Image recognition-based display color difference intelligent adjustment method and system

Abstract

The application discloses a display color difference intelligent adjustment method and system based on image recognition, and relates to the technical field of displays, which comprises the following steps: driving a display to display a standard pattern and collecting an image, calculating and outputting a pixel-level color difference map; inputting a full-screen RGB response matrix and the color difference map after channel-level splicing into an improved D-LKA feature extraction network for multi-scale feature extraction, and outputting a D-LKA multi-scale feature set; obtaining a physical color gamut degradation region set, a region label map and a soft edge weighting coefficient map; generating a display region consistency feature set; fitting to obtain a three-dimensional lookup table parameter set and a primary color gain parameter set corresponding to each physical color gamut degradation region; generating a screen-level color difference compensation matrix; and writing the screen-level color difference compensation matrix into a color gamut correction register or a non-volatile storage area of a built-in control chip of the display through a correction control interface. The application improves the screen color distortion topology construction precision.

Description

Technical Field

[0001] This invention relates to the field of display technology, and in particular to a display color difference intelligent calibration method and system based on image recognition. Background Technology

[0002] With the increasing application of OLED, Mini-LED and high color gamut LCD displays in professional imaging, automotive displays and industrial inspection scenarios, display color difference calibration has placed higher demands on full-screen color consistency, local color gamut accuracy and hardware writing stability. Existing display calibration is usually achieved through standard pattern acquisition, fixed partition measurement and writing of area calibration parameters, but it still has obvious shortcomings when dealing with color gamut caused by panel aging, backlight decay, heat accumulation and process differences.

[0003] Traditional full-screen uniform calibration methods are difficult to reflect the high-dimensional color gamut degradation at the screen edges, corners, and local pixel areas, which can easily lead to local color gamut volume shrinkage that cannot be accurately compensated.

[0004] Traditional multi-region correction methods typically rely on fixed grid division, with each grid independently calculating 3D lookup tables or primary color gain parameters. This can easily lead to parameter abrupt changes at the boundaries of regions, resulting in color breaks, stripe artifacts, and metamerism failures in large-area gradient images.

[0005] While existing deep feature extraction networks can extract spatial features from screen images, ordinary convolutional networks have fixed receptive fields, making it difficult to adapt to irregular color difference boundaries. Self-attention networks have high computational costs, making it difficult to meet the stability and efficiency requirements of production line calibration. Summary of the Invention

[0006] One objective of this invention is to propose a display color difference intelligent calibration method and system based on image recognition. This invention avoids boundary omissions and feature aliasing caused by traditional fixed sampling, and improves the accuracy of screen color distortion topology construction.

[0007] A display color difference intelligent calibration method based on image recognition according to an embodiment of the present invention includes: The system drives the display to show a standard pattern and captures images, generates a full-screen RGB response matrix, compares it pixel by pixel with the target matrix, calculates and outputs a pixel-level color difference map. The full-screen RGB response matrix and color difference map are concatenated at the channel level and then input into the improved D-LKA feature extraction network for multi-scale feature extraction, and the output is a D-LKA multi-scale feature set. The screen color distortion topology map is calculated based on the D-LKA multi-scale feature set. Adaptive clustering is performed on the screen color distortion topology map to obtain the physical color gamut degradation region set, region label map and soft edge weighting coefficient map. Based on the region label map, the D-LKA multi-scale feature set is modulated pixel by pixel to generate the display region consistency feature set; Based on the display area consistency feature set and the physical color gamut degradation area set, a three-dimensional lookup table parameter set and a primary color gain parameter set corresponding to each physical color gamut degradation area are obtained by fitting. Based on the soft edge weighting coefficient map, the three-dimensional lookup table parameter set and the primary color gain parameter set of each physical color gamut degradation region are jointly spatially interpolated and fused to generate a screen-level color difference compensation matrix. The screen-level color difference compensation matrix is ​​written into the color gamut correction register or non-volatile memory area of ​​the monitor's built-in control chip through the calibration control interface, thus completing the intelligent color difference adjustment of the monitor.

[0008] Optionally, the driving display shows a standard pattern and acquires an image, generates a full-screen RGB response matrix, and compares it pixel-by-pixel with the target matrix, including: The display to be calibrated is driven to emit light according to a preset standard grayscale and color test pattern. The image acquisition device is used to capture the full screen image synchronously, generate a full-screen RGB response matrix, compare the full-screen RGB response matrix with the reference RGB target matrix of the same size pixel by pixel, calculate the pixel-level color difference map and output the color difference map.

[0009] Optionally, the step of concatenating the full-screen RGB response matrix and color difference map at the channel level and then inputting them into an improved D-LKA feature extraction network for multi-scale feature extraction includes: The full-screen RGB response matrix and color difference map are stitched together according to the channel dimension to generate the display color difference input features; Perform shallow convolution on the color difference input features of the display to increase the feature dimensionality and generate a shallow feature map; The shallow feature map is input into the color difference-driven bias predictor to generate a deformable sampling bias matrix; Perform large-kernel dilated convolution with a set kernel size on the shallow feature map to generate a large-kernel feature map; Based on the large kernel feature map and the deformable sampling bias matrix, deformable sampling operation is performed to generate deformable sampling features. Softmax normalization is then performed on the deformable sampling features to generate the first-scale attention feature map. Pointwise convolution is used to perform channel mapping on the first-scale attention feature map to generate a first-scale projection feature map; The first-scale projected feature map and the shallow feature map are subjected to residual fusion to generate the first-scale residual fused feature map; The residual fusion feature map of the first scale is used as the input for the next scale until feature extraction of all scales is completed, generating a D-LKA multi-scale feature set.

[0010] Optionally, the calculation of the screen color distortion topology map based on the D-LKA multi-scale feature set includes: Scale alignment and channel mapping are performed on the residual fusion feature maps of each scale in the D-LKA multi-scale feature set to generate a unified scale distortion feature map. Scale weight fusion is performed on the uniform-scale distortion feature map to generate a screen color distortion topology map; Based on the screen color distortion topology map, construct the clustering feature vector corresponding to each screen pixel position; Based on the screen color distortion topology map, determine the number of physical color gamut degradation regions for adaptive clustering; Adaptive clustering is performed on the screen color distortion topology map based on the clustering feature vector and the number of physical color gamut degradation regions to generate the region belonging probability corresponding to each screen pixel position; Based on the region attribution probability, generate a set of physical color gamut degradation regions and a region label map; A soft-edge weighted coefficient map is generated based on the region affiliation probability and the region label map.

[0011] Optionally, the step of pixel-by-pixel modulation of the D-LKA multi-scale feature set based on the region label map includes: Construct a unique-hot encoding map of regions based on the region label map; Based on the spatial dimensions of the residual fusion feature maps at each scale in the D-LKA multi-scale feature set, scale synchronization is performed on the regional one-hot coding maps to generate regional coding maps at each scale. Perform region embedding mapping on the region encoding maps at each scale to generate region embedding feature maps at each scale; The residual fusion feature maps at each scale are connected pixel-by-pixel with the region embedding feature maps at the corresponding scale to generate joint feature maps of regions at each scale. Based on the joint feature map of each scale region, generate a consistent modulation coefficient map of each scale region. Based on the consistent modulation coefficient map of each scale region, pixel-wise modulation is performed on the residual fusion feature map of each scale in the D-LKA multi-scale feature set to generate consistent attention feature maps of each scale region. Summarize the region consistency attention feature maps corresponding to all scales to generate a display region consistency feature set.

[0012] Optionally, the step of fitting a three-dimensional lookup table parameter set and a primary color gain parameter set corresponding to each physical color gamut degradation region based on the display area consistency feature set and the physical color gamut degradation region set includes: Scale alignment and scale fusion are performed on the display area consistency feature set to generate a screen-level area consistency feature map. Extract the region consistency description vector corresponding to each physical gamut degradation region from the screen-level region consistency feature map; Based on the physical color gamut degradation region, the full-screen RGB response matrix, and the target matrix, a set of regional color fitting samples corresponding to each physical color gamut degradation region is constructed. Based on the region consistency description vector, generate the initial parameters of the three-dimensional lookup table and the initial parameters of the primary color gain for each physical color gamut degradation region; Based on the regional color fitting sample set and the initial parameters of the three-dimensional lookup table, fit the parameters of the three-dimensional lookup table corresponding to each physical color gamut degradation region. Based on the regional color fitting sample set, the three-dimensional lookup table parameters and the initial parameters of the primary color gain, the primary color gain parameters corresponding to each physical color gamut degradation region are fitted. Summarize the three-dimensional lookup table parameters and primary color gain parameters corresponding to all physical color gamut degradation regions, and generate a set of three-dimensional lookup table parameters and a set of primary color gain parameters.

[0013] Optionally, the step of performing joint spatial interpolation fusion of the three-dimensional lookup table parameter set and the primary color gain parameter set for each physical color gamut degradation region according to the soft edge weighting coefficient map includes: The three-dimensional lookup table parameter set is spatially interpolated and fused pixel by pixel according to the soft edge weighting coefficient map to generate pixel-level three-dimensional lookup table compensation terms corresponding to each screen pixel position. The set of primary color gain parameters is spatially interpolated and fused pixel by pixel according to the soft edge weighting coefficient map to generate a pixel-level primary color gain compensation term corresponding to each screen pixel position; The pixel-level 3D lookup table compensation item and the pixel-level primary color gain compensation item are jointly encapsulated to generate the pixel-level joint compensation item corresponding to each screen pixel position. A color difference compensation mapping relationship is established for each screen pixel position based on the pixel-level joint compensation term; The pixel-level joint compensation terms are arranged according to the positions of all screen pixels in the complete screen image of the display to be calibrated, and a screen-level color difference compensation matrix is ​​generated.

