Liquid crystal screen hdr color dynamic enhancement method and system based on image recognition

Abstract

The application discloses an image recognition-based liquid crystal screen HDR color dynamic enhancement method and system, and aims at solving the problem of edge light emission halo caused by mutual interference of backlight partition and pixel enhancement. The application generates a mask, a boundary and a confidence degree through semantic segmentation, performs partition weighting aggregation, models spatial variable anisotropic PSF light leakage based on panel state driving, generates a risk heat map based on boundary coupling attention, performs double-layer buffer band layered limiting and projects to a boundary safety set through a differentiable joint optimization, controls backlight and pixels under a unified timing, and realizes the technical effects of improving HDR brightness and color, suppressing halo, maintaining boundary stability and readability, and enhancing partition continuity.

Description

Technical Field

[0001] This invention relates to the field of HDR image processing for liquid crystal displays, and more particularly to a method and system for dynamic HDR color enhancement of a liquid crystal screen based on image recognition. Background Technology

[0002] With the increasing prevalence of HDR in LCD monitors, the industry commonly employs local dimming (FALD, Mini-LED) and pixel-level tone mapping to collaboratively enhance contrast and color performance. Existing solutions often generate local dimming settings based on image brightness statistics (mean, peak, histogram) and approximate panel light leakage using fixed or isotropic kernels. Edge areas are typically handled using heuristic strategies such as threshold limiting and edge detection suppression. Backlight control and pixel rendering are mostly executed separately in independent modules and timing sequences, lacking end-to-end modeling and joint optimization for edge scenes.

[0003] However, existing technologies still have shortcomings:

[0004] 1. The lack of joint optimization between backlight zoning and pixel enhancement makes it easy for mutual interference to occur at high contrast boundaries, resulting in cross-boundary brightness overshoot and halo.

[0005] 2. Panel light leakage models generally use static, spatially invariant, and isotropic point diffusion functions, without considering partition location, driving current, temperature, aging, and other conditions, resulting in large light leakage prediction deviations and insufficient edge suppression.

[0006] 3. Boundary risk assessment relies heavily on simple gradient or edge detection, lacks coupling with semantic boundaries and directional consistency, and does not have confidence adjustment. Furthermore, it does not have a layered buffer, which impairs the readability and hue stability of sensitive categories such as subtitles and faces.

[0007] Therefore, a method and system for dynamic color enhancement of LCD screens with HDR that can overcome the shortcomings of the prior art is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0008] One objective of this invention is to propose a dynamic color enhancement method for HDR LCD screens based on image recognition. Addressing the problems in existing technologies such as lack of joint optimization between backlight partitioning and pixel enhancement, static isotropic light leakage models, and coarse edge risk assessment leading to edge halos and hue shift, this invention proposes a technical solution involving semantic segmentation and linearized input, panel state-driven spatially variable anisotropic PSF light leakage modeling, PSF boundary-coupled attention generation of halo risk heatmaps, layered limiting with a double-layer buffer along the diffusion axis, differentiable joint optimization and iterative projection to a boundary safety set defined by cross-boundary brightness gradients and hue shift thresholds, and unified temporal joint control of backlight and pixels. This invention achieves the technical effects of enhancing HDR brightness and color while suppressing halos, reducing cross-boundary overshoot and hue shift, protecting the readability of sensitive categories such as subtitles and faces, maintaining partitioning continuity and control stability, and exhibiting adaptive robustness to panel state changes.

[0009] A method for dynamic color enhancement of a liquid crystal display screen based on image recognition according to an embodiment of the present invention is characterized by comprising:

[0010] S1. Perform semantic segmentation and linearization on the input image frame to generate a semantic mask map, boundary map, confidence map, and brightness map;

[0011] S2. The brightness map is weighted and aggregated according to the backlight partition mapping table, and combined with the semantic mask map, boundary map and confidence map for weighting to generate backlight partition brightness candidate map;

[0012] S3. Based on the candidate image of backlight zone brightness and panel state parameters, conditional modeling is performed to generate a spatially variable, anisotropic point spread function kernel map.

[0013] S4. Convolve and superimpose the candidate backlight partition brightness maps using the point spread function kernel map to generate a light leakage distribution map;

[0014] S5. Based on the leakage distribution map, boundary map, confidence map and semantic mask map, a halo risk heat map is generated by weighted fusion through a boundary coupling attention mechanism.

[0015] S6. Generate a double-layer buffer zone map based on the halo risk heat map, wherein the inner ring buffer zone is used to limit the target brightness of the backlight zone, and the outer ring buffer zone is used to limit the pixel-end tone enhancement gain.

[0016] S7. Using the semantic mask map as a condition, construct a differentiable joint optimization objective function based on the candidate backlight partition brightness map, halo risk heat map and double-layer buffer zone map, and jointly solve the target brightness of the backlight partition and the pixel tone mapping parameters to obtain the initial backlight partition target brightness map and the initial pixel tone mapping parameter map.

[0017] S8. Under the constraints of the halo risk heat map and the double-layer buffer zone map, perform differentiable joint optimization iteration on the initial backlight partition target brightness map and the initial pixel end tone mapping parameter map, and adaptively adjust the constraint strength according to the confidence map. After each iteration, perform differentiable constraint projection to the boundary safety set defined by the cross-boundary brightness gradient and the boundary hue offset threshold to obtain the final backlight partition target brightness map and the final pixel end tone mapping parameter map.

