 85% of the base brightness.

[0110] The backlight units of an LCD screen are typically arranged in a matrix. However, the number and arrangement of the backlight units may not perfectly match the image block division. For example, a certain LCD screen uses a 32×18 backlight unit array, with a total of 576 backlight units. During image analysis, the image is divided into 24×16=384 image blocks. In order to accurately map the 384 compensation values ​​onto the 576 backlight units, the initial backlight compensation matrix needs to be dimensionally expanded.

[0111] Bilinear interpolation is performed on the initial backlight compensation matrix, expanding its dimensions from 24×16 to 32×18 to match the number of backlight units. Bilinear interpolation is a method of interpolation in two-dimensional space. It considers the values ​​of four points within a 2×2 neighborhood of the point to be interpolated, and calculates the value of the interpolated point using a distance-weighted average. For each position (x, y) in the expanded matrix, its corresponding floating-point coordinates (x', y') in the original matrix are found. This correspondence is determined by a linear mapping: x' = x × (original width - 1) / (new width - 1), y' = y × (original height - 1) / (new height - 1). In this example, x' = x × 23 / 31, y' = y × 15 / 17.

[0112] Since the calculated x' and y' are usually not integers and cannot directly correspond to elements in the original matrix, interpolation is required. Find the smallest rectangle containing (x', y'), which consists of four integer coordinate points in the original matrix: (x1, y1), (x1, y2), (x2, y1), and (x2, y2), where x1 = floor(x'), x2 = floor(x'), y1 = floor(y'), and y2 = floor(y'). Calculate the relative distances between (x', y') and the four corner points, converting these distances into weights. The weights are inversely proportional to the distances. Let dx = x' - x1 and dy = y' - y1, then the weights of the four corner points are (1 - dx) × (1 - dy), (1 - dx) × dy, dx × (1 - dy), and dx × dy, respectively. Multiply the values ​​of the original matrix at the four corner points by their corresponding weights and sum them to obtain the value at the interpolation point (x', y').

[0113] For example, suppose we need to calculate the value at position (10, 5) in the expanded matrix, which corresponds to the position in the original matrix as (10×23 / 31, 5×15 / 17)≈(7.42, 4.41). The four integer coordinates containing this point are (7, 4), (7, 5), (8, 4), and (8, 5). Calculate the relative distances: dx = 7.42 - 7 = 0.42, dy = 4.41 - 4 = 0.41. The weights of the four corner points are (1 - 0.42) × (1 - 0.41) = 0.3422, (1 - 0.42) × 0.41 = 0.2378, 0.42 × (1 - 0.41) = 0.2478, and 0.42 × 0.41 = 0.1722. Assuming the original matrix has values ​​of 0.75, 0.78, 0.72, and 0.76 at these four positions, the interpolation result is 0.75×0.3422+0.78×0.2378+0.72×0.2478+0.76×0.1722=0.7532.

[0114] Through bilinear interpolation, the initial backlight compensation matrix is ​​expanded into a 32×18 backlight compensation matrix, where each element corresponds to the compensation value of a backlight unit in the LCD screen. The compensation values ​​in the backlight compensation matrix are mapped to the corresponding backlight units in the LCD screen backlight unit array in a one-to-one correspondence. During the mapping process, the value in the i-th row and j-th column of the backlight compensation matrix is ​​directly applied to the unit in the i-th row and j-th column of the backlight unit array. For example, the value 0.7532 at position (10, 5) of the backlight compensation matrix will be applied to the backlight unit at position (10, 5) in the backlight unit array, adjusting the brightness of that unit to 75.32% of the reference brightness.

[0115] In practical applications, the compensation value needs to be converted into a control signal format that the backlight control system can recognize. Most LCD backlight control systems use pulse width modulation (PWM) signals to control backlight brightness, with signal ranges typically between 0-255 or 0-1023. Therefore, the compensation value in the 0-1 range needs to be linearly mapped to the corresponding PWM control range. For example, mapping 0.7532 to the 0-255 range yields a PWM value of 0.7532 × 255 ≈ 192. This PWM control value is then sent to the backlight drive circuit through the display's control interface to achieve precise adjustment of the backlight brightness.

[0116] In this embodiment, by arranging the backlight compensation values ​​according to the spatial position relationship of image blocks in the original image, the spatial correspondence between the compensation result and the image content can be maintained, thereby ensuring that the distribution of compensation values ​​matches the brightness changes of the original image. Based on the initial backlight compensation matrix, bilinear interpolation is performed to expand the dimension to be consistent with the number of backlight units of the liquid crystal display screen. This can achieve a smooth transition of compensation values ​​while ensuring that the overall brightness trend remains unchanged, avoiding boundary abruptness or graininess caused by direct mapping. The expanded compensation matrix is ​​mapped to the backlight unit array in a one-to-one correspondence manner, so that each backlight unit can obtain a precise compensation value, thereby achieving fine control of the display screen backlight. This can effectively improve the continuity and accuracy of the backlight compensation in spatial distribution, significantly improve backlight uniformity, avoid local bright spots or dark areas, and improve the overall consistency and visual comfort of the display screen.