[0014] Optionally, the process of generating the data frame by writing the color gamut correction register to the data frame or writing the non-volatile memory area includes: Extract the pixel-level 3D lookup table compensation item and pixel-level primary color gain compensation item corresponding to each screen pixel position from the screen-level color difference compensation matrix. Based on the number of 3D lookup table nodes, the number of primary color gain channels and the data bit width supported by the display's built-in control chip, define the hardware mapping rules of the screen-level color difference compensation matrix. According to the hardware mapping rules, the three-dimensional lookup table writing sequence and the primary color gain writing sequence are subjected to fixed-point quantization processing to generate fixed-point three-dimensional lookup table writing data and fixed-point primary color gain writing data. When the display's built-in control chip uses a color gamut correction register to store correction parameters, the data written to the fixed-point three-dimensional lookup table and the data written to the fixed-point primary color gain are arranged according to the color gamut correction register address table to generate a color gamut correction register write data frame. When the display's built-in control chip uses a non-volatile memory area to store correction parameters, the fixed-point three-dimensional lookup table write data and the fixed-point primary color gain write data are arranged according to the non-volatile memory area page address, block address, and offset address to generate a non-volatile memory area write data frame. The system sends a write enable command to the display's built-in control chip through the calibration control interface, reads the write status flag returned by the display's built-in control chip, and when the write status flag indicates that writing is allowed, writes the color gamut correction register to the color gamut correction register using a write data frame, or writes the non-volatile memory area to the non-volatile memory area using a write data frame. After writing is complete, the readback correction data in the built-in control chip of the display is read through the correction control interface. The readback correction data is then compared with the data frame written to the color gamut correction register or the data frame written to the non-volatile memory area for consistency verification. If the consistency verification passes, a correction write completion flag is generated. Based on the calibration completion flag, the monitor's built-in control chip calls the written 3D lookup table compensation items and primary color gain compensation items to perform color difference compensation processing on the monitor's subsequent display input signals, thus completing the monitor's intelligent color difference calibration.

[0015] A display color difference intelligent calibration system based on image recognition, used to execute a display color difference intelligent calibration method, including: The acquisition module is used to drive the display to show a standard pattern and acquire images, generate a full-screen RGB response matrix, compare it with the target matrix pixel by pixel, calculate and output a pixel-level color difference map; The extraction module is used to input the full-screen RGB response matrix and color difference map into the improved D-LKA feature extraction network after channel-level concatenation, and output the D-LKA multi-scale feature set. The clustering module is used to calculate the screen color distortion topology map and perform adaptive clustering to obtain a set of physical color gamut degradation regions, a region label map, and a soft edge weighting coefficient map. The modulation module is used to modulate the D-LKA multi-scale feature set pixel by pixel according to the region label map to generate a display region consistency feature set. The fitting module is used to fit a three-dimensional lookup table parameter set and a primary color gain parameter set based on the display area consistency feature set and the physical color gamut degradation area set. The output module is used to perform joint spatial interpolation fusion according to the soft edge weighted coefficient map, generate a screen-level color difference compensation matrix, and write it into the display's built-in control chip through the calibration control interface.

[0016] The beneficial effects of this invention are: This invention employs an improved D-LKA feature extraction network, which can accurately capture the complex, irregular, and continuously changing color distortion space features of the display while maintaining the advantage of a large receptive field. This allows macroscopic low-frequency chromaticity Mura and local high-frequency color gamut degradation to be effectively expressed simultaneously, avoiding boundary omissions and feature aliasing caused by traditional fixed sampling, and improving the accuracy of screen color distortion topology construction.

[0017] This invention establishes a pixel-wise association between the region label map and the D-LKA multi-scale feature set. Through region one-hot encoding, region embedding mapping, region consistency modulation, and pixel-wise region attention enhancement, it ensures consistent color response expression within the same physical color gamut degradation region. At the same time, it effectively suppresses feature interference between different physical color gamut degradation regions. It can effectively eliminate color breaks, banding phenomena, and metamerism caused by parameter abrupt changes at region boundaries, enabling the display to maintain a smooth transition in scenes with gradient sky and continuous skin color. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a display color difference intelligent calibration method based on image recognition proposed in this invention; Figure 2 This is a diagram of the improved D-LKA feature extraction network structure of an image recognition-based intelligent color difference calibration method for displays proposed in this invention. Figure 3 This is a magnified view of the boundary of the traditional 8x8 fixed grid correction area in Embodiment 4 of the intelligent color difference adjustment method and system for displays based on image recognition proposed in this invention. Figure 4 This is a magnified view of the boundary of the area being corrected by the present invention in Embodiment 4 of the intelligent color difference adjustment method and system for displays based on image recognition proposed in this invention; Figure 5 This is a color difference heatmap of traditional 8x8 fixed grid correction in Embodiment 4 of the intelligent color difference adjustment method and system for displays based on image recognition proposed in this invention. Figure 6 This is the color difference heatmap corrected by the present invention in Embodiment 4 of the intelligent color difference adjustment method and system for display based on image recognition proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figure 1 As shown in Example 1: This example provides a display color difference intelligent calibration method based on image recognition, including: The system drives the display to show a standard pattern and captures images, generates a full-screen RGB response matrix, compares it pixel by pixel with the target matrix, calculates and outputs a pixel-level color difference map. The full-screen RGB response matrix and color difference map are concatenated at the channel level and then input into the improved D-LKA feature extraction network for multi-scale feature extraction, and the output is a D-LKA multi-scale feature set. The screen color distortion topology map is calculated based on the D-LKA multi-scale feature set. Adaptive clustering is performed on the screen color distortion topology map to obtain the physical color gamut degradation region set, region label map and soft edge weighting coefficient map. Based on the region label map, the D-LKA multi-scale feature set is modulated pixel by pixel to generate the display region consistency feature set; Based on the display area consistency feature set and the physical color gamut degradation area set, a three-dimensional lookup table parameter set and a primary color gain parameter set corresponding to each physical color gamut degradation area are obtained by fitting. Based on the soft edge weighting coefficient map, the three-dimensional lookup table parameter set and the primary color gain parameter set of each physical color gamut degradation region are jointly spatially interpolated and fused to generate a screen-level color difference compensation matrix. The screen-level color difference compensation matrix is ​​written into the color gamut correction register or non-volatile memory area of ​​the monitor's built-in control chip through the calibration control interface, thus completing the intelligent color difference adjustment of the monitor.

[0021] In this embodiment, the display is driven to show a standard pattern and acquire an image, generating a full-screen RGB response matrix, and comparing it pixel-by-pixel with the target matrix, including: The display to be calibrated is driven to emit light according to a preset standard grayscale and color test pattern. The image acquisition device is used to capture the full screen image synchronously, generate a full-screen RGB response matrix, compare the full-screen RGB response matrix with the reference RGB target matrix of the same size pixel by pixel, calculate the pixel-level color difference map and output the color difference map.

[0022] In this embodiment, the full-screen RGB response matrix and color difference map are concatenated at the channel level and then input into an improved D-LKA feature extraction network for multi-scale feature extraction, including: The full-screen RGB response matrix and color difference map are stitched together according to the channel dimension to generate the display color difference input features; In Example 1, the display color difference input feature simultaneously maintains the full-screen RGB response information and pixel-level color difference information of the display to be calibrated. It is obtained by connecting the red response channel, green response channel, and blue response channel in the full-screen RGB response matrix with the pixel-level color difference response channel in the color difference diagram according to the corresponding pixel positions.

[0023] Perform shallow convolution on the color difference input features of the display to increase the feature dimensionality and generate a shallow feature map; In Example 1, shallow convolution is used to perform convolution mapping on the color difference input features of the display, and nonlinear activation is applied to the convolution mapping result to obtain a shallow feature map.

[0024] The shallow feature map is input into the color difference-driven bias predictor to generate a deformable sampling bias matrix; In Example 1, the color difference driven bias predictor includes a color difference saliency extraction unit, a bias feature mapping unit, and a two-dimensional offset output unit.

[0025] The color difference saliency extraction unit extracts color difference saliency features formed by the propagation of pixel-level color difference information from the shallow feature map. The bias feature mapping unit generates large kernel sampling bias features corresponding to each screen pixel position based on the color difference saliency features. The two-dimensional offset output unit converts the large kernel sampling bias features into horizontal and vertical pixel coordinate offsets to obtain a deformable sampling bias matrix. The deformable sampling bias matrix describes the two-dimensional sampling offset of each large kernel sampling point corresponding to each screen pixel position relative to the regular sampling position. ; in, This represents the deformable sampling bias matrix, used to describe the two-dimensional sampling offset of each big kernel sampling point relative to the regular sampling position at each screen pixel location. , This indicates the number of pixels in the vertical direction of the full-screen image of the monitor to be calibrated. This indicates the number of pixels in the horizontal direction of the full-screen image of the monitor to be calibrated. This indicates the set kernel size used in large-kernel dilated convolution. This represents the number of large kernel sampling points corresponding to each screen pixel position, and 2 represents the offset components of the deformable sampling bias matrix in the horizontal and vertical pixel coordinates. This represents a shallow feature map, used to carry shallow joint features obtained by mapping the full-screen RGB response matrix and the color difference map. This indicates that the color difference saliency extraction unit and the bias feature mapping unit in the color difference driven bias predictor perform a first bias convolution mapping operation on the shallow feature map. The first bias convolution mapping operation is used to extract the color difference saliency features formed by the propagation of pixel-level color difference information from the shallow feature map, and generate large kernel sampling bias features corresponding to each screen pixel position based on the color difference saliency features. This represents the convolution weights of the first bias convolution mapping operation. This represents the bias term of the first bias convolution mapping operation. This represents a non-linear activation operation. This indicates the second bias convolution mapping operation performed by the two-dimensional offset output unit in the color difference-driven bias predictor on the large kernel sampling bias features. The second bias convolution mapping operation is used to convert the large kernel sampling bias features into horizontal and vertical pixel coordinate offsets. This represents the convolution weights of the second bias convolution mapping operation. This represents the bias term of the second bias convolution mapping operation.