[0018] S9. Based on the final backlight partition target brightness map and the final pixel end tone mapping parameter map, generate backlight partition control instructions and pixel rendering parameters, and send them to the backlight driver and pixel rendering engine under unified timing conditions to achieve HDR color dynamic enhancement and suppress edge halo on the input image frame.

[0019] Optionally, step S1 specifically includes:

[0020] The input image frame is normalized in color space and subjected to inverse electro-optical transfer function transformation to convert the pixels of the input image frame into a linear luminance representation to generate a luminance map;

[0021] Semantic segmentation is performed on the input image frame to generate a semantic mask map, which is a pixel-level label map that assigns a category label to each pixel according to a predefined set of categories;

[0022] A confidence map is generated based on the pixel category probabilities obtained from semantic segmentation. The confidence map is the maximum category probability or equivalent confidence value of each pixel.

[0023] A boundary map is generated based on the category boundary of the semantic mask map and the gradient response of the brightness map. The boundary map is a pixel-by-pixel boundary probability map used to identify the boundary position of the semantic object.

[0024] Output semantic mask image, boundary image, confidence image and brightness image.

[0025] Optionally, step S2 specifically includes:

[0026] Assign a backlight zone number to each pixel of the brightness map based on the backlight zone mapping table;

[0027] A semantic weight is assigned to each pixel based on the semantic mask image. The semantic weight is set according to a predefined category priority to improve readability protection in categories such as subtitles, faces and text interfaces.

[0028] Based on the boundary map, a boundary proximity attenuation weight is applied to pixels located in the region near the boundary of a semantic object to reduce the cross-boundary brightness contribution.

[0029] Pixels with confidence scores below a threshold are weighted based on the confidence map.

[0030] The semantic weights, boundary proximity attenuation weights, and confidence weights are multiplied to obtain pixel weighting coefficients, and the brightness map is weighted and partitioned to obtain the weighted average brightness value of each backlight partition.

[0031] A robust limiting is performed on the weighted average brightness value of each backlight zone. The robust limiting is set based on the high quantile threshold and low quantile threshold of the zone to suppress the influence of abnormally bright or abnormally dark pixels on the zone result.

[0032] Generate candidate images of backlight zone brightness.

[0033] Terminology definition:

[0034] The brightness map is a two-dimensional matrix of linear brightness values ​​obtained by normalizing the input image frame in color space and inversely transforming it with the electro-optical transfer function. Its resolution is consistent with the pixel grid of the panel.

[0035] The semantic mask image is a graph that provides a semantic category label or category probability for each pixel, and is used to provide category information;

[0036] The boundary map is a pixel-wise probability or intensity map that identifies the boundary location of a semantic object and is used to identify the region adjacent to the boundary.

[0037] The confidence graph is a graph of the maximum class probability or equivalent confidence measure of semantic segmentation at each pixel;

[0038] The backlight partition mapping table is a preset correspondence that maps panel pixel coordinates to backlight partition numbers, and can be a lookup table or an equivalent function;

[0039] The backlight zone number is an index identifier used to distinguish different backlight zones;

[0040] The weighted partition aggregation is a calculation that sums and normalizes the pixel brightness of each backlight partition according to a given weight to obtain the partition brightness statistics.

[0041] The predefined category priority is a relative importance order or weight coefficient pre-set for different semantic categories;

[0042] The cross-boundary brightness contribution is the amount of influence of pixels located in the region near the object boundary on the brightness aggregation result of adjacent semantic regions or adjacent backlight partitions.

[0043] The confidence threshold is a numerical threshold used to distinguish between high-confidence and low-confidence pixels;

[0044] The weighted average brightness value is a brightness statistic obtained by averaging the brightness map according to pixel weights within a backlight zone.

[0045] The high quantile threshold is a value corresponding to the upper quantile selected based on the pixel brightness distribution within the partition, used to determine the upper limit of the amplitude limit;

[0046] The lower quantile threshold is a value corresponding to the lower quantile selected based on the pixel brightness distribution within the partition, used to determine the lower limit of the amplitude limiting.

[0047] The backlight partition brightness candidate map is a set representation composed of candidate brightness values ​​of each backlight partition, which can be represented in the partition index space or mapped to the pixel grid for subsequent processing.

[0048] Optionally, step S3 specifically includes:

[0049] Based on the candidate backlight brightness map, calculate the brightness statistics of each backlight zone, including weighted average brightness, peak brightness, and brightness gradient within the zone;

[0050] A partition condition vector is constructed based on the partition location, adjacency relationship, drive current, temperature, and aging parameters in the panel status parameters;

[0051] The point spread function parameter set for the backlight zone is generated by using the partition brightness statistical characteristics and partition condition vector through a parameter mapping function, including the spread radius, directionality coefficient and energy attenuation coefficient.

[0052] The point spread function parameters are weighted and combined using a basic kernel set to generate an anisotropic point spread function kernel;

[0053] Energy normalization and nonnegativity constraints are applied to the generated anisotropic point spread function kernel, and continuity constraints are applied to the point spread function parameters between adjacent backlight partitions to avoid abrupt changes across partitions.

[0054] Boundary clipping is applied to the point spread function kernel of the backlight zone located at the edge of the panel to suppress light leakage to the outside of the panel;

[0055] The point spread function kernels of all backlight zones are organized into a spatially variable point spread function kernel diagram according to the zone number;

[0056] Output spatially variable point diffusion function kernel diagram.