[0117] In one alternative implementation,

[0118] The ambient light intensity and operating temperature of the LCD screen are collected, and the ambient light compensation gain matrix and temperature compensation gain matrix are calculated. Based on the backlight compensation matrix and the ambient light compensation gain matrix, the ambient light correction matrix is ​​determined. The ambient light correction matrix and the temperature compensation gain matrix are multiplied by a dot product to obtain the ambient light temperature correction matrix, which includes:

[0119] The ambient light intensity and operating temperature of the liquid crystal display screen are collected. The operating temperature data is recorded in multiple consecutive sampling periods and the operating temperature difference between adjacent sampling periods is calculated. The operating temperature difference is matched with the preset liquid crystal molecule temperature response curve to determine the response hysteresis coefficient of the liquid crystal molecule and the temperature response hysteresis period is calculated based on the response hysteresis coefficient.

[0120] The operating temperature and the temperature response hysteresis period are substituted into a pre-set nonlinear hysteresis compensation function. The amplitude of the temperature response is compensated by the amplitude compensation term in the nonlinear hysteresis compensation function, and the time delay of the temperature response is compensated by the phase compensation term, so as to obtain the compensated actual operating temperature.

[0121] Calculate the deviation between the compensated actual operating temperature and the preset optimal operating temperature, substitute the deviation into the hysteresis compensation gain function, adjust the nonlinear attenuation parameter of the hysteresis compensation gain function according to the magnitude of the deviation, and calculate the temperature compensation gain matrix.

[0122] The ambient light intensity is divided by the reference ambient light intensity to obtain the ambient light intensity ratio. The logarithm of the ambient light intensity ratio is calculated. The ambient light compensation gain matrix is ​​calculated based on the logarithm. The backlight compensation matrix is ​​multiplied by the ambient light compensation gain matrix to obtain the ambient light correction matrix.

[0123] The hysteresis characteristic compensation factor is calculated based on the temperature response hysteresis period. The ambient light correction matrix and the temperature compensation gain matrix are multiplied to obtain the first matrix. This first matrix is ​​then multiplied with the hysteresis characteristic compensation factor to obtain the ambient light temperature correction matrix.

[0124] The system collects ambient light intensity and operating temperature data for the LCD screen. Ambient light intensity is collected by an ambient light sensor integrated into the screen bezel, with a sampling frequency of once per second, a measurement range of 0-10000 lux, and a resolution of 1 lux. Operating temperature is collected by a temperature sensor array attached to the back panel of the screen. The array consists of 8×4 temperature sensors evenly distributed across the back panel area, with a measurement range of -10℃ to 80℃ and an accuracy of ±0.5℃. The sampling frequency is set to once every 5 seconds, with 100 consecutive sampling cycles to form an 8×4×100 temperature data tensor.

[0125] The system records operating temperature data over multiple consecutive sampling periods and calculates the temperature difference between adjacent sampling periods. For each temperature sensor location, the temperature difference between adjacent time points is calculated. For example, if a sensor's temperature at time t is 36.5℃ and its temperature at time t+1 is 37.2℃, the temperature difference at that location is 0.7℃. This process is repeated for all temperature sensor locations, resulting in an 8×4×99 temperature difference tensor. The temperature difference is then matched against a pre-defined liquid crystal molecule temperature response curve. This response curve is a set of experimentally measured response characteristic curves of liquid crystal molecules at different rates of temperature change. The curve set contains multiple curves, each corresponding to a temperature change rate range, describing the relationship between the response time and temperature of the liquid crystal molecules within that range.

[0126] The matching process is implemented using an interpolation algorithm. For the calculated temperature difference, the closest curve is found in a preset curve set, and the corresponding response characteristic parameters are extracted. The response hysteresis coefficient of the liquid crystal molecules is determined, which represents the degree of delay in the liquid crystal molecules' response to temperature changes. For example, when the temperature rise rate is 0.7℃ / 5 seconds, the corresponding response hysteresis coefficient is 1.35. The temperature response hysteresis period is calculated based on the response hysteresis coefficient by multiplying the response hysteresis coefficient by the reference response time (usually 16.7 milliseconds, i.e., one display refresh cycle). In this example, the temperature response hysteresis period is 1.35 × 16.7 = 22.5 milliseconds.

[0127] The operating temperature and temperature response hysteresis period are substituted into a pre-set nonlinear hysteresis compensation function, which consists of two parts: an amplitude compensation term and a phase compensation term. The amplitude compensation term adjusts the amplitude of the temperature response; it is an exponentially decaying function with parameters related to the operating temperature. The phase compensation term adjusts the time delay of the temperature response, using a sinusoidal function with parameters related to the temperature response hysteresis period. Assuming the current operating temperature is 36.5℃ and the temperature response hysteresis period is 22.5 milliseconds, after substituting these into the nonlinear hysteresis compensation function, the amplitude of the temperature response is increased by 12% through the amplitude compensation term, and the time delay of the temperature response is reduced by 5.3 milliseconds through the phase compensation term, resulting in a compensated actual operating temperature of 35.8℃.