[0026] The convolution weights of the first bias convolutional mapping operation and the second bias convolutional mapping operation are initialized using He normality and updated through backpropagation during the training process of display color difference samples based on the color difference significance features and the region boundary consistency error. The update method is to subtract the product of the learning rate and the corresponding error gradient from the current convolution weights and then limit them to a preset weight range, as shown in Example 1. The schematic value matrix is ​​as follows The schematic value matrix is ​​as follows , The range of values ​​is limited to The range of values ​​is limited to The selection criteria are to ensure that the first bias convolutional mapping operation retains the significant color difference characteristics formed by the propagation of pixel-level color difference information, and to keep the horizontal and vertical pixel coordinate offsets output by the second bias convolutional mapping operation within the range that can be stably interpolated at the large kernel sampling points.

[0027] Perform large-kernel dilated convolution with a set kernel size on the shallow feature map to generate a large-kernel feature map; In Example 1, the large kernel dilated convolution includes a large kernel sampling position unfolding unit, a dilated sampling response calculation unit, and a sampling response integration output unit.

[0028] The large kernel sampling position unfolding unit takes each screen pixel position as the center, determines multiple regular sampling positions corresponding to that screen pixel position according to the set convolution kernel size, and expands the pixel interval between adjacent regular sampling positions according to the set hole rate. The hole sampling response calculation unit extracts the shallow local features corresponding to each regular sampling position in the shallow feature map, and performs weighted calculation on the shallow local features and the large kernel convolution weights corresponding to the regular sampling positions. The sampling response integration output unit arranges the weighted calculation results of multiple regular sampling positions corresponding to each screen pixel position according to the sampling position order to generate a large kernel feature map. The large kernel feature map represents the color response change trend of each screen pixel position of the display to be calibrated within the large receptive field.

[0029] Set the kernel size The number of pixels is calculated based on the shorter side of the full-screen image of the monitor to be calibrated, and is not less than [amount missing]. The nearest odd number and restricted to Within the odd range, set the void ratio. according to Calculate and limit to Within the integer range, the resolution in Example 1 is Time to take The selection criteria are to ensure that the effective receptive field of the large kernel hole convolution covers the low-frequency spatial variation trend of the chromaticity Mura over a large area, and to avoid excessive smoothing of the local high-frequency color gamut degradation boundary due to an excessively large sampling span.

[0030] Based on the large kernel feature map and the deformable sampling bias matrix, deformable sampling operation is performed to generate deformable sampling features. Softmax normalization is then performed on the deformable sampling features to generate the first-scale attention feature map. Pointwise convolution is used to perform channel mapping on the first-scale attention feature map to generate a first-scale projection feature map; The first-scale projected feature map and the shallow feature map are subjected to residual fusion to generate the first-scale residual fused feature map; The residual fusion feature map of the first scale is used as the input for the next scale until feature extraction of all scales is completed, generating a D-LKA multi-scale feature set.

[0031] In Example 1, the spatial resolution of the feature map remains unchanged during the feature extraction process at each scale. High- and low-frequency composite color difference features at different receptive field depths are extracted by cascading repeated execution.

[0032] The process of repeatedly performing shallow convolutional feature upscaling, generating deformable sampling bias matrix, generating large kernel feature map, generating first-scale attention feature map, generating first-scale projection feature map, and generating first-scale residual fusion feature map is repeated, using the first-scale residual fusion feature map as input for the next scale, until feature extraction at all scales is completed.

[0033] refer to Figure 2 As shown, the improved D-LKA feature extraction network is used to receive the display color difference input features formed by the concatenation of the full-screen RGB response matrix and the color difference map at the channel level. The improved D-LKA feature extraction network includes a shallow convolutional feature enhancement layer, a color difference-driven bias predictor, a large-kernel dilated convolutional layer, a deformable sampling layer, a Softmax normalization layer, a pointwise convolutional channel mapping layer, and a residual fusion layer. Four scale levels are repeatedly set in a multi-scale cascade manner. Each scale level outputs a residual fusion feature map of the corresponding scale, forming a D-LKA multi-scale feature set. The shallow convolutional feature enhancement layer maps the display color difference input features. The shallow feature map is generated by a color difference-driven bias predictor, which generates a deformable sampling bias matrix based on the color difference significance features formed by the propagation of pixel-level color difference information in the shallow feature map. The large kernel dilated convolutional layer extracts the color response change trend of the display to be calibrated within a large receptive field according to the set convolutional kernel size and set dilation rate. The deformable sampling layer adjusts the spatial coordinates of the large kernel sampling points according to the deformable sampling bias matrix and performs bilinear interpolation sampling on the shallow feature map. The point-by-point convolutional channel mapping layer generates the first-scale projection feature map. The residual fusion layer performs residual fusion between the first-scale projection feature map and the shallow feature map to form the input for the next scale.

[0034] The output dimension of the improved D-LKA feature extraction network is ,in , , , The numbers represent the number of channels in the residual fusion feature maps at scales 1 to 4, respectively. In Example 1, these numbers are 64, 128, 256, and 512. The network consists of one input splicing layer, four shallow convolutional feature upscaling layers, four color difference-driven bias predictors, four large-kernel dilated convolutional layers, four deformable sampling layers, four Softmax normalization layers, four pointwise convolutional channel mapping layers, and four residual fusion layers. The parameter size is controlled between 8.6M and 13.8M. The training method is supervised training, using the display's color difference input features as training input. The color difference input features measured by a high-precision color analyzer are used as the training input. The obtained screen chromaticity shift distribution, chromaticity Mura annotation map, and regional gamut degradation annotation results are used as training labels. End-to-end training is performed using a joint loss function composed of screen color distortion topology map prediction error, regional boundary consistency error, and multi-scale feature smoothing constraint error. The training data comes from complete screen images, corresponding full-screen RGB response matrices, reference RGB target matrices, color difference maps, and true chromaticity measurement results obtained by color analyzer calibration from different batches of OLED displays, Mini-LED displays, and LCD displays under preset standard grayscale and color test patterns.

[0035] Compared with existing D-LKA feature extraction networks, the improved D-LKA feature extraction network in this embodiment uses the full-screen RGB response matrix and color difference map as input, enabling the network to obtain pixel-level color difference information in the initial stage of feature extraction. In the bias prediction stage, a color difference-driven bias predictor is set, which allows the deformable sampling bias matrix to adaptively shift around the irregular chromaticity Mura boundary, local color gamut degradation boundary, and backlight non-uniformity boundary. In the large kernel dilated convolution stage, the convolution kernel size and dilation rate are adaptively determined according to the full screen image size of the display to be corrected. This allows the network to simultaneously retain the spatial variation trend of large-area low-frequency chromaticity non-uniformity and the boundary features of local high-frequency color gamut degradation, improving the accuracy and continuity of screen color distortion topology map calculation, physical color gamut degradation region division, and screen-level color difference compensation matrix generation.

[0036] In this embodiment, the screen color distortion topology map is calculated based on the D-LKA multi-scale feature set, including: Scale alignment and channel mapping are performed on the residual fusion feature maps of each scale in the D-LKA multi-scale feature set to generate a unified scale distortion feature map. In Example 1, upsampling scale alignment is performed on the residual fusion feature map output at each scale to adjust the residual fusion feature map output at each scale to the same spatial size corresponding to the full screen image of the display to be corrected. Then, channel mapping convolution operation is performed on the upsampling scale aligned residual fusion feature map, followed by nonlinear activation operation, to generate a unified scale distortion feature map corresponding to the scale, which is used to carry the color response change features extracted at the corresponding scale under the full screen image size.

[0037] Scale weight fusion is performed on the uniform-scale distortion feature map to generate a screen color distortion topology map; In Example 1, each uniform-scale distortion feature map in the uniform-scale distortion feature map set is configured with a corresponding learnable scale weight parameter, and exponential normalization is performed on all learnable scale weight parameters to generate the scale fusion weight corresponding to each scale.

[0038] Each uniform-scale distortion feature map is multiplied by its corresponding scale fusion weight and then summed to generate a fused distortion feature map. A topological mapping convolution operation and a nonlinear activation operation are then performed on the fused distortion feature map to generate a screen color distortion topology map, which is used to describe the spatial distortion topology distribution in the complete screen image of the display to be corrected, which is formed by large-area chromaticity Mura, local color gamut degradation and backlight non-uniformity.

[0039] The learnable scale weight parameters are initialized with zero mean and updated during training by backpropagation based on the prediction error of the screen color distortion topology map and the region boundary consistency error. In Example 1, the learnable scale weight parameters corresponding to the four scales are η1=0.12, η2=0.28, η3=0.41, and η4=0.19, with the value range limited to [−1.00, 1.00].

[0040] Based on the screen color distortion topology map, construct the clustering feature vector corresponding to each screen pixel position; In Example 1, the topological features corresponding to each screen pixel position in the screen color distortion topology map are used as the color distortion representation of the screen pixel position. The horizontal pixel coordinates of the screen pixel position in the full screen image coordinate system are divided by the number of pixels in the horizontal direction of the full screen image of the display to be corrected to obtain the horizontal normalized coordinates of the screen pixel position. The vertical normalized coordinates are calculated using the same method.

[0041] The topological features corresponding to the screen pixel positions are multiplied by the horizontal normalized coordinates and the vertical normalized coordinates, and then concatenated to generate clustering feature vectors corresponding to the screen pixel positions. These vectors are used to simultaneously represent the color distortion topological features and normalized screen spatial positions of the screen pixel positions during adaptive clustering.