[0057] Terminology definition:

[0058] The panel state parameters are panel operating and optical state quantities acquired or estimated in the current or recent frame period, including the partition position, adjacency relationship, driving current, panel temperature and aging parameters of each backlight partition, which are used as conditional inputs for generating point spread function parameters.

[0059] The conditional modeling is a process of taking a conditional vector composed of backlight partition brightness candidate images and panel state parameters as input, generating point spread function parameters through a parameter mapping function, and constructing a point spread function kernel accordingly.

[0060] The point spread function kernel is a discrete two-dimensional weighted matrix that represents the light leakage from a single backlight zone to each pixel in the panel pixel coordinate system. Its elements are non-negative and reflect the spatial distribution of the leaked energy.

[0061] The dot spread function kernel map is a spatial mapping of a set of dot spread function kernels organized by backlight partitions, recording the kernel corresponding to each partition so that the kernel parameters at different locations can be different;

[0062] The space can be transformed into the property that the parameters of the point spread function kernel change spatially with the partition position or panel state;

[0063] The anisotropy is the property that the point spread function kernel has different diffusion scales in different directions, and is usually characterized by the ratio of the diffusion radius of the principal axis to that of the secondary axis.

[0064] The partition brightness statistics feature is a set of brightness statistics calculated from the backlight partition brightness candidate image in each partition, which is used to drive the generation of point spread function parameters;

[0065] The brightness peak value is the maximum linear brightness of a pixel within a certain backlight zone, used to characterize the brightness of that zone.

[0066] The brightness gradient within a partition is a statistical analysis of the gradient magnitude formed by the first-order difference of the brightness map within that partition in spatial coordinates, used to reflect the intensity and direction of brightness changes.

[0067] The partition position refers to the geometric position parameters of the backlight partition in the panel coordinate system, including the partition center coordinates and the partition shape range;

[0068] The adjacency relationship is the topological adjacency information between backlight partitions, which is used to constrain the continuity of the point spread function parameters of adjacent partitions and evaluate their mutual influence.

[0069] The driving current is the set or measured value of the driving current of the backlight partition in the current control cycle, which is used to affect the energy and diffusion parameters of the point spread function.

[0070] The temperature is a measured or estimated value of the operating temperature of the panel or backlight module, used to reflect the light leakage characteristics as temperature changes.

[0071] The aging parameter is a performance characterization parameter caused by the service life and optical component degradation, and is used to correct the change of point spread function over time.

[0072] The partitioning condition vector is a numerical vector composed of partitioning position, adjacency relationship, driving current, temperature and aging parameters in a fixed order or structure, which serves as the input to the parameter mapping function.

[0073] The parameter mapping function is a deterministic function or model that maps the zoning brightness statistical features and zoning condition vectors into a set of point spread function parameters. It can be a linear / nonlinear regression, a neural network, or a lookup table.

[0074] The point spread function parameter set is a set of parameters used to construct the point spread function kernel, including the diffusion radius, directionality coefficient, and energy decay coefficient;

[0075] The diffusion radius is the diffusion scale parameter of the point diffusion function kernel in a given direction, and the unit is a pixel or equivalent spatial unit;

[0076] The directionality coefficient is a parameter describing the degree of nuclear anisotropy, and is usually defined as the ratio of the primary axis diffusion radius to the secondary axis diffusion radius or an equivalent index.

[0077] The energy attenuation coefficient is an intensity parameter describing the decrease of light leakage with distance, used to control the radial or axial attenuation rate of the kernel weight.

[0078] The basic kernel set is a predefined, weighted, combinable kernel base dictionary, such as anisotropic Gaussian kernels, exponential decay kernels, etc., used to construct the target point diffusion function kernel;

[0079] The energy normalization is a process of normalizing the point spread function kernel so that the sum of its weights is 1 or a preset energy constant, in order to ensure energy conservation.

[0080] The nonnegativity constraint is a constraint that ensures all weights of the point spread function kernel are not less than zero;

[0081] The continuity constraint is a constraint that limits the variation of the point spread function parameter between adjacent backlight zones in order to avoid abrupt changes across zones;

[0082] The panel edge is the boundary range of the panel pixel grid, used to define the effective calculation area of ​​the kernel;

[0083] The boundary clipping process involves clipping the kernel weights of the point diffusion function located outside the panel edge to zero to prevent light leakage calculations from exceeding the limits.

[0084] The spatially variable point spread function kernel map is a set of point spread function kernels for all backlight zones organized by zone number and represented on the panel coordinate system, making the kernels spatially variable and anisotropic.

[0085] Optionally, step S4 specifically includes:

[0086] Anisotropic convolution is performed on the backlight partition brightness candidate map on the panel pixel grid. For each backlight partition, the corresponding point spread function kernel is selected according to the partition number of the backlight partition in the spatially variable point spread function kernel map, and the candidate brightness value of the backlight partition is used as the amplitude of the convolution signal.

[0087] The convolution results of each backlight zone are superimposed pixel by pixel on the panel pixel grid to form a light leakage distribution map;

[0088] During convolution and stacking, the energy normalization, non-negativity constraint and boundary clipping implemented in the spatially variable point diffusion function kernel map are followed, so that the value of the light leakage distribution map is non-negative and does not include pixels outside the panel.

[0089] The light leakage distribution map is numerically limited to match the dynamic range of the brightness map;

[0090] Generate a light leakage distribution map.