[0128] The deviation between the compensated actual operating temperature and the preset optimal operating temperature is calculated. The preset optimal operating temperature of the LCD screen is typically 25℃, which is the temperature point where the liquid crystal molecules are most stable. In this embodiment, the compensated actual operating temperature is 35.8℃, and the deviation from the preset optimal operating temperature is 10.8℃. The deviation value is then substituted into the hysteresis compensation gain function. The hysteresis compensation gain function is a piecewise continuous nonlinear function, and different calculation rules are used for different deviation ranges. When the deviation value is less than 5℃, a linear gain is used; when the deviation value is between 5℃ and 15℃, a quadratic gain is used; and when the deviation value is greater than 15℃, an exponential gain is used.

[0129] The nonlinear attenuation parameter of the hysteresis compensation gain function is adjusted according to the magnitude of the deviation. The larger the deviation, the larger the nonlinear attenuation parameter, and the stronger the compensation effect. In this embodiment, the deviation is 10.8℃, which is in the range of 5℃ to 15℃. A quadratic gain is used, and the nonlinear attenuation parameter is set to 0.015. The temperature compensation gain matrix is ​​calculated, with a size consistent with the backlight unit array, which is 32×18. The value of each element in the matrix is ​​calculated using the hysteresis compensation gain function, representing the temperature compensation gain coefficient of the corresponding backlight unit. In the calculated temperature compensation gain matrix, the values ​​of each element are between 0.85 and 1.15, with a higher gain coefficient in the central area and a lower gain coefficient in the edge area, which is consistent with the temperature distribution characteristics of the liquid crystal display screen.

[0130] The ambient light intensity is divided by the reference ambient light intensity to obtain the ambient light intensity ratio. The reference ambient light intensity is set to 300 lux, representing the lighting level of a standard office environment. Assuming the current ambient light intensity is 1200 lux, the ambient light intensity ratio is 1200 / 300 = 4. The logarithm of the ambient light intensity ratio is calculated using a base-2 logarithm, resulting in a logarithm of 2 (because 2² = 4). The ambient light compensation gain matrix is ​​calculated based on this logarithm. The calculation method is to multiply the logarithm by a preset ambient light gain coefficient of 0.1, and then add 1, resulting in a base ambient light gain of 1.2. Considering the different effects of ambient light on different areas of the display screen, an ambient light distribution weight matrix is ​​introduced. This matrix is ​​calculated using the location information of the ambient light sensor and also has a size of 32×18. The ambient light compensation gain matrix is ​​obtained by multiplying the base ambient light gain by the ambient light distribution weight matrix.

[0131] The ambient light correction matrix is ​​obtained by multiplying the backlight compensation matrix and the ambient light compensation gain matrix. The multiplication operation is element-wise multiplication; that is, the element in the i-th row and j-th column of the ambient light correction matrix is ​​equal to the element in the i-th row and j-th column of the backlight compensation matrix multiplied by the element in the i-th row and j-th column of the ambient light compensation gain matrix. For example, if the value at position (5, 8) in the backlight compensation matrix is ​​0.75, and the value at the corresponding position in the ambient light compensation gain matrix is ​​1.15, then the value at that position in the ambient light correction matrix is ​​0.75 × 1.15 = 0.8625.

[0132] The hysteresis compensation factor is calculated based on the temperature response hysteresis period. The calculation method is to divide the temperature response hysteresis period by the reference response time and then take the inverse ratio. In this embodiment, the temperature response hysteresis period is 22.5 milliseconds, the reference response time is 16.7 milliseconds, and the hysteresis compensation factor is 16.7 / 22.5 = 0.742. The ambient light correction matrix and the temperature compensation gain matrix are multiplied element-wise to obtain the first matrix. The first matrix is ​​then multiplied by the hysteresis compensation factor to obtain the ambient light temperature correction matrix. During the multiplication operation, the hysteresis compensation factor is treated as a scalar and multiplied by each element in the first matrix. For example, if the value at position (5, 8) in the first matrix is ​​0.92 and the hysteresis compensation factor is 0.742, then the value at that position in the ambient light temperature correction matrix is ​​0.92 × 0.742 = 0.683.

[0133] The ambient light temperature correction matrix, a 32×18-inch control matrix used for backlight uniformity compensation, corresponds one-to-one with the backlight unit array of the LCD screen. Each value in the matrix represents the brightness adjustment coefficient of the corresponding backlight unit, which is directly applied to the control signal of the backlight drive circuit.

[0134] In this embodiment, by collecting the ambient light intensity and operating temperature of the liquid crystal display screen and matching the temperature change within a continuous sampling period with the temperature response curve of the liquid crystal molecules, the response hysteresis characteristics of the liquid crystal molecules can be accurately obtained. This allows for the calculation of the temperature response hysteresis period, enabling modeling of the dynamic temperature characteristics of the liquid crystal. A nonlinear hysteresis compensation function is used to perform dual compensation for the amplitude and phase of the operating temperature, effectively correcting response amplitude deviations and time delays caused by temperature changes. This yields a compensated operating temperature closer to the actual operating state. By calculating the deviation between the compensated actual operating temperature and the optimal operating temperature, and by introducing a hysteresis compensation gain function to dynamically adjust the nonlinear attenuation parameter, adaptive adjustment of temperature compensation can be achieved, thereby improving the accuracy and stability of compensation. By comprehensively considering the influence of ambient light and temperature hysteresis characteristics on the performance of the liquid crystal display, precise compensation under multi-factor coupling is achieved. This effectively eliminates the influence of temperature hysteresis and ambient light changes on the backlight uniformity and brightness consistency of the liquid crystal display screen, ensuring a stable, uniform, and clear visual effect under different temperature and lighting conditions. This significantly improves the adaptability and reliability of the liquid crystal display screen in complex environments.