[0042] Based on the screen color distortion topology map, determine the number of physical color gamut degradation regions for adaptive clustering; In Example 1, the topological change intensity of the screen color distortion topology map is calculated. The topological change intensity is used to describe the degree of spatial change of color response distortion boundaries and color gamut degradation boundaries in the complete screen image of the display to be corrected. ; in, This represents the intensity of topological changes in the screen color distortion topology map. Represents the topology of screen color distortion In the Topological differential response at each screen pixel location along the horizontal pixel coordinate direction Represents the topology of screen color distortion In the Topological differential response at each screen pixel location along the vertical pixel coordinate direction This represents the operation of calculating the 1-norm.

[0043] Preset base area quantity Determine based on the area of ​​the complete screen image of the monitor to be calibrated. This is used to ensure a basic partition density in regular chroma Mura scenes, and the adjustment coefficient is... according to Calculations are used to increase the number of physical gamut degradation regions based on the intensity of topological changes, where Take 16 to 24. Take 6 to 12. Take 4 to 18, the resolution in Example 1 is Time to take The selection criteria are to enable the basic partition to cover a large area of ​​low-frequency chromaticity Mura, and to increase the number of regions when the screen color distortion topology changes strongly, in order to preserve the local high-frequency color gamut degradation boundary.

[0044] Adaptive clustering is performed on the screen color distortion topology map based on the clustering feature vector and the number of physical color gamut degradation regions to generate the region belonging probability corresponding to each screen pixel position; In Example 1, the clustering feature vector corresponding to each screen pixel position is used as the clustering input, and the number of physical color gamut degradation regions is used as the number of clusters. A corresponding cluster center is configured for each physical color gamut degradation region, and adaptive clustering is performed to obtain the region affiliation probability of the screen pixel position belonging to each physical color gamut degradation region. The region affiliation probability represents the probability degree of the screen pixel position belonging to the corresponding physical color gamut degradation region.

[0045] After calculating the region attribution probability, the cluster feature vector corresponding to each screen pixel position is multiplied by the region attribution probability of the screen pixel position belonging to the corresponding physical color gamut degradation region. The product results under the same physical color gamut degradation region are summed and divided by the sum of the region attribution probability corresponding to the physical color gamut degradation region to update the cluster center corresponding to the physical color gamut degradation region.

[0046] Based on the region attribution probability, generate a set of physical color gamut degradation regions and a region label map; In Example 1, the final region assignment probabilities of each screen pixel position belonging to each physical color gamut degradation region are compared. The physical color gamut degradation region with the highest final region assignment probability is selected as the region label corresponding to the screen pixel position. Screen pixel positions with the same region label are merged into the same physical color gamut degradation region, and a physical color gamut degradation region set is generated according to all physical color gamut degradation regions.

[0047] Write the region label corresponding to each screen pixel position into the corresponding position of that screen pixel position in the full screen image to generate a region label map, which is used to record the physical color gamut degradation region number corresponding to each screen pixel position.

[0048] A soft-edge weighted coefficient map is generated based on the region affiliation probability and the region label map.

[0049] In Example 1, the maximum value of the final region assignment probability of each screen pixel position belonging to each physical color gamut degradation region is selected, and the region boundary uncertainty is obtained by subtracting the maximum value of the final region assignment probability corresponding to the screen pixel position from the value of 1.

[0050] Generate a soft-edge weighted coefficient map based on the region boundary uncertainty: ; in, Indicates the first The position of the screen pixel relative to the first The soft edge weighting coefficients for each physical color gamut degradation region Indicates the first Clustering feature vectors corresponding to each screen pixel position Indicates the first step after adaptive clustering. The final cluster centers corresponding to each physical color gamut degradation region Indicates the first step after adaptive clustering. The final cluster centers corresponding to each physical color gamut degradation region This indicates the number of physical color gamut degradation regions obtained by adaptive clustering. The dimensionless soft-edge temperature coefficient represents the degree of smoothness controlled by the soft-edge weighting coefficient. Indicates the first The uncertainty of the region boundary corresponding to each screen pixel position is used to indicate the degree to which the screen pixel position is inside the physical color gamut degradation region or at the boundary of adjacent physical color gamut degradation regions. This represents the dimensionless boundary smoothing amplification term formed based on the uncertainty of the region boundary. This indicates exponentiation.

[0051] In this embodiment, pixel-by-pixel modulation of the D-LKA multi-scale feature set based on the region label map includes: Construct a unique-hot encoding map of regions based on the region label map; In Example 1, the region label value corresponding to each screen pixel position in the region label map is read, and the region label value is compared one by one with the physical color gamut degradation region number: If the region label value corresponding to the screen pixel position is equal to the current physical color gamut degradation region number, then the region one-hot encoding value of the screen pixel position relative to the current physical color gamut degradation region is set to 1.

[0052] If the region label value corresponding to the screen pixel position is not equal to the current physical color gamut degradation region number, then the region one-hot encoding value of the screen pixel position relative to the current physical color gamut degradation region is set to 0.

[0053] The region one-hot encoding values ​​are arranged according to the positions of all screen pixels and the numbers of all physical color gamut degradation regions to generate a region one-hot encoding map.

[0054] Based on the spatial dimensions of the residual fusion feature maps at each scale in the D-LKA multi-scale feature set, scale synchronization is performed on the regional one-hot coding maps to generate regional coding maps at each scale. In Example 1, the vertical and horizontal pixel counts of the residual fusion feature map at each scale in the D-LKA multi-scale feature set are read sequentially. According to the vertical and horizontal pixel counts of the residual fusion feature map at the current scale, the nearest neighbor scale synchronization is performed on the region one-hot coding map, so that the region one-hot coding value in the region one-hot coding map is mapped to the spatial position corresponding to the current scale residual fusion feature map, and the region coding map corresponding to the current scale is generated. This is used to maintain the physical color gamut degradation region attribution information corresponding to each feature position in the residual fusion feature map space at the current scale.

[0055] Perform region embedding mapping on the region encoding maps at each scale to generate region embedding feature maps at each scale; In Example 1, a region embedding convolutional mapping is performed on the region encoding map corresponding to each scale. The region embedding convolutional mapping maps the physical color gamut degradation region attribution information in the region encoding map to a region modulation feature with the same number of channels as the residual fusion feature map of the current scale. A nonlinear activation operation is performed on the region embedding convolutional mapping result to generate the region embedding feature map corresponding to the current scale.

[0056] The residual fusion feature maps at each scale are connected pixel-by-pixel with the region embedding feature maps at the corresponding scale to generate joint feature maps of regions at each scale. Based on the joint feature map of each scale region, generate a consistent modulation coefficient map of each scale region. In Example 1, the color response change features and physical gamut degradation region attribution features in the region joint feature map are converted into region consistency modulation response. The region consistency modulation response is subjected to Sigmoid normalization to generate the region consistency modulation coefficient map corresponding to the current scale. This is used to enhance the consistency of color response change features of the same physical gamut degradation region within the current scale and suppress feature interference between different physical gamut degradation regions.

[0057] Based on the consistent modulation coefficient map of each scale region, pixel-wise modulation is performed on the residual fusion feature map of each scale in the D-LKA multi-scale feature set to generate consistent attention feature maps of each scale region. In Example 1, each modulation coefficient in the region consistency modulation coefficient map corresponding to each scale is incremented by 1 to obtain the region enhancement coefficient corresponding to the current scale. The region enhancement coefficient corresponding to the current scale and the residual fusion feature map output by the current scale are multiplied pixel by pixel and channel by channel at the same screen space position and the same channel position to generate the region consistency attention feature map corresponding to the current scale.

[0058] Summarize the region consistency attention feature maps corresponding to all scales to generate a display region consistency feature set.

[0059] In this embodiment, based on the display area consistency feature set and the physical color gamut degradation region set, a three-dimensional lookup table parameter set and a primary color gain parameter set corresponding to each physical color gamut degradation region are fitted and obtained, including: Scale alignment and scale fusion are performed on the display area consistency feature set to generate a screen-level area consistency feature map. In Example 1, the screen-level region consistency feature map carries the color response change characteristics after the region label map is modulated under the full screen image size of the display to be calibrated: ; in, This represents a screen-level region consistency feature map. Indicates the first The region-consistency feature fusion weights corresponding to each scale Indicates the first The scale alignment operation involves upsampling the region consistency attention feature maps corresponding to each scale to the full-screen image size of the display to be corrected. Indicates the first Region-consistency feature channel mapping operation corresponding to each scale Indicates the first The convolution weights for the region-consistency feature channel mapping operation corresponding to each scale. Indicates the first The bias term of the region consistency feature channel mapping operation corresponding to each scale.

[0060] The region consistency feature fusion weights are obtained by exponentially normalizing the learnable fusion parameters corresponding to each scale. The convolution weights of the region consistency feature channel mapping operation are initialized using He normality and updated via backpropagation during the same training process based on the mapping error from the region consistency attention feature map at the corresponding scale to the screen-level region consistency feature map. In Example 1, the region consistency feature fusion weights corresponding to the four scales are shown as follows: The range of values ​​for the weights in the regional consistency feature fusion is limited to... Furthermore, the sum of the region consistency feature fusion weights across all scales is 1, and the convolution weights of the region consistency feature channel mapping operation... The schematic value matrix is ​​as follows Convolution weights for region consistency feature channel mapping operation The range of values ​​is limited to The selection criteria are as follows: the shallow region-consistent attention feature map retains the local high-frequency color gamut degradation boundary, the intermediate scale region-consistent attention feature map enhances the color response transition expression at the region boundary, and the deep region-consistent attention feature map retains the large area low-frequency color Mura space trend.