[0091] Terminology definition:

[0092] The backlight partition is an independently driveable and controllable light-emitting area unit on the backlight module, and its geometric range can be mapped to a group of panel pixels in the panel coordinate system.

[0093] The panel pixel grid is a two-dimensional discrete sampling grid with the actual pixel arrangement of the panel as the coordinate system, and the coordinate unit is pixels. It is used to carry the computational domain of image and kernel operations.

[0094] The anisotropic convolution is a discrete convolution operation performed on the panel pixel grid using a direction-dependent point spread function kernel, and the kernel has different spread scales and shapes in different directions.

[0095] The candidate brightness value is the scalar brightness value corresponding to each backlight partition in the backlight partition brightness candidate image, which is used to represent the signal intensity to be convolved in that partition.

[0096] The amplitude of the convolution signal is a scalar coefficient used to scale the output of the point spread function kernel, taken from the candidate brightness value of the corresponding backlight zone.

[0097] The pixel-by-pixel superposition is an operation that sums the convolution results from each backlight zone on the panel pixel grid according to the same pixel coordinates;

[0098] The light leakage distribution map is a two-dimensional map on the panel pixel grid representing the light leakage intensity generated at the full-screen pixels after each backlight zone is convolved by a point spread function and superimposed.

[0099] The pixels outside the panel are coordinate points located outside the panel pixel grid boundary and do not participate in convolution, superposition, and display calculations;

[0100] The numerical range is defined as the processing of cropping and / or linearly scaling the values ​​of the light leakage distribution map so that its upper and lower limits are consistent with the representation range of the brightness map.

[0101] The dynamic range is the range of minimum and maximum brightness values ​​that the brightness map can represent under given positioning depth and linearization conditions.

[0102] Optionally, step S5 specifically includes:

[0103] Boundary coupling attention weights are constructed based on the light leakage distribution map, boundary map, confidence map, and semantic mask map. The boundary coupling attention weights are obtained by combining the following weights at each pixel: boundary proximity weights are generated based on the distance transformation of the boundary map, and pixels closer to the boundary are given higher weights using a distance decreasing function.

[0104] Boundary direction weights are generated based on the consistency of the gradient vectors of the boundary map and the gradient vectors of the light leakage distribution map. The weights are increased when the main light leakage diffusion direction is consistent with the boundary normal direction.

[0105] Confidence weights are generated based on the confidence map, and the weights of pixels with confidence scores below a threshold are reduced.

[0106] Generate category priority weights based on the semantic mask image, giving higher weights to the edges of predefined categories;

[0107] The above weights are normalized and combined to form a boundary coupling attention weight graph;

[0108] The leakage distribution map is weighted and fused pixel by pixel using the boundary coupling attention weight map to generate an initial risk map;

[0109] Noise suppression and numerical range normalization are performed on the initial risk map to improve continuity and stability;

[0110] Generate and output a halo risk heatmap.

[0111] Terminology definition:

[0112] The boundary coupling attention mechanism is a process that combines the boundary map, the light leakage distribution map, the semantic mask map and the confidence map to calculate attention weights and use them for risk measurement, so as to highlight the pixel contributions related to the halo at the boundary.

[0113] The boundary coupling attention weight is a scalar obtained by normalizing and combining multiple sub-weights at each pixel, used to measure the halo correlation of the pixel;

[0114] The boundary coupling attention weight map is a two-dimensional scalar map composed of boundary coupling attention weights on the panel pixel grid;

[0115] The distance transformation is an operation that calculates the distance from each pixel to the nearest boundary position based on the boundary map. The distance metric can be Euclidean, Manhattan, or an equivalent metric.

[0116] The boundary proximity weight is a weight obtained by mapping the distance from the pixel to the boundary using a distance-decreasing function; the closer the distance, the greater the weight.

[0117] The distance decreasing function is a function that monotonically and non-increasingly maps the distance value to [0,1] or an equivalent range, and can take the form of exponential, reciprocal, piecewise linear or equivalent.

[0118] The gradient vector is a vector of local change direction and magnitude obtained by calculating the first spatial derivative of the image or probability map on the pixel grid. The calculation operator can be Sobel, Prewitt, Scharr or an equivalent method.

[0119] The boundary normal direction is the unit direction of the boundary map gradient vector, indicating the normal direction in which the boundary strength increases;

[0120] The main diffusion direction of light leakage is the unit vector of the main change direction of the light leakage distribution map in a local area, which can be approximated by the gradient vector or the direction of its principal components of the light leakage distribution map.

[0121] The directional consistency is a measure of the directional similarity between the boundary normal direction and the main diffusion direction of light leakage, which can be represented by the normalized vector dot product, the cosine of the included angle, or an equivalent function.

[0122] The boundary direction weight is a weight that maps the direction similarity to [0,1] or an equivalent range based on the direction consistency. The more consistent the directions are, the greater the weight.

[0123] The confidence weight is an adjustment weight assigned to a pixel based on the confidence map, with the weight reduced for low-confidence pixels and the weight maintained or increased for high-confidence pixels.

[0124] The category priority weight is a mapping based on the semantic mask image to increase the weight of predefined high-priority categories (such as subtitles, faces, text interfaces, etc.) at their edges;

[0125] The normalization combination is the process of obtaining a single weight by scaling and normalizing multiple sub-weights and then following a set rule (such as weighted summation, product, or Softmax normalization).