[0135] In one alternative implementation,

[0136] The historical compensation effect of each element in the ambient light temperature correction matrix is ​​calculated using Kalman filtering, and the elements of the ambient light temperature correction matrix are weighted and adjusted to obtain the optimized backlight compensation matrix, which includes:

[0137] The ambient light temperature correction matrix is ​​received, and the ambient light temperature correction matrix is ​​divided into multiple sampling periods. In each sampling period, the historical compensation effect of each element in the ambient light temperature correction matrix is ​​obtained. The difference between the historical compensation effect of the current sampling period and the estimated compensation effect of the previous sampling period is calculated. The difference is multiplied by a preset Kalman gain to update the estimated compensation effect of the current sampling period and update the estimation error covariance, so as to obtain the historical compensation effect of each element in the ambient light temperature correction matrix.

[0138] The spatial correlation of each element in the ambient light temperature correction matrix is ​​calculated to obtain the kernel function value and construct the probability distribution of historical compensation effect. The mean vector and covariance matrix of the probability distribution are iteratively optimized by variational inference method to obtain the optimized covariance matrix. The square root of the diagonal element of the optimized covariance matrix is ​​calculated to obtain the uncertainty index. The uncertainty index is added to the preset sampling base to determine the sampling number. Each element in the ambient light temperature correction matrix is ​​sampled to obtain the sampled value and the confidence interval is calculated.

[0139] The uncertainty index of the sampled values ​​is calculated. The uncertainty index is substituted into a pre-set exponential function and normalized to obtain a weighted adjustment coefficient. The weighted adjustment coefficient is multiplied by the sampled values ​​and summed to obtain a weighted average value. The compensation direction is determined according to the sign of the confidence interval. The weighted average value is multiplied by a preset compensation intensity adjustment factor and then superimposed to obtain an optimized backlight compensation matrix.

[0140] The system receives an ambient light temperature correction matrix, which corresponds one-to-one with the backlight unit array of the LCD screen. The matrix is ​​divided into multiple sampling periods, each set to 30 seconds. Within each sampling period, the historical compensation effect of each element in the ambient light temperature correction matrix is ​​acquired. This historical compensation effect is obtained by collecting actual brightness uniformity data of the display screen after compensation, and is expressed as a percentage improvement in uniformity. For example, if the element at position (5, 8) in the ambient light temperature correction matrix had a historical compensation effect of 17.5% in the previous sampling period, it indicates that the backlight uniformity at that position has improved by 17.5%.

[0141] Calculate the difference between the historical compensation effect of the current sampling period and the estimated compensation effect of the previous sampling period. Assuming the estimated compensation effect of the element at position (5, 8) in the previous sampling period was 16.2%, and the historical compensation effect of the current sampling period is 17.5%, the difference is 17.5% - 16.2% = 1.3%. Multiply the difference by the preset Kalman gain to update the estimated compensation effect of the current sampling period. The Kalman gain is a value between 0 and 1, controlling the weight of new observation data; in this embodiment, it is set to 0.4. Therefore, the estimated compensation effect of the current sampling period is updated to 16.2% + 0.4 × 1.3% = 16.72%. Update the estimation error covariance, which represents the uncertainty of the compensation effect estimation. Assuming the estimation error covariance of the previous sampling period was 0.85, the preset process noise covariance was 0.2, and the measurement noise covariance was 0.3, the estimated error covariance of the current sampling period is updated to (1 - 0.4) × 0.85 = 0.51.

[0142] The spatial correlation of each element in the ambient light temperature correction matrix is ​​calculated to obtain the kernel function value. Spatial correlation is calculated using radial basis functions (RBFs), which consider the distance between elements in the matrix; the closer the distance, the higher the correlation. For any two elements (i, j) and (m, n) in the matrix, the Euclidean distance d is calculated, and then the kernel function value K = exp(-d) is calculated using the RDFs. 2 / σ 2 ), where σ is the bandwidth parameter, set to 3.5. For example, the distance between position (5, 8) and position (6, 8) is 1, and the kernel function value is exp(-1 / 12.25)≈0.922; while the distance to position (10, 15) is approximately 8.6, and the kernel function value is exp(-8.6).2 ( / 12.25)≈0.002.

[0143] A probability distribution of historical compensation effects is constructed based on the kernel function values. This probability distribution is a multivariate Gaussian distribution, with the mean vector representing the estimated compensation effect of each element, and the covariance matrix consisting of the kernel function values ​​and the estimation error covariance. The mean vector and covariance matrix of this probability distribution are iteratively optimized using variational inference. Variational inference is an approximate Bayesian inference method that iteratively updates the distribution parameters by minimizing the KL divergence. A coordinate descent algorithm is used, iterating 50 times, updating a subset of elements of the mean vector and covariance matrix each time. After iteration, the optimized covariance matrix is ​​obtained, where the diagonal elements represent the variance of the compensation effect of each backlight unit.