[0061] Specifically, the formula originates from the multi-scale feature pyramid fusion theory and convolutional feature mapping mechanism. The original form is used to uniformly map feature maps of different spatial resolutions and semantic levels to the same scale and perform weighted fusion. In Example 1, the ordinary image feature map is replaced with a region consistency attention feature map obtained by pixel-wise modulation of the region label map, enabling the multi-scale fusion process to apply to the region consistency color response features in intelligent color difference adjustment scenarios for displays. The upsampling scale alignment operation is used to eliminate spatial size differences between region consistency attention feature maps of different scales. The region consistency feature channel mapping operation is used to unify the channel expression dimension of features at different scales. The region consistency feature fusion weight is used to limit the contribution ratio of the color response change feature at the s-th scale in the screen-level region consistency feature map. A screen-level region consistency feature map is formed by weighted summation of the region consistency attention feature maps corresponding to all scales. The multi-scale feature fusion is derived from the ordinary visual feature enhancement form into a screen-level region consistency feature construction structure suitable for the joint expression of large-area low-frequency chromaticity Mura, local high-frequency color gamut degradation boundaries, and color response transitions at regional boundaries in displays.

[0062] Extract the region consistency description vector corresponding to each physical gamut degradation region from the screen-level region consistency feature map; In Example 1, the feature vectors of all screen pixel positions within the physical color gamut degradation region are read in the screen-level region consistency feature map. The feature vectors corresponding to all screen pixel positions within the physical color gamut degradation region are summed and divided by the number of screen pixel positions contained in the physical color gamut degradation region to obtain the region consistency description vector corresponding to the physical color gamut degradation region.

[0063] Based on the physical color gamut degradation region, the full-screen RGB response matrix, and the target matrix, a set of regional color fitting samples corresponding to each physical color gamut degradation region is constructed. In Example 1, for each screen pixel position, the red, green and blue response channel values ​​and target channel values ​​corresponding to the screen pixel position are read from the full-screen RGB response matrix and target matrix, respectively, and then the RGB response vector and RGB target vector corresponding to the screen pixel position are formed in order.

[0064] For each physical color gamut degradation region, the RGB response vector, RGB target vector, and the region consistency description vector corresponding to the physical color gamut degradation region are combined to generate a set of region color fitting samples.

[0065] Based on the region consistency description vector, generate the initial parameters of the three-dimensional lookup table and the initial parameters of the primary color gain for each physical color gamut degradation region; In Example 1, the region consistency description vector corresponding to the physical gamut degradation region is input into the three-dimensional lookup table parameter mapping unit. The three-dimensional lookup table parameter mapping unit performs parameter mapping on the region consistency description vector and rearranges the parameter mapping result into a three-dimensional lookup table node tensor to obtain the initial parameters of the three-dimensional lookup table. These parameters are used to provide the initial three-dimensional color mapping nodes for the physical gamut degradation region. Each three-dimensional lookup table node contains a red output channel, a green output channel, and a blue output channel.

[0066] The region consistency description vector corresponding to the physical color gamut degradation region is input into the primary color gain parameter mapping unit. The primary color gain parameter mapping unit performs parameter mapping on the region consistency description vector, performs hyperbolic tangent normalization on the parameter mapping result, multiplies it by the upper limit of the primary color gain adjustment range, and adds it to the three-dimensional unity gain vector composed of three values ​​of 1 to obtain the initial parameters of the primary color gain, which are used to provide the initial values ​​of the red primary color gain, green primary color gain and blue primary color gain of the physical color gamut degradation region.

[0067] Based on the regional color fitting sample set and the initial parameters of the three-dimensional lookup table, fit the parameters of the three-dimensional lookup table corresponding to each physical color gamut degradation region. In Example 1, the three-dimensional lookup table parameters are used to map the RGB response vectors within the physical gamut degradation region to a corrected output that approximates the RGB target vector: ; in, Indicates the first Parameters of the three-dimensional lookup table corresponding to each physical color gamut degradation region. This represents the variable of the three-dimensional lookup table to be optimized during the fitting process of the three-dimensional lookup table. This represents a three-dimensional lookup table trilinear interpolation mapping operation. Indicates a variable in a three-dimensional lookup table. Next The output vector obtained by trilinear interpolation of the RGB response vector corresponding to each screen pixel position. This represents the operation of calculating the square of the norm. The initial parameters, representing the weights maintained during the fitting process of the 3D lookup table, are adaptively calculated based on the sample size of the region color fitting sample set, the mean of the region boundary uncertainty, and the mean of the region color difference. Indicates the first Initial parameters for the 3D lookup table corresponding to each physical color gamut degradation region Represents the RGB response vector. This represents the RGB target vector.

[0068] The formula originates from least squares fitting theory and regularized parameter estimation methods. Its original form is used to construct a mapping function with minimum error between known input samples and target samples. In this embodiment 1, the ordinary input samples are replaced with the RGB response vector corresponding to each screen pixel position within the physical color gamut degradation region, and the ordinary target samples are replaced with the RGB target vector corresponding to the same screen pixel position. This allows least squares fitting to be applied to the three-dimensional color mapping space of the physical color gamut degradation region of the display. The three-dimensional lookup table trilinear interpolation mapping operation represents the nonlinear color correction process of the three-dimensional lookup table parameters on the RGB response vector, thus addressing the physical color gamut degradation. The region limits the range of screen pixel positions involved in the fitting. The initial parameters of the 3D lookup table constrain the fitting result to remain stable relative to the initial 3D color mapping node generated by the region consistency description vector. The initial parameters maintain a weight balance between the region color difference fitting accuracy and the smoothness of the 3D lookup table parameters. The 3D lookup table parameters corresponding to the nth physical color gamut degradation region are solved by using the error terms between all RGB response vectors and RGB target vectors in the region, as well as the maintenance terms of the 3D lookup table parameters relative to the initial parameters. The least squares parameter fitting form is derived into a region 3D lookup table parameter optimization structure suitable for multi-region color gamut degradation correction of displays.

[0069] Based on the regional color fitting sample set, the three-dimensional lookup table parameters and the initial parameters of the primary color gain, the primary color gain parameters corresponding to each physical color gamut degradation region are fitted. In Example 1, the primary color gain parameter represents the red, green, and blue primary color gain in the physical color gamut degradation region. The calculation process of the primary color gain parameter is the same as that of the three-dimensional lookup table parameter.

[0070] Summarize the three-dimensional lookup table parameters and primary color gain parameters corresponding to all physical color gamut degradation regions, and generate a set of three-dimensional lookup table parameters and a set of primary color gain parameters.

[0071] In this embodiment, joint spatial interpolation fusion is performed on the three-dimensional lookup table parameter set and the primary color gain parameter set of each physical color gamut degradation region according to the soft edge weighting coefficient map, including: The three-dimensional lookup table parameter set is spatially interpolated and fused pixel by pixel according to the soft edge weighting coefficient map to generate pixel-level three-dimensional lookup table compensation terms corresponding to each screen pixel position. The set of primary color gain parameters is spatially interpolated and fused pixel by pixel according to the soft edge weighting coefficient map to generate a pixel-level primary color gain compensation term corresponding to each screen pixel position; The pixel-level 3D lookup table compensation item and the pixel-level primary color gain compensation item are jointly encapsulated to generate the pixel-level joint compensation item corresponding to each screen pixel position. A color difference compensation mapping relationship is established for each screen pixel position based on the pixel-level joint compensation term; In Example 1, a pixel-level three-dimensional lookup table compensation term corresponding to the screen pixel position is used to perform three-dimensional lookup table trilinear interpolation mapping on the RGB response vector corresponding to the screen pixel position to obtain the three-dimensional lookup table mapping output vector.

[0072] A diagonal gain matrix composed of pixel-level primary color gain compensation terms is used to perform primary color gain correction on the three-dimensional lookup table mapped output vector to obtain the compensated RGB output vector, which is used to represent the color difference compensation result after the combined effect of the three-dimensional lookup table compensation terms and the primary color gain compensation terms.

[0073] The pixel-level joint compensation terms are arranged according to the positions of all screen pixels in the complete screen image of the display to be calibrated, and a screen-level color difference compensation matrix is ​​generated.

[0074] In Example 1, for all screen pixel positions in the complete screen image of the display to be calibrated, the corresponding pixel-level joint compensation terms are read respectively, and the pixel-level joint compensation terms corresponding to all screen pixel positions are arranged according to their spatial positions in the complete screen image of the display to be calibrated to obtain the screen-level color difference compensation matrix.

[0075] The screen-level color difference compensation matrix records the pixel-level joint compensation terms corresponding to each screen pixel position in the complete screen image of the display to be calibrated. The screen-level color difference compensation matrix contains both three-dimensional lookup table compensation terms and primary color gain compensation terms, and has pixel-level continuous transition characteristics.

[0076] In this embodiment, the process of generating a data frame by writing the color gamut correction register or the non-volatile memory area includes: Extract the pixel-level 3D lookup table compensation item and pixel-level primary color gain compensation item corresponding to each screen pixel position from the screen-level color difference compensation matrix. Based on the number of 3D lookup table nodes, the number of primary color gain channels and the data bit width supported by the display's built-in control chip, define the hardware mapping rules of the screen-level color difference compensation matrix. In Example 1, the hardware mapping rule is to convert the screen-level color difference compensation matrix into a correspondence that the control chip can recognize and write data, based on the number of three-dimensional lookup table nodes, the number of primary color gain channels, the data bit width, the register address order, and the non-volatile memory area page address structure supported by the display's built-in control chip.

[0077] Hardware mapping rules include 3D lookup table compensation item mapping rules, primary color gain compensation item mapping rules, fixed-point quantization rules, and address arrangement rules.

[0078] The 3D lookup table compensation item mapping rule arranges the pixel-level 3D lookup table compensation items into a 3D lookup table writing sequence in ascending order of nodes on the red input axis, green input axis, and blue input axis.

[0079] The primary color gain compensation term mapping rule is used to arrange the pixel-level primary color gain compensation terms in the channel order of red primary color gain, green primary color gain, and blue primary color gain as the primary color gain write sequence.