[0126] The pixel-by-pixel weighted fusion is a process of multiplying and aggregating the light leakage distribution map at the corresponding pixel position with boundary coupling attention weights to form a risk measure;

[0127] The initial risk map is a two-dimensional halo risk intensity map obtained by pixel-by-pixel weighted fusion, which has not yet been denoised and range-normalized.

[0128] The noise suppression is a denoising and smoothing process applied to the initial risk map to improve spatial continuity and stability. Low-pass filtering, bilateral filtering, morphological opening and closing, or equivalent methods can be used.

[0129] The numerical range normalization is a scaling transformation process that maps the risk map to a preset range (such as [0,1] or target depth).

[0130] The halo risk heatmap is a two-dimensional intensity map output after noise suppression and numerical range normalization. It is used to characterize the risk intensity of halo artifacts in each pixel and is used by subsequent modules.

[0131] Optionally, step S6 specifically includes:

[0132] The pixel values ​​of the halo risk heatmap are used as the risk intensity, and the ratio of the diffusion principal axis direction to the anisotropy is calculated on the panel pixel grid based on the spatially variable point diffusion function kernel map.

[0133] Anisotropic expansion is performed on the perimeter of the high-value region of risk intensity according to the diffusion principal axis direction and anisotropy ratio to generate an inner ring buffer band mask. The bandwidth of the inner ring buffer band mask is adaptively determined according to a preset function of risk intensity and anisotropy ratio, and is clipped at the edge of the panel to avoid exceeding the boundary.

[0134] Anisotropic expansion is performed along the diffusion principal axis on the outer side of the inner ring buffer band mask, and a non-overlapping constraint is applied to generate an outer ring buffer band mask. The bandwidth of the outer ring buffer band mask is adaptively determined based on a risk intensity decreasing function and a preset function of anisotropy ratio.

[0135] Based on the risk intensity of the halo risk heat map, the backlight zone target brightness limiting coefficient map is generated using the inner ring buffer band mask, so that the limiting coefficient of the backlight zone target brightness tightens as the risk intensity increases;

[0136] Based on the risk intensity of the halo risk heatmap, a pixel-end tone enhancement gain limiting coefficient map is generated using the outer ring buffer band mask, so that the limiting coefficient of pixel-end tone enhancement gain tightens as the risk intensity increases;

[0137] The inner ring buffer band mask and its backlight zone target brightness limiting coefficient map are combined with the outer ring buffer band mask and its pixel end tone enhancement gain limiting coefficient map to form a double-layer buffer band map.

[0138] Terminology definition:

[0139] The risk intensity is a scalar value of the halo risk heatmap at each pixel, used to measure the relative risk of halo artifacts occurring in that pixel;

[0140] The diffusion principal axis direction is a unit direction determined based on the local principal axis orientation of the spatially variable point diffusion function kernel, indicating the main direction of light leakage diffusion;

[0141] The anisotropy ratio is the ratio of the diffusion scale of the principal axis to the secondary axis of the point diffusion function kernel, which is used to quantify the degree of diffusion imbalance of the kernel in different directions.

[0142] The anisotropic expansion is a morphological expansion operation performed on the target region by a directional structural element along the diffusion principal axis and set according to the anisotropic ratio.

[0143] The inner ring buffer band mask is a first-layer strip binary mask generated around the perimeter of the high-risk intensity region, used to identify the spatial range in which the brightness of the target backlight zone needs to be limited.

[0144] The outer ring buffer band mask is a second-layer strip binary mask located outside the inner ring buffer band mask and not overlapping with it, used to identify the spatial range in which the pixel-end tone enhancement gain needs to be limited.

[0145] The bandwidth is the pixel thickness or equivalent spatial width of the buffer band mask in the normal direction, and is adaptively determined by a preset function of risk intensity and anisotropy ratio.

[0146] The non-overlapping constraint is a mutual exclusion condition applied to the inner ring buffer band mask and the outer ring buffer band mask, ensuring that their spatial intersection is an empty set;

[0147] The target brightness of the backlight zones is the target brightness value to be set for each backlight zone, which is used as the basic parameter for joint optimization and drive setting.

[0148] The backlight zone target brightness limiting coefficient map is a coefficient mapping defined on the panel pixel grid, used to limit the allowable upper limit of the backlight zone target brightness associated with the pixel, and the coefficient decreases as the risk intensity increases;

[0149] The pixel-side tone enhancement gain is a set of gain parameters used on the pixel rendering side to improve visual effects, including at least brightness gain and saturation gain and their combined effects;

[0150] The pixel-end tone enhancement gain limiting coefficient map is a coefficient mapping defined on the panel pixel grid, used to limit the allowable upper limit of pixel-end tone enhancement gain, and the coefficient decreases as the risk intensity increases;

[0151] The dual-layer buffer band diagram is a composite representation that combines the inner ring buffer band mask and its backlight zone target brightness limiting coefficient diagram with the outer ring buffer band mask and its pixel-end tone enhancement gain limiting coefficient diagram, and is used for unified reference in subsequent optimization and control.

[0152] Optionally, step S7 specifically includes:

[0153] The optimization variables for joint optimization are set as the target brightness of the backlight zone and the pixel-end tone mapping parameters, where the pixel-end tone mapping parameters include brightness gain parameters, saturation gain parameters and hue shift parameters;

[0154] A brightness fidelity term is constructed based on the backlight partition brightness candidate map, so that the target brightness of the backlight partition approximates the backlight partition brightness candidate map in each backlight partition, and the fidelity of the predefined category is improved by using the category priority of the semantic mask map as the weight.