[0144] The square root of the diagonal elements of the optimized covariance matrix is ​​calculated to obtain the uncertainty index. For example, the diagonal element of the optimized covariance matrix corresponding to position (5, 8) is 0.0625, and the square root is 0.25, meaning the uncertainty index for this position is 0.25. The uncertainty index is added to the preset sampling base to determine the sampling quantity. The preset sampling base is 10, so the sampling quantity for position (5, 8) is 10 + 0.25 / 0.05 = 15 (the uncertainty index is divided by 0.05, rounded up, and then added to the base). Each element in the ambient light temperature correction matrix is ​​sampled. The sampling method is to randomly draw a specified number of samples from a normal distribution whose mean is the current compensation effect estimate and whose standard deviation is the uncertainty index. For example, the estimated compensation effect at position (5, 8) is 16.72%, with an uncertainty index of 0.25. A sample set of values ​​was obtained after 15 samplings: 16.54%, 16.63%, 16.89%, 16.75%, 16.42%, 16.97%, 16.81%, 16.59%, 16.70%, 16.85%, 16.77%, 16.68%, 16.90%, 16.66%, 16.72%.

[0145] The confidence interval was calculated based on the sampled values, with a confidence level set at 95%. The sample values ​​were sorted from smallest to largest, and the 2nd value (16.54%) and the 14th value (16.90%) were taken as the lower and upper limits of the confidence interval. Since both the lower and upper limits of the confidence interval are positive, the compensation direction was determined to be positive compensation, i.e., enhancing backlight brightness.

[0146] The uncertainty index of the sampled values ​​is calculated by calculating the sample standard deviation, which is 0.15. Substituting this uncertainty index into the pre-set exponential function f(x) = exp(-5x), we get exp(-5 × 0.15) ≈ 0.472. This value is calculated and normalized for all sampled points to obtain a weighted adjustment coefficient. The exponential function value for each point is divided by the sum of the exponential function values ​​for all points. For example, if the sum of the exponential function values ​​for 15 sampled points is 7.08, then the weighted adjustment coefficient for position (5, 8) is 0.472 / 7.08 ≈ 0.0667. The weighted adjustment coefficient is multiplied by the sampled values ​​and summed to obtain the weighted average. The weighted average is calculated as: 16.54% × 0.0667 + 16.63% × 0.0667 + ... + 16.72% × 0.0667 = 16.73%.

[0147] The compensation direction is determined based on the sign of the confidence interval. In this embodiment, the confidence interval is [16.54%, 16.90%], all of which are positive, so the compensation direction is positive. The weighted average value is multiplied by a preset compensation intensity adjustment factor and then superimposed on the corresponding element in the ambient light temperature correction matrix. The preset compensation intensity adjustment factor is 0.05, so the compensation value at position (5, 8) is 16.73% × 0.05 = 0.837%. If the original value at position (5, 8) in the ambient light temperature correction matrix is ​​0.683, then the optimized value is 0.683 × (1 + 0.00837) ≈ 0.689. This operation is performed on all elements in the ambient light temperature correction matrix to obtain the optimized backlight compensation matrix.

[0148] In this embodiment, by dividing the ambient light temperature correction matrix into multiple sampling periods and updating the estimated compensation effect using Kalman gain, recursive correction of historical compensation effects can be achieved in dynamic environments. This effectively reduces the impact of measurement noise and random fluctuations, thereby ensuring the continuity and stability of the compensation effect estimation. A kernel function is constructed based on the spatial correlation of matrix elements, and the probability distribution of historical compensation effects is iteratively optimized using variational inference methods. This not only obtains more accurate mean and covariance matrices but also calculates the uncertainty index through the square root of the diagonal elements of the covariance, providing a reliable basis for subsequent sampling and confidence interval calculation. By combining the uncertainty index with a preset sampling base to determine the sampling quantity, adaptive sample size adjustment can be achieved, thereby balancing computational efficiency and estimation accuracy. The compensation direction is determined by the sign of the confidence interval and then adjusted by the compensation intensity adjustment factor, avoiding over-compensation or under-compensation and ensuring the rationality and convergence of the compensation value. This not only achieves accurate estimation and robust optimization of backlight compensation values ​​but also takes into account real-time performance and stability, significantly improving the display consistency and visual quality of the LCD screen in complex environments.

[0149] Figure 2This is a flowchart illustrating the optimization of the ambient light temperature correction matrix for a deep learning-based liquid crystal display backlight uniformity compensation method according to an embodiment of the present invention.

[0150] A second aspect of the present invention provides a deep learning-based backlight uniformity compensation system for a liquid crystal display screen, comprising:

[0151] The first unit is used to acquire the real-time display image of the liquid crystal display screen and convert it into a grayscale image, divide the grayscale image into multiple image blocks and calculate the grayscale mean value of each image block;

[0152] The second unit is used to extract the channel dimension features and spatial dimension features corresponding to each image block, perform pooling operations at different scales on each image block to obtain local features at multiple scales, and weight and fuse the channel dimension features, the spatial dimension features and the local features at multiple scales to obtain a comprehensive feature vector. The comprehensive feature vector is input into a deep neural network, and the backlight compensation value corresponding to each image block is output by minimizing the feature representation difference and contrast loss.