[0080] Fixed-point quantization rules are used to convert the three-dimensional lookup table write sequence and the primary color gain write sequence into fixed-point write data according to the data bit width of the display's built-in control chip.

[0081] Address arrangement rules are used to arrange fixed-point write data into gamut correction register write data frames or non-volatile memory area write data frames according to the gamut correction register address table or the page address, block address, and offset address of the non-volatile memory area.

[0082] According to the hardware mapping rules, the three-dimensional lookup table writing sequence and the primary color gain writing sequence are subjected to fixed-point quantization processing to generate fixed-point three-dimensional lookup table writing data and fixed-point primary color gain writing data. When the display's built-in control chip uses a color gamut correction register to store correction parameters, the data written to the fixed-point three-dimensional lookup table and the data written to the fixed-point primary color gain are arranged according to the color gamut correction register address table to generate a color gamut correction register write data frame. When the display's built-in control chip uses a non-volatile memory area to store correction parameters, the fixed-point three-dimensional lookup table write data and the fixed-point primary color gain write data are arranged according to the non-volatile memory area page address, block address, and offset address to generate a non-volatile memory area write data frame. The system sends a write enable command to the display's built-in control chip through the calibration control interface, reads the write status flag returned by the display's built-in control chip, and when the write status flag indicates that writing is allowed, writes the color gamut correction register to the color gamut correction register using a write data frame, or writes the non-volatile memory area to the non-volatile memory area using a write data frame. After writing is complete, the readback correction data in the built-in control chip of the display is read through the correction control interface. The readback correction data is then compared with the data frame written to the color gamut correction register or the data frame written to the non-volatile memory area for consistency verification. If the consistency verification passes, a correction write completion flag is generated. Based on the calibration completion flag, the monitor's built-in control chip calls the written 3D lookup table compensation items and primary color gain compensation items to perform color difference compensation processing on the monitor's subsequent display input signals, thus completing the monitor's intelligent color difference calibration.

[0083] Example 2: This example provides an image recognition-based intelligent color difference calibration system for displays, used to execute an image recognition-based intelligent color difference calibration method for displays, including: The acquisition module is used to drive the display to show a standard pattern and acquire images, generate a full-screen RGB response matrix, compare it with the target matrix pixel by pixel, calculate and output a pixel-level color difference map; The extraction module is used to input the full-screen RGB response matrix and color difference map into the improved D-LKA feature extraction network after channel-level concatenation, and output the D-LKA multi-scale feature set. The clustering module is used to calculate the screen color distortion topology map and perform adaptive clustering to obtain a set of physical color gamut degradation regions, a region label map, and a soft edge weighting coefficient map. The modulation module is used to modulate the D-LKA multi-scale feature set pixel by pixel according to the region label map to generate a display region consistency feature set. The fitting module is used to fit a three-dimensional lookup table parameter set and a primary color gain parameter set based on the display area consistency feature set and the physical color gamut degradation area set. The output module is used to perform joint spatial interpolation fusion according to the soft edge weighted coefficient map, generate a screen-level color difference compensation matrix, and write it into the display's built-in control chip through the calibration control interface.

[0084] Example 3: This example provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of any of the methods described above.

[0085] The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), registers, hard disks, optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof, or any other form of computer-readable storage medium in the art.

[0086] An exemplary storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and storage medium can reside in an application-specific integrated circuit (ASIC). In embodiments of the invention, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device.

[0087] Example 4: The present invention was applied to the intelligent color difference calibration process of a batch of high color gamut displays. The displays included OLED displays, Mini-LED displays, and high color gamut LCD displays, with screen resolutions covering three categories: 1920×1080, 2560×1440, and 3840×2160. The calibration station was set up in a closed darkroom environment with an ambient illuminance of less than 1 lx. After the displays were warmed up for 30 minutes, the calibration process began. The image acquisition device was fixed in front of the display and aligned with the center normal of the screen. The acquisition distance was maintained at 1.5 times the length of the screen diagonal. The acquisition device output a 16-bit linear RGB image. An external color analyzer was used to verify and measure the CIEXYZ tristimulus values ​​of the selected sampling points.

[0088] The monitor to be calibrated exhibits large-area chromaticity mura, local color gamut degradation, and color discontinuity at area boundaries. Specifically, the white field in the upper left area of ​​the screen is greenish, the blue channel response in the lower right edge area is attenuated, and there is a continuous slow drift in brightness and chromaticity from the center to the edge of the screen. After traditional 8×8 fixed grid partition calibration, vertical segmented color bands, stripe artifacts, and color breaks at area boundaries can still be observed in the grayscale gradient display map, skin tone gradient display map, and sky blue gradient display map.

[0089] In this embodiment, the driven display sequentially displays a standard grayscale pattern, a red primary color pattern, a green primary color pattern, a blue primary color pattern, a cyan pattern, a magenta pattern, a yellow pattern, and a medium-low saturation skin tone pattern. Each type of pattern has 17 grayscale or color levels, and a total of 136 complete screen images are acquired. The image acquisition device performs dark field subtraction, lens shading correction, and coordinate registration on each complete screen image to generate a full-screen RGB response matrix corresponding to the pixel position of the display, and compares it pixel by pixel with the target matrix to obtain a pixel-level color difference map. In a verification prototype of a 2560×1440 Mini-LED display, the average white field color difference ΔE00 before correction is 4.86, the maximum color difference ΔE00 is 10.72, the average color difference ΔE00 in the upper left area of ​​the screen is 6.18, the average color difference ΔE00 in the lower right edge area is 7.43, and the average color difference ΔE00 in the center area of ​​the screen is 3.12. This indicates that the verification prototype has both large-area low-frequency chromaticity mura and local high-frequency color gamut degradation at the edges.

[0090] The training data comes from display sample data collected from similar calibration workstations. The training set includes 680 OLED displays, 520 Mini-LED displays, and 460 high color gamut LCD displays. Each display captured 136 standard pattern images, resulting in a total of 225,760 complete screen images. Each display simultaneously saves the full-screen RGB response matrix, target matrix, pixel-level color difference map, local true colorimetric sampling results measured by a color analyzer, manually verified colorimetric Mura region annotations, and local color gamut degradation region annotations. The training, validation, and test sets are divided in a 7:2:1 ratio. The training set contains 158,032 complete screen images, the validation set contains 45,152 complete screen images, and the test set contains 22,576 complete screen images. In the training samples, the average color difference ΔE00 of the white field ranged from 1.25 to 8.92, the maximum color difference ΔE00 in the edge region ranged from 3.80 to 14.65, the local color gamut volume shrinkage rate ranged from 4.6% to 18.7%, and the proportion of color breakage samples at the fixed grid boundaries was 31.4%. During training, the improved D-LKA feature extraction network used the display color difference input features obtained by stitching together the full-screen RGB response matrix and color difference map as input, and output a D-LKA multi-scale feature set, with screen color distortion topology map, physical color gamut degradation region boundaries, and region assignment probabilities as supervision constraints. The network was trained for 80 epochs with a batch size of 8 and an initial learning rate of 0.0003. After training, the average structural similarity of the screen color distortion topology map on the validation set was 0.941, the average localization error of the physical color gamut degradation region boundaries was 2.7 pixels, and the average localization error of the boundaries of the traditional unimproved D-LKA network was 4.4 pixels.

[0091] In actual calibration, the full-screen RGB response matrix and color difference map are concatenated at the channel level and then input into the improved D-LKA feature extraction network. Shallow convolution generates a shallow feature map, and the color difference-driven bias predictor extracts the saliency features of color difference formed by the propagation of pixel-level color difference information from the shallow feature map and generates a deformable sampling bias matrix. For the aforementioned 2560×1440 Mini-LED verification prototype, the kernel size of the large-kernel dilated convolution is set to 17, and the dilation rate is set to 2. The improved D-LKA forms a continuous low-frequency response in the upper left greenish area, a local high-frequency response at the lower right blue light attenuation edge, and a smooth transition response in the gradient area from the center to the edge of the screen. In the screen color distortion topology map calculated by the D-LKA multi-scale feature set, the upper left chromaticity Mura area covers 18.6% of the screen area, the lower right local color gamut degradation area covers 9.4% of the screen area, and the top edge backlight non-uniform area covers 6.7% of the screen area. Compared with the real abnormal areas obtained by manual sampling from a color analyzer, the area overlap rate of the screen color distortion topology map is 92.8%. The traditional 8×8 fixed grid method can only roughly cover 76.5% of the abnormal areas. Moreover, the boundary of the traditional 8×8 fixed grid area is a regular rectangular boundary, which is significantly inconsistent with the boundary of the real color difference.

[0092] Adaptive clustering is performed on the screen color distortion topology map. For the verification prototype, the preset number of basic regions is 8. After calculating the topology change intensity, the number of physical color gamut degradation regions is adaptively increased to 17. In the clustering results, the upper left greenish region is divided into 3 consecutive physical color gamut degradation regions, the lower right blue channel attenuation region is divided into 4 physical color gamut degradation regions, the low-frequency gradient region in the center of the screen is divided into 5 physical color gamut degradation regions, and the remaining regions are divided into 5 low-degradation regions. The region label map records the physical color gamut degradation region number at each screen pixel position. The soft edge weighting coefficient map forms a continuous transition weight at the boundary of adjacent regions. Taking the boundary region corresponding to the local magnified image at the boundary of the region as an example, in the traditional 8×8 fixed grid correction result, there are clear vertical and horizontal boundaries between the four rectangular partitions in the magnified area, and the color transition on both sides of the boundary is abrupt. In the soft edge multi-region color gamut correction result of this invention, a continuous mixed transition is formed between adjacent physical color gamut degradation regions in the magnified area, and the boundary no longer shows a hard switch. The transition width of the boundary region is approximately 34 pixels. Within this transition range, the weight of the main region gradually decreases from 0.91 to 0.18, while the weight of the adjacent region gradually increases from 0.07 to 0.79. Compared with the hard switching from 0 to 1 in traditional fixed grid correction, it has continuous interpolation characteristics.