[0155] A halo suppression penalty term is constructed based on the halo risk heat map and the double-layer buffer zone map. In the inner ring buffer zone, a barrier function for the cross-boundary brightness gradient is set, and in the outer ring buffer zone, a barrier function for the pixel-end tone mapping parameter gain is set, so that the penalty coefficient is larger at pixels with higher risk intensity.

[0156] Based on the target brightness limiting coefficient map of the backlight zone and the color enhancement gain limiting coefficient map of the pixel end in the double-layer buffer zone map, a limiting constraint is set so that the target brightness of the backlight zone and the color mapping parameters of the pixel end do not exceed the corresponding limiting upper limit within their respective buffer zones.

[0157] A spatial smoothing regularization term is constructed based on the adjacency relationship between backlight zones to constrain the target brightness difference between adjacent backlight zones to maintain continuity.

[0158] The brightness fidelity term, halo suppression penalty term, amplitude limiting constraint and spatial smoothing regularization term are combined into a differentiable joint optimization objective function, and the initial solution is generated by the joint solution of weighted least squares and barrier function.

[0159] Based on the amplitude limiting constraint in the inner ring buffer zone, the candidate brightness map of the backlight partition is risk-weighted and shrunken to generate the initial target brightness map of the backlight partition.

[0160] Based on the amplitude limiting constraints in the outer ring buffer band and combined with the category priority of the semantic mask map, the pixel-end tone mapping parameters are risk-weighted to generate an initial pixel-end tone mapping parameter map.

[0161] Output the initial backlight zone target brightness map and the initial pixel-level tone mapping parameter map.

[0162] Terminology definition:

[0163] The joint optimization is a process of simultaneously solving for the target brightness of the backlight zone and the pixel-level tone mapping parameters under the same objective function;

[0164] The optimization variables are the set of parameters that need to be estimated in the joint optimization, including the target brightness of the backlight partition and the pixel-end tone mapping parameters;

[0165] The pixel-side tone mapping parameters are a set of parameters used to adjust the display effect on the pixel side, including brightness gain parameters, saturation gain parameters, and hue shift parameters;

[0166] The brightness gain parameter is a multiplicative coefficient that acts on the linear brightness of the pixel, and is used to increase or decrease the pixel brightness.

[0167] The saturation gain parameter is a multiplicative coefficient that acts on the pixel color saturation, used to enhance or reduce color purity.

[0168] The hue offset parameter is an additive offset that acts on the pixel hue angle and is used to fine-tune the dominant hue.

[0169] The brightness fidelity term is a cost term that measures the deviation between the target brightness of the backlight partition and the candidate brightness map of the backlight partition, and is weighted by the category priority of the semantic mask map to improve the fidelity of the predefined categories.

[0170] The halo suppression penalty term is a cost term constructed based on the halo risk heat map and the double-layer buffer zone map, which is used to suppress cross-boundary brightness overshoot and hue drift in high-risk areas.

[0171] The cross-boundary brightness gradient is a measure of the rate of change of brightness calculated along the boundary normal direction at the semantic boundary, used to characterize the degree of brightness step or overshoot across the boundary.

[0172] The barrier function is a penalty function that tends to infinity when the parameter or gradient approaches a set threshold, used to prevent the optimization variable from going out of bounds;

[0173] The pixel-end tone mapping parameter gain is a gain measure of the pixel-end tone mapping parameter relative to the baseline setting, including luminance gain, saturation gain and their combined intensity;

[0174] The limiting constraint is an upper limit imposed on the target brightness of the backlight zone and the pixel-end tone mapping parameters based on the limiting coefficient of the double-layer buffer band map, so that it does not exceed the allowable range;

[0175] The spatial smoothing regularization term is a term that penalizes the difference in target brightness between adjacent backlight zones, in order to constrain spatial continuity and reduce abrupt changes across zones.

[0176] The differentiable joint optimization objective function is a total cost function composed of a brightness fidelity term, a halo suppression penalty term, an amplitude limiting constraint, and a spatial smoothing regularization term, and whose gradient can be calculated with respect to the optimization variables.

[0177] The weighted least squares method is a solution that adjusts the influence of each error term with weights and minimizes the sum of squared weighted errors, and is used to generate an initial solution;

[0178] The initial solution is the initial value of the optimization variable obtained by jointly solving the weighted least squares and the barrier function, and is used as the starting point for subsequent iterations;

[0179] The risk-weighted shrinkage is a calculation that reduces the candidate brightness map of the backlight partition based on the amplitude limiting constraint and risk intensity in the inner ring buffer zone, so that the candidate brightness in the high-risk area is suppressed more strongly.

[0180] The initial backlight zone target brightness map is a two-dimensional representation of the brightness target, which is formed by organizing the risk-weighted shrinkage results according to the backlight zones and mapping them to the panel coordinate system.

[0181] The risk weighting setting is a process of weighting the pixel-end tone mapping parameters based on the amplitude limiting constraints in the outer ring buffer band and in combination with the category priority of the semantic mask map.

[0182] The initial pixel-end tone mapping parameter map is a two-dimensional or multi-channel map of pixel-end tone mapping parameters represented on the panel pixel grid, used as the initial state for iterative optimization.