[0153] The third unit is a backlight unit used to generate a backlight compensation matrix based on the backlight compensation value and map it to the liquid crystal display screen.

[0154] The fourth unit is used to collect the ambient light intensity and operating temperature of the LCD screen and calculate the ambient light compensation gain matrix and the temperature compensation gain matrix. Based on the backlight compensation matrix and the ambient light compensation gain matrix, the ambient light correction matrix is ​​determined. The ambient light correction matrix and the temperature compensation gain matrix are multiplied to obtain the ambient light temperature correction matrix. The historical compensation effect of each element in the ambient light temperature correction matrix is ​​calculated by Kalman filtering and the elements of the ambient light temperature correction matrix are weighted and adjusted to obtain the optimized backlight compensation matrix.

[0155] The fifth unit is used to perform backlight compensation on the liquid crystal display screen based on the optimized backlight compensation matrix.

[0156] A third aspect of the present invention provides an electronic device, comprising:

[0157] A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke instructions stored in the memory to perform the aforementioned method.

[0158] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0159] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0160] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for compensating backlight uniformity of a liquid crystal display based on deep learning, characterized in that, The method comprises the following steps: acquiring a real-time display image of a liquid crystal display and converting it into a grayscale image, dividing the grayscale image into multiple image blocks and calculating the grayscale mean value of each image block; extracting the channel dimension feature and the spatial dimension feature corresponding to each image block, performing a pooling operation of different scales on each image block, acquiring local features at multiple scales, weighting and fusing the channel dimension feature, the spatial dimension feature and the local features at multiple scales to obtain a comprehensive feature vector, inputting the comprehensive feature vector into a deep neural network and outputting a backlight compensation value corresponding to each image block by minimizing the feature representation difference and contrast loss; generating a backlight compensation matrix according to the backlight compensation value and mapping it to the backlight unit of the liquid crystal display; acquiring the ambient light intensity and working temperature of the liquid crystal display and calculating an ambient light compensation gain matrix and a temperature compensation gain matrix, determining an ambient light correction matrix based on the backlight compensation matrix and the ambient light compensation gain matrix, performing a point multiplication operation on the ambient light correction matrix and the temperature compensation gain matrix to obtain an ambient light temperature correction matrix, calculating the historical compensation effect of each element in the ambient light temperature correction matrix by Kalman filtering and weighting and adjusting the elements of the ambient light temperature correction matrix to obtain an optimized backlight compensation matrix; performing backlight compensation on the liquid crystal display based on the optimized backlight compensation matrix.