[0093] In the region consistency attention processing, a one-hot region coding map is constructed based on the region label map, and scale synchronization is performed according to the spatial dimensions of the residual fusion feature maps at each scale in the D-LKA multi-scale feature set. After region embedding mapping, the residual fusion feature maps at each scale are concatenated pixel-by-pixel with the corresponding scale region embedding feature maps to generate a region joint feature map. This map is then normalized using Sigmoid to generate a region consistency modulation coefficient map. This process enhances the consistency of color response variation features within the same physical color gamut degradation region and suppresses feature interference between degradation regions of different physical color gamuts. In testing, the feature variance within the same region decreased from 0.084 before modulation to 0.049, and the feature mutation index at the boundary of adjacent regions decreased from 0.137 to 0.061, indicating that the pixel-by-pixel modulation of the D-LKA multi-scale feature set by the region label map makes the region more stable and the region boundaries more continuous.

[0094] Within each physical gamut degradation region, a corresponding 3D lookup table parameter set and primary color gain parameter set are fitted using the region consistency attention feature set and the physical gamut degradation region set. For this validation prototype, the 3D lookup table nodes adopt a 17×17×17 structure. 3D lookup table parameters are fitted to each physical gamut degradation region, and the red, green, and blue primary color gains are fitted simultaneously. In the lower right blue attenuation region, the average blue primary color gain parameter is increased to 1.083, the red primary color gain parameter is 0.997, and the green primary color gain parameter is 1.011; in the upper left greenish region, the green primary color gain parameter is adjusted to 0.962, the red primary color gain parameter is 1.018, and the blue primary color gain parameter is 1.006. The 3D lookup table parameters are used to correct nonlinear gamut shifts, and the primary color gain parameters are used to correct the strength of the primary color response within the region. With their combined effect, the average residual of the region fitting samples is reduced from 1.08 when using the 3D lookup table alone to 0.63.

[0095] According to the soft edge weighted coefficient map, the three-dimensional lookup table parameter set and the primary color gain parameter set of each physical color gamut degradation region are jointly spatially interpolated and fused to generate a screen-level color difference compensation matrix. This screen-level color difference compensation matrix records the pixel-level three-dimensional lookup table compensation item and the pixel-level primary color gain compensation item corresponding to each screen pixel position, and has pixel-level continuous transition characteristics. Traditional 8×8 fixed grid correction will show direct switching of compensation matrices between adjacent regions at the grid boundary. As shown in the grayscale gradient display image, skin tone gradient display image, and sky blue gradient display image on the left, there are segmented color bands in the grayscale gradient, vertical stripes in the skin tone gradient, and blocky discontinuous transitions in the sky blue gradient. As shown in the correction result of this invention on the right, the grayscale gradient display image shows a continuous transition, the stripes in the skin tone gradient display image are significantly reduced, and no obvious abrupt change in the grid boundary appears in the sky blue gradient display image. Figure 3After correction using a traditional 8×8 fixed grid, the maximum ΔE00 at the region boundary is 4.71, the color breakage index at the region boundary is 0.132, the number of visible stripes in the skin gradient map is 7.6, and the number of visible stripes in the sky gradient map is 6.9. After correction using the method of this invention, Figure 4 The maximum ΔE00 at the boundary of the central region decreased to 2.18, the color breakage index at the boundary of the region decreased to 0.047, the number of visible stripes in the skin gradient map decreased to 1.1, and the number of visible stripes in the sky gradient map decreased to 0.8.

[0096] See Figure 5 The color difference heatmap in the image shows a clear 8×8 grid pattern with high color difference residue, with high yellow to red areas at grid intersections and edges. The color difference distribution exhibits a regular enhancement at the fixed grid boundaries. (See also...) Figure 6 The color difference heatmap corresponding to the soft-edge multi-region color gamut correction results of this invention is dominated by low-value blue areas, retaining only a few low-intensity scattered points, and does not form a regular grid-like high color difference residue. Statistical analysis of the ΔE00 curves at the region boundaries shows that the traditional 8×8 fixed grid correction results exhibit spikes at multiple boundary locations, with the highest peak reaching 4.71; the ΔE00 curve of the correction results of this invention is generally smooth, with a peak value of 2.18, and no periodic spikes corresponding to the fixed grid boundaries appear. Therefore, the difference between the traditional method and the method of this invention is not only reflected in the number of visual stripes, but also in the simultaneous decrease of the maximum ΔE00 and color breakage index at the region boundaries.

[0097] After the screen-level color difference compensation matrix is ​​generated, hardware mapping rules are generated based on the number of 3D lookup table nodes, the number of primary color gain channels, and the data bit width supported by the display's built-in control chip. The control chip of this verification prototype supports 17×17×17 3D lookup table nodes, 3 channels of primary color gain, and a 12-bit fixed-point data bit width. The system arranges the pixel-level 3D lookup table compensation items in ascending order of red, green, and blue input axes as a 3D lookup table write sequence, and arranges the pixel-level primary color gain compensation items in channel order of red, green, and blue primary color gain as a primary color gain write sequence. A color gamut correction register write data frame is generated according to the register address table. Before writing, the write status flag returned by the control chip is read; after confirming that writing is allowed, the register write is completed. After writing is completed, a readback consistency check is performed. The verification batch tested 120 monitors. The average length of the data frame written to the color gamut correction register was 37.2KB, the average writing time per monitor was 1.6 seconds, and the readback consistency verification pass rate was 100%. No data misalignment or primary color gain channel sequence errors were found.

[0098] To verify the feasibility and effectiveness of this invention, this embodiment compares the method of this invention, the traditional full-screen unified 3D-LUT correction method, and the traditional 8×8 fixed grid multi-region correction method. The comparison sample includes 120 monitors from the same batch, including 40 OLED monitors, 40 Mini-LED monitors, and 40 high color gamut LCD monitors. For each monitor, standard grayscale images, primary color images, skin tone gradient images, and sky gradient images were collected before correction, after traditional full-screen unified 3D-LUT correction, after traditional 8×8 fixed grid multi-region correction, and after correction using the method of this invention, respectively. A total of 81 sampling points were selected at the center of the screen, the four corners, the midpoint of the edge, and the high response area of ​​the color difference topology to measure ΔE00.

[0099] The comparison results show that before calibration, the average ΔE00 of the 120 monitors was 4.72, the average maximum ΔE00 was 11.38, and the average local color gamut volume shrinkage was 10.6%. After traditional full-screen unified 3D-LUT calibration, the average ΔE00 decreased to 2.14, the average maximum ΔE00 was 6.82, and the recovery rate of local color gamut volume shrinkage in the edge area was 38.5%, indicating that full-screen unified calibration is effective for large-area color casts, but insufficient compensation for local color gamut degradation in the corners and edges. After traditional 8×8 fixed grid multi-area calibration, the average ΔE00 decreased to 1.58, and the recovery rate of local color gamut volume shrinkage in the edge area was 63.2%, but the maximum ΔE00 at the area boundary was still 4.71, the color breakage index at the area boundary was 0.132, an average of 7.6 visible stripes were detected in the skin gradient image, and an average of 6.9 visible stripes were detected in the sky gradient image. After adopting the method of the present invention, the average ΔE00 is reduced to 0.86, the local color gamut volume shrinkage recovery rate in the edge region reaches 84.7%, the maximum ΔE00 at the region boundary is reduced to 2.18, the color breakage index at the region boundary is reduced to 0.047, the average number of visible stripes in the skin gradient image is reduced to 1.1, and the average number of visible stripes in the sky gradient image is reduced to 0.8.

[0100] In terms of calibration efficiency, the traditional full-screen unified 3D-LUT calibration takes an average of 42.5 seconds per unit, while the traditional 8×8 fixed grid multi-region calibration takes an average of 96.8 seconds per unit. The method of this invention, by generating a screen color distortion topology map through an improved D-LKA feature extraction network and adaptively determining the number of physical color gamut degradation regions, avoids repeated high-density measurements on all fixed grids, resulting in an average time of 68.4 seconds per unit. Although this invention takes longer than the full-screen unified calibration, the average ΔE00 is reduced by 59.8%, and the maximum ΔE00 at the region boundary is reduced from 4.71 after the traditional 8×8 fixed grid multi-region calibration to 2.18. Compared to the traditional 8×8 fixed grid multi-region calibration, this invention reduces the time per unit by approximately 29.3%, the average ΔE00 by 45.6%, the maximum ΔE00 at the region boundary by 53.7%, and the color breakage index at the region boundary by 64.4%. In the subjective review of continuous color images, 31 prototypes using the traditional 8×8 fixed grid multi-area correction method were marked as having obvious area boundary stripes, while only 4 prototypes using the method of this invention were marked as having slight transition marks, and none of them affected the normal display experience.

[0101] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A display color difference intelligent calibration method based on image recognition, characterized in that, include: The system drives the display to show a standard pattern and captures images, generates a full-screen RGB response matrix, compares it pixel by pixel with the target matrix, calculates and outputs a pixel-level color difference map. The full-screen RGB response matrix and color difference map are concatenated at the channel level and then input into the improved D-LKA feature extraction network for multi-scale feature extraction, and the output is a D-LKA multi-scale feature set. The screen color distortion topology map is calculated based on the D-LKA multi-scale feature set. Adaptive clustering is performed on the screen color distortion topology map to obtain the physical color gamut degradation region set, region label map and soft edge weighting coefficient map. Based on the region label map, the D-LKA multi-scale feature set is modulated pixel by pixel to generate the display region consistency feature set; Based on the display area consistency feature set and the physical color gamut degradation area set, a three-dimensional lookup table parameter set and a primary color gain parameter set corresponding to each physical color gamut degradation area are obtained by fitting. Based on the soft edge weighting coefficient map, the three-dimensional lookup table parameter set and the primary color gain parameter set of each physical color gamut degradation region are jointly spatially interpolated and fused to generate a screen-level color difference compensation matrix. The screen-level color difference compensation matrix is ​​written into the color gamut correction register or non-volatile memory area of ​​the monitor's built-in control chip through the calibration control interface, thus completing the intelligent color difference adjustment of the monitor.