[0183] Optionally, step S8 specifically includes:

[0184] Using the initial backlight partition target brightness map and the initial pixel-end tone mapping parameter map as the initial state for joint optimization, a differentiable joint optimization iterative process is constructed.

[0185] In each iteration, the target brightness of the backlight partition and the pixel tone mapping parameters are updated according to the gradient of the joint optimization objective function, and the update step size is scaled in high-risk areas according to the halo risk heatmap, so that the update step size is smaller for pixel positions with higher risk intensity.

[0186] During the update process, the target brightness of the backlight zone in the inner ring buffer zone is smoothed and limited according to the double-layer buffer zone diagram, the brightness gain parameter and saturation gain parameter in the outer ring buffer zone are smoothed and limited, and the rate of change of the hue shift parameter is suppressed, so that the parameter changes in the buffer zone are limited and continuous.

[0187] After each iteration, based on the risk intensity of the halo risk heatmap and the limiting coefficient in the double-layer buffer zone map, the cross-boundary brightness gradient threshold and the boundary hue offset threshold are determined. The boundary safety set is constructed as a set of parameters that satisfy the cross-boundary brightness gradient not exceeding the cross-boundary brightness gradient threshold and the boundary hue offset not exceeding the boundary hue offset threshold.

[0188] Perform approximate and differentiable constraint projection to project the iteratively updated backlight partition target brightness and pixel-end tone mapping parameters onto the boundary safety set;

[0189] Based on the confidence map, the projection intensity is reduced in high-confidence regions and increased in low-confidence regions, so that the constraint intensity adapts to the recognition confidence.

[0190] The iteration is terminated when the upper limit of the number of iterations is met or the parameter update magnitude is lower than the convergence threshold. The final backlight partition target brightness map and the final pixel end tone mapping parameter map are generated and output.

[0191] Terminology definition:

[0192] The differentiable joint optimization iterative process is a solution process that repeatedly updates the optimization variables under a gradient-calculated objective function and stops when the termination condition is met.

[0193] The update step size is the scaling factor for updating the optimization variables along the gradient correlation direction in each iteration;

[0194] The step size scaling is an operation that adjusts the update step size by multiplying it by a location-related factor based on the spatial location of the halo risk heat map, so that the step size of high-risk locations is reduced and the step size of low-risk locations is relatively increased.

[0195] The high-risk area is a set of pixels in the halo risk heatmap with a risk intensity not lower than a preset threshold or in a high quantile range.

[0196] The smoothing limit is a limiting mechanism that sets an upper limit on variables within the buffer band and implements it through a spatial smoothing or continuous transition function, so as to avoid spatial discontinuity caused by hard truncation.

[0197] The rate of change suppression is a mechanism that sets an upper limit or penalty on the incremental magnitude of the hue shift parameter between adjacent iterations, used to suppress sudden changes in the parameter and maintain continuity;

[0198] The cross-boundary brightness gradient threshold is a threshold that limits the maximum allowed brightness gradient at the semantic boundary, and its magnitude is determined based on the risk intensity and the limiting coefficient.

[0199] The boundary hue offset is a measure of the change in pixel hue angle caused by pixel-end tone mapping in the region near the semantic boundary.

[0200] The boundary hue offset threshold is the threshold for the maximum allowable hue angle change in the boundary area, and its value is determined based on the risk intensity and the limiting coefficient.

[0201] The boundary safety set is a set of optimized variable values ​​that simultaneously satisfy the conditions that the cross-boundary brightness gradient does not exceed the cross-boundary brightness gradient threshold and the boundary hue shift does not exceed the boundary hue shift threshold.

[0202] The approximately differentiable constraint projection is a projection operator that maps the updated variables to the boundary safety set using a smooth approximation or a near-end operator, and its mapping of variables is numerically differentiable or approximately differentiable.

[0203] The projection intensity is a weighting coefficient for the contraction magnitude of the control variables toward the boundary safety set when performing constrained projection, and is used to adjust the projection intensity.

[0204] The high-confidence region is the set of pixels in the confidence map with a confidence level not lower than a set threshold;

[0205] The low-confidence region is the set of pixels in the confidence map whose confidence level is lower than a set threshold;

[0206] The constraint strength is a measure of the degree of restriction imposed on the update of the optimization variables, and can be adaptively adjusted according to the risk intensity and confidence level.

[0207] The limiting coefficient is a scalar coefficient used at the corresponding pixel in the double-layer buffer band map to determine the upper limit of the variable or the suppression ratio.

[0208] The upper limit of the number of iterations is the maximum number of iteration steps set to prevent infinite iteration;

[0209] The parameter update magnitude is a measure of the change in the optimization variable between two adjacent iterations, and can be taken as the norm of the vector difference or an equivalent index.

[0210] The convergence threshold is a numerical threshold used to determine whether the parameter update magnitude is small enough to terminate the iteration.

[0211] The final backlight zone target brightness map is a two-dimensional representation of the backlight target brightness in the panel coordinate system obtained after iteration and constraint projection, which is used to generate backlight zone control commands.

[0212] The final pixel-end tone mapping parameter diagram is a two-dimensional or multi-channel representation of the pixel-end tone mapping parameters obtained after iteration and constrained projection on the panel pixel grid, which is used to generate pixel rendering parameters.

[0213] Optionally, step S9 specifically includes:

[0214] Based on the final target brightness map of the backlight zones, the drive settings for each backlight zone are calculated using the response curve and drive limit of the backlight drive. The drive settings include the drive current setting value, duty cycle and current change slope. The drive settings are then quantized and limited to generate backlight zone control commands.