2. The method of claim 1, wherein, The method comprises the following steps: acquiring a real-time display image of a liquid crystal display and converting it into a grayscale image, dividing the grayscale image into multiple image blocks and calculating the grayscale mean value of each image block; acquiring a real-time display image of a liquid crystal display and converting it into a grayscale image, dividing the grayscale image into multiple image blocks and calculating the grayscale mean value of each image block; 3. The method of claim 1, wherein, acquiring the ambient light intensity and working temperature of the liquid crystal display and calculating an ambient light compensation gain matrix and a temperature compensation gain matrix, determining an ambient light correction matrix based on the backlight compensation matrix and the ambient light compensation gain matrix, performing a point multiplication operation on the ambient light correction matrix and the temperature compensation gain matrix to obtain an ambient light temperature correction matrix, calculating the historical compensation effect of each element in the ambient light temperature correction matrix by Kalman filtering and weighting and adjusting the elements of the ambient light temperature correction matrix to obtain an optimized backlight compensation matrix; performing backlight compensation on the liquid crystal display based on the optimized backlight compensation matrix. The method comprises the following steps: acquiring a real-time display image of a liquid crystal display and converting it into a grayscale image, dividing the grayscale image into multiple image blocks and calculating the grayscale mean value of each image block; acquiring a real-time display image of a liquid crystal display and converting it into a grayscale image, dividing the grayscale image into multiple image blocks and calculating the grayscale mean value of each image block; acquiring the ambient light intensity and working temperature of the liquid crystal display and calculating an ambient light compensation gain matrix and a temperature compensation gain matrix, determining an ambient light correction matrix based on the backlight compensation matrix and the ambient light compensation gain matrix, performing a point multiplication operation on the ambient light correction matrix and the temperature compensation gain matrix to obtain an ambient light temperature correction matrix, calculating the historical compensation effect of each element in the ambient light temperature correction matrix by Kalman filtering and weighting and adjusting the elements of the ambient light temperature correction matrix to obtain an optimized backlight compensation matrix; performing backlight compensation on the liquid crystal display based on the optimized backlight compensation matrix. The method comprises the following steps: acquiring a real-time display image of a liquid crystal display and converting it into a grayscale image, dividing the grayscale image into multiple image blocks and calculating the grayscale mean value of each image block; acquiring a real-time display image of a liquid crystal display and converting it into a grayscale image, dividing the grayscale image into multiple image blocks and calculating the grayscale mean value of each image block; acquiring the ambient light intensity and working temperature of the liquid crystal display and calculating an ambient light compensation gain matrix and a temperature compensation gain matrix, determining an ambient light correction matrix based on the backlight compensation matrix and the ambient light compensation gain matrix, performing a point multiplication operation on the ambient light correction matrix and the temperature compensation gain matrix to obtain an ambient light temperature correction matrix, calculating the historical compensation effect of each element in the ambient light temperature correction matrix by Kalman filtering and weighting and adjusting the elements of the ambient light temperature correction matrix to obtain an optimized backlight compensation matrix; performing backlight compensation on the liquid crystal display based on the optimized backlight compensation matrix. The method comprises the following steps: acquiring a real-time display image of a liquid crystal display and converting it into a grayscale image, dividing the grayscale image into multiple image blocks and calculating the grayscale mean value of each image block; acquiring a real-time display image of a liquid crystal display and converting it into a grayscale image, dividing the grayscale image into multiple image blocks and calculating the grayscale mean value of each image block; acquiring the ambient light intensity and working temperature of the liquid crystal display and calculating an ambient light compensation gain matrix and a temperature compensation gain matrix, determining an ambient light correction matrix based on the backlight compensation matrix and the ambient light compensation gain matrix, performing a point multiplication operation on the ambient light correction matrix and the temperature compensation gain matrix to obtain an ambient light temperature correction matrix, calculating the historical compensation effect of each element in the ambient light temperature correction matrix by Kalman filtering and weighting and adjusting the elements of the ambient light temperature correction matrix to obtain an optimized backlight compensation matrix; performing backlight compensation on the liquid crystal display based on the optimized backlight compensation matrix. The image block is divided into a plurality of grid regions, pixel values in the grid regions are summed and normalized, response values in the grid regions and response value differences of adjacent grid regions are calculated, a response transfer matrix is constructed, and response transfer directions and response transfer strengths are determined, response values of a current grid region and response values of adjacent grid regions are fused based on the response transfer directions and the response transfer strengths, spatial dimension features are generated, different scale pooling operations are performed on the image block, and multi-scale local features are obtained; The channel dimension features, the spatial dimension features, and the multi-scale local features are weighted and summed based on preset initial fusion weights to obtain fusion features, the initial fusion weights are iteratively adjusted to minimize differences between two adjacent fusion features, optimized fusion weights are obtained, and the channel dimension features, the spatial dimension features, and the multi-scale local features are weighted and summed again to obtain a comprehensive feature vector.

4. The method of claim 1, wherein, The comprehensive feature vector is input into a deep neural network, and a backlight compensation value corresponding to each image block is output by minimizing feature representation difference and contrast loss, including: The comprehensive feature vector is input into a deep neural network, and initial features are obtained by feature extraction based on convolution layers and activation layers in the deep neural network, frequency domain features are obtained by performing two-dimensional discrete Fourier transform on the initial features, and amplitude spectrum features and phase spectrum features are extracted from the frequency domain features; Statistical features of the initial features are calculated, spatial domain feature difference values of the initial features are calculated based on the statistical features, Euclidean distances between amplitude spectrums of the amplitude spectrum features and the comprehensive feature vector in the frequency domain and cosine differences between phase spectrums are calculated respectively, and the calculated Euclidean distance values and cosine difference values are combined to obtain frequency domain feature difference values; Feature pairs of the same image block are selected from the initial features to construct positive sample pairs, feature pairs of different image blocks are selected to construct negative sample pairs, cosine similarities of the positive sample pairs and the negative sample pairs are calculated, a contrast loss value of the cosine similarities is calculated based on a preset normalization temperature parameter, and spatial domain feature difference values, frequency domain feature difference values, and the contrast loss value are weighted and combined to obtain a joint optimization target value; The initial features are input into a fully connected layer, the initial features are transformed by a weight matrix and a bias vector in the fully connected layer, a backlight compensation preliminary value is obtained by optimizing the weight matrix and the bias vector according to the joint optimization target value, an absolute difference value between the backlight compensation preliminary values of spatially adjacent positions is calculated to obtain a smoothing constraint value, and the joint optimization target value and the smoothing constraint value are weighted and combined and optimized to minimize to obtain a backlight compensation value corresponding to the image block.

5. The method of claim 1, wherein, A backlight compensation matrix is generated based on the backlight compensation value and is mapped to backlight units of a liquid crystal display screen, including: The backlight compensation values are arranged according to spatial positional relationships of the image blocks in an original image to generate an initial backlight compensation matrix corresponding to a block number of the original image; A bilinear interpolation operation is performed on the initial backlight compensation matrix, a dimension of the initial backlight compensation matrix is expanded to a backlight compensation matrix matching a number of backlight units of the liquid crystal display screen, and the compensation values in the backlight compensation matrix are mapped to backlight units at corresponding positions in a backlight unit array of the liquid crystal display screen in a one-to-one manner.