2. The intelligent color difference adjustment method for a display based on image recognition according to claim 1, characterized in that, The driving display shows a standard pattern and acquires an image, generates a full-screen RGB response matrix, and compares it pixel-by-pixel with the target matrix, including: The display to be calibrated is driven to emit light according to a preset standard grayscale and color test pattern. The image acquisition device is used to capture the full screen image synchronously, generate a full-screen RGB response matrix, compare the full-screen RGB response matrix with the reference RGB target matrix of the same size pixel by pixel, calculate the pixel-level color difference map and output the color difference map.

3. The intelligent color difference adjustment method for displays based on image recognition according to claim 1, characterized in that, The process of concatenating the full-screen RGB response matrix and color difference map at the channel level and then inputting them into an improved D-LKA feature extraction network for multi-scale feature extraction includes: The full-screen RGB response matrix and color difference map are stitched together according to the channel dimension to generate the display color difference input features; Perform shallow convolution on the color difference input features of the display to increase the feature dimensionality and generate a shallow feature map; The shallow feature map is input into the color difference-driven bias predictor to generate a deformable sampling bias matrix; Perform large-kernel dilated convolution with a set kernel size on the shallow feature map to generate a large-kernel feature map; Based on the large kernel feature map and the deformable sampling bias matrix, deformable sampling operation is performed to generate deformable sampling features. Softmax normalization is then performed on the deformable sampling features to generate the first-scale attention feature map. Pointwise convolution is used to perform channel mapping on the first-scale attention feature map to generate a first-scale projection feature map; The first-scale projected feature map and the shallow feature map are subjected to residual fusion to generate the first-scale residual fused feature map; The residual fusion feature map of the first scale is used as the input for the next scale until feature extraction of all scales is completed, generating a D-LKA multi-scale feature set.

4. The intelligent color difference adjustment method for a display based on image recognition according to claim 1, characterized in that, The calculation of the screen color distortion topology map based on the D-LKA multi-scale feature set includes: Scale alignment and channel mapping are performed on the residual fusion feature maps of each scale in the D-LKA multi-scale feature set to generate a unified scale distortion feature map. Scale weight fusion is performed on the uniform-scale distortion feature map to generate a screen color distortion topology map; Based on the screen color distortion topology map, construct the clustering feature vector corresponding to each screen pixel position; Based on the screen color distortion topology map, determine the number of physical color gamut degradation regions for adaptive clustering; Adaptive clustering is performed on the screen color distortion topology map based on the clustering feature vector and the number of physical color gamut degradation regions to generate the region belonging probability corresponding to each screen pixel position; Based on the region attribution probability, generate a set of physical color gamut degradation regions and a region label map; A soft-edge weighted coefficient map is generated based on the region affiliation probability and the region label map.

5. The intelligent color difference adjustment method for a display based on image recognition according to claim 1, characterized in that, The step of pixel-by-pixel modulation of the D-LKA multi-scale feature set based on the region label map includes: Construct a unique-hot encoding map of regions based on the region label map; Based on the spatial dimensions of the residual fusion feature maps at each scale in the D-LKA multi-scale feature set, scale synchronization is performed on the regional one-hot coding maps to generate regional coding maps at each scale. Perform region embedding mapping on the region encoding maps at each scale to generate region embedding feature maps at each scale; The residual fusion feature maps at each scale are connected pixel-by-pixel with the region embedding feature maps at the corresponding scale to generate joint feature maps of regions at each scale. Based on the joint feature map of each scale region, generate a consistent modulation coefficient map of each scale region. Based on the consistent modulation coefficient map of each scale region, pixel-wise modulation is performed on the residual fusion feature map of each scale in the D-LKA multi-scale feature set to generate consistent attention feature maps of each scale region. Summarize the region consistency attention feature maps corresponding to all scales to generate a display region consistency feature set.

6. The intelligent color difference adjustment method for a display based on image recognition according to claim 1, characterized in that, The method, based on the display area consistency feature set and the physical color gamut degradation region set, fits and obtains the three-dimensional lookup table parameter set and primary color gain parameter set corresponding to each physical color gamut degradation region, including: Scale alignment and scale fusion are performed on the display area consistency feature set to generate a screen-level area consistency feature map. Extract the region consistency description vector corresponding to each physical gamut degradation region from the screen-level region consistency feature map; Based on the physical color gamut degradation region, the full-screen RGB response matrix, and the target matrix, a set of regional color fitting samples corresponding to each physical color gamut degradation region is constructed. Based on the region consistency description vector, generate the initial parameters of the three-dimensional lookup table and the initial parameters of the primary color gain for each physical color gamut degradation region; Based on the regional color fitting sample set and the initial parameters of the three-dimensional lookup table, fit the parameters of the three-dimensional lookup table corresponding to each physical color gamut degradation region. Based on the regional color fitting sample set, the three-dimensional lookup table parameters and the initial parameters of the primary color gain, the primary color gain parameters corresponding to each physical color gamut degradation region are fitted. Summarize the three-dimensional lookup table parameters and primary color gain parameters corresponding to all physical color gamut degradation regions, and generate a set of three-dimensional lookup table parameters and a set of primary color gain parameters.

7. The intelligent color difference adjustment method for a display based on image recognition according to claim 1, characterized in that, The step of performing joint spatial interpolation fusion of the three-dimensional lookup table parameter set and the primary color gain parameter set for each physical color gamut degradation region according to the soft edge weighting coefficient map includes: The three-dimensional lookup table parameter set is spatially interpolated and fused pixel by pixel according to the soft edge weighting coefficient map to generate pixel-level three-dimensional lookup table compensation terms corresponding to each screen pixel position. The set of primary color gain parameters is spatially interpolated and fused pixel by pixel according to the soft edge weighting coefficient map to generate a pixel-level primary color gain compensation term corresponding to each screen pixel position; The pixel-level 3D lookup table compensation item and the pixel-level primary color gain compensation item are jointly encapsulated to generate the pixel-level joint compensation item corresponding to each screen pixel position. A color difference compensation mapping relationship is established for each screen pixel position based on the pixel-level joint compensation term; The pixel-level joint compensation terms are arranged according to the positions of all screen pixels in the complete screen image of the display to be calibrated, and a screen-level color difference compensation matrix is ​​generated.

8. The intelligent color difference adjustment method for a display based on image recognition according to claim 1, characterized in that, The process of generating the data frame by writing the color gamut correction register or the data frame by writing the non-volatile memory area includes: Extract the pixel-level 3D lookup table compensation item and pixel-level primary color gain compensation item corresponding to each screen pixel position from the screen-level color difference compensation matrix. Based on the number of 3D lookup table nodes, the number of primary color gain channels and the data bit width supported by the display's built-in control chip, define the hardware mapping rules of the screen-level color difference compensation matrix. According to the hardware mapping rules, the three-dimensional lookup table writing sequence and the primary color gain writing sequence are subjected to fixed-point quantization processing to generate fixed-point three-dimensional lookup table writing data and fixed-point primary color gain writing data. When the display's built-in control chip uses a color gamut correction register to store correction parameters, the data written to the fixed-point three-dimensional lookup table and the data written to the fixed-point primary color gain are arranged according to the color gamut correction register address table to generate a color gamut correction register write data frame. When the display's built-in control chip uses a non-volatile memory area to store correction parameters, the fixed-point three-dimensional lookup table write data and the fixed-point primary color gain write data are arranged according to the non-volatile memory area page address, block address, and offset address to generate a non-volatile memory area write data frame. The system sends a write enable command to the display's built-in control chip through the calibration control interface, reads the write status flag returned by the display's built-in control chip, and when the write status flag indicates that writing is allowed, writes the color gamut correction register to the color gamut correction register using a write data frame, or writes the non-volatile memory area to the non-volatile memory area using a write data frame. After writing is complete, the readback correction data in the built-in control chip of the display is read through the correction control interface. The readback correction data is then compared with the data frame written to the color gamut correction register or the data frame written to the non-volatile memory area for consistency verification. If the consistency verification passes, a correction write completion flag is generated. Based on the calibration completion flag, the monitor's built-in control chip calls the written 3D lookup table compensation items and primary color gain compensation items to perform color difference compensation processing on the monitor's subsequent display input signals, thus completing the monitor's intelligent color difference calibration.

9. A display color difference intelligent calibration system based on image recognition, used to execute the display color difference intelligent calibration method based on image recognition as described in any one of claims 1-8, characterized in that, include: The acquisition module is used to drive the display to show a standard pattern and acquire images, generate a full-screen RGB response matrix, compare it with the target matrix pixel by pixel, calculate and output a pixel-level color difference map; The extraction module is used to input the full-screen RGB response matrix and color difference map into the improved D-LKA feature extraction network after channel-level concatenation, and output the D-LKA multi-scale feature set. The clustering module is used to calculate the screen color distortion topology map and perform adaptive clustering to obtain a set of physical color gamut degradation regions, a region label map, and a soft edge weighting coefficient map. The modulation module is used to modulate the D-LKA multi-scale feature set pixel by pixel according to the region label map to generate a display region consistency feature set. The fitting module is used to fit a three-dimensional lookup table parameter set and a primary color gain parameter set based on the display area consistency feature set and the physical color gamut degradation area set. The output module is used to perform joint spatial interpolation fusion according to the soft edge weighted coefficient map, generate a screen-level color difference compensation matrix, and write it into the display's built-in control chip through the calibration control interface.

Images