6. The method of claim 1, wherein, Collecting ambient light intensity and working temperature of the liquid crystal display screen and calculating ambient light compensation gain matrix and temperature compensation gain matrix, determining ambient light correction matrix based on backlight compensation matrix and ambient light compensation gain matrix, and performing point multiplication operation between ambient light correction matrix and temperature compensation gain matrix to obtain ambient light temperature correction matrix, including: Collecting ambient light intensity and working temperature of the liquid crystal display screen, recording working temperature data in continuous multiple sampling periods and calculating working temperature difference value of adjacent sampling periods, matching working temperature difference value with preset liquid crystal molecule temperature response curve to determine response hysteresis coefficient of the liquid crystal molecule and calculating temperature response hysteresis period according to the response hysteresis coefficient; Substituting working temperature and temperature response hysteresis period into preset nonlinear hysteresis compensation function, compensating amplitude of temperature response through amplitude compensation term in nonlinear hysteresis compensation function, compensating time delay of temperature response through phase compensation term, and obtaining compensated actual working temperature; Calculating deviation value of compensated actual working temperature and preset optimal working temperature, substituting deviation value into hysteresis compensation gain function, adjusting nonlinear attenuation parameter of hysteresis compensation gain function according to size of deviation value, and calculating temperature compensation gain matrix; Dividing ambient light intensity by reference ambient light intensity to obtain ambient light intensity ratio, calculating logarithmic value of ambient light intensity ratio, calculating ambient light compensation gain matrix according to the logarithmic value, and multiplying backlight compensation matrix and ambient light compensation gain matrix to obtain ambient light correction matrix; Calculating hysteresis characteristic compensation factor according to temperature response hysteresis period, performing point multiplication between ambient light correction matrix and temperature compensation gain matrix to obtain first matrix, and performing point multiplication operation between the first matrix and the hysteresis characteristic compensation factor to obtain ambient light temperature correction matrix.

7. The method of claim 1, wherein, Calculating historical compensation effect of each element in the ambient light temperature correction matrix through Kalman filtering and weighting adjusting elements of the ambient light temperature correction matrix to obtain optimized backlight compensation matrix, including: Receiving ambient light temperature correction matrix, dividing ambient light temperature correction matrix into multiple sampling periods, obtaining historical compensation effect of each element in ambient light temperature correction matrix in each sampling period, calculating difference value between historical compensation effect of current sampling period and compensation effect estimation value of previous sampling period, multiplying the difference value by preset Kalman gain to update compensation effect estimation value of current sampling period and update estimation error covariance to obtain historical compensation effect of each element in ambient light temperature correction matrix; Calculating spatial correlation of each element in ambient light temperature correction matrix to obtain kernel function value and constructing probability distribution of historical compensation effect, obtaining optimized covariance matrix through variational inference method to iteratively optimize mean vector and covariance matrix of the probability distribution, calculating square root of diagonal elements of the optimized covariance matrix to obtain uncertainty index, adding the uncertainty index and preset sampling base to determine sampling number, sampling each element in ambient light temperature correction matrix to obtain sampling value and calculate confidence interval; The uncertainty index of the sampling values is calculated, the uncertainty index is substituted into a preset exponential function and normalized to obtain a weighted adjustment coefficient, the weighted adjustment coefficient is multiplied by the sampling values to obtain a weighted average value, a compensation direction is determined according to the sign of a confidence interval, and the weighted average value is multiplied by a preset compensation intensity adjustment factor to obtain an optimized backlight compensation matrix.

8. A deep learning based compensation system for liquid crystal display backlight uniformity, for implementing the method of any of the preceding claims 1-7, characterized in that, The method comprises the following steps: A first unit is configured to acquire a real-time display image of a liquid crystal display and convert the real-time display image into a grayscale image, divide the grayscale image into a plurality of image blocks, and calculate a grayscale mean value of each image block; A second unit is configured to extract a channel dimension feature and a spatial dimension feature corresponding to each image block, perform a pooling operation of different scales on each image block, acquire local features at a plurality of scales, and perform weighted fusion on the channel dimension feature, the spatial dimension feature, and the local features at the plurality of scales to obtain a comprehensive feature vector, input the comprehensive feature vector into a deep neural network, and output a backlight compensation value corresponding to each image block by minimizing a feature representation difference and a contrast loss; A third unit is configured to generate a backlight compensation matrix according to the backlight compensation value and map the backlight compensation matrix to a backlight unit of the liquid crystal display; A fourth unit is configured to acquire an ambient light intensity and a working temperature of the liquid crystal display, calculate an ambient light compensation gain matrix and a temperature compensation gain matrix, determine an ambient light correction matrix based on the backlight compensation matrix and the ambient light compensation gain matrix, perform a point multiplication operation on the ambient light correction matrix and the temperature compensation gain matrix to obtain an ambient light temperature correction matrix, calculate a historical compensation effect of each element in the ambient light temperature correction matrix by Kalman filtering, and perform weighted adjustment on the elements of the ambient light temperature correction matrix to obtain an optimized backlight compensation matrix; A fifth unit is configured to perform backlight compensation on the liquid crystal display based on the optimized backlight compensation matrix.

9. An electronic device, comprising: The method comprises the following steps: A processor; A memory for storing processor-executable instructions; The processor is configured to invoke the instructions stored in the memory to execute the method of any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon computer program instructions, wherein, The computer program instructions are executed by the processor to implement the method of any one of claims 1 to 7.

Images