[0215] Based on the final pixel tone mapping parameter map, calculate the brightness gain map, saturation gain map and hue shift map required for pixel rendering, and quantize and correct the numerical range according to the bit depth of the pixel rendering engine to generate pixel rendering parameters.

[0216] Based on the unified timing conditions, the same frame synchronization identifier is assigned to the backlight partition control command and the pixel rendering parameters, and phase alignment and delay compensation are set according to the panel row scanning order so that the backlight emission time window and the pixel switching time window are aligned in the same frame.

[0217] The backlight zoning control command is sent to the LCD backlight driver, and the pixel rendering parameters are sent to the pixel rendering engine to complete HDR color dynamic enhancement and suppress edge halos on the input image frame, and output the backlight zoning control command and pixel rendering parameters.

[0218] Terminology definition:

[0219] The response curve of the backlight driver is a calibration function that maps the driving current and duty cycle to the backlight optical brightness output. It is usually monotonic and nonlinear and can be corrected with temperature and aging.

[0220] The driving limits are the allowable ranges of driving parameters for backlight driving hardware and safety specifications, including upper and lower limits of current, upper and lower limits of duty cycle, and upper limit of current change slope.

[0221] The driving settings are a set of driving parameters calculated for each backlight zone in the current frame, including at least the driving current setting value, duty cycle and current change slope.

[0222] The driving current setting value is the target current amplitude of the backlight zone light-emitting unit, which is used to determine the output brightness reference.

[0223] The duty cycle is a parameter representing the proportion of the on-time within one cycle of pulse width modulation, used to adjust the average optical output.

[0224] The slope of the current change is the maximum allowable rate parameter of the change in the driving current per unit time, used to suppress overshoot and flicker;

[0225] The quantization is the process of mapping continuous parameters into discrete code values ​​based on the bit depth and resolution of the backlight driver or pixel rendering engine.

[0226] The limiting is a process that trims the calculation parameters to a preset upper and lower limit range to prevent exceeding hardware or security limitations.

[0227] The backlight partition control command is a control data frame containing the settings of each partition driver and encoded according to the protocol, used to instruct the backlight driver to execute;

[0228] The brightness gain map is a two-dimensional map that encodes the brightness gain parameters of each pixel on a pixel grid;

[0229] The saturation gain map is a two-dimensional map that encodes the saturation gain parameters of each pixel on a pixel grid.

[0230] The hue shift map is a two-dimensional or multi-channel map that encodes the hue shift parameters of each pixel on a pixel grid.

[0231] The pixel rendering parameters are a set of parameters that drive the pixel rendering engine to perform tone mapping, including at least a brightness gain map, a saturation gain map, and a hue shift map;

[0232] The bit depth is the number of binary bits of the parameter code value supported by the pixel rendering engine or backlight driver, used to limit the number of discrete levels that can be represented.

[0233] The numerical range correction involves scaling, offsetting, and cropping the pixel rendering parameters according to the target bit depth and the legal range.

[0234] The unified timing condition is a constraint that makes backlight and pixel control follow a unified time base and take effect synchronously within the same frame period;

[0235] The frame synchronization identifier is a synchronization sequence number or timestamp that marks the same video frame or control cycle, used to ensure that the backlight command and pixel parameters are effective in the same frame.

[0236] The panel row scanning order is the timing sequence information of the LCD panel driver updating pixels by row or column, which is used to calculate phase alignment and delay compensation;

[0237] The phase alignment is a control that adjusts the backlight emission time window relative to the pixel switching time window on the time axis to a desired phase relationship;

[0238] The delay compensation is a control that corrects the time offset caused by signal processing, interface transmission and liquid crystal response, so that the backlight and the pixel actually overlap.

[0239] The backlight emission time window is the time interval during which the backlight is allowed to be lit and output luminous flux within one frame period.

[0240] The pixel switching time window is a sampling and holding interval within one frame period during which the pixel voltage is updated and stabilized.

[0241] The LCD backlight driver is a hardware and software module that controls the backlight zone current and duty cycle, including a driver IC, control logic and interface;

[0242] The pixel rendering engine is a hardware and software module that applies pixel rendering parameters and generates panel driver code values ​​in the display pipeline.

[0243] On the other hand, the present invention also provides an image recognition-based HDR color dynamic enhancement system for liquid crystal displays, comprising:

[0244] The recognition module is used to generate brightness maps, semantic mask maps, boundary maps, and confidence maps;

[0245] The partition aggregation module is used to weight and aggregate the brightness map according to the backlight partition mapping table to obtain the backlight partition brightness candidate map.

[0246] The light leakage module is used to generate a spatially variable, anisotropic point spread function kernel map based on the candidate map and panel state parameters, and to generate a light leakage distribution map.

[0247] The buffer zone module is used to generate halo risk heatmaps based on boundary coupling attention and to generate a two-layer buffer zone map.

[0248] The joint optimization module is used to construct a differentiable joint optimization, generate an initial backlight partition target brightness map and an initial pixel-end tone mapping parameter map, iterate and project to a boundary safety set defined by the cross-boundary brightness gradient and the boundary hue offset threshold, and generate the final backlight partition target brightness map and the final pixel-end tone mapping parameter map.

[0249] The control timing module is used to generate backlight partition control commands and pixel rendering parameters based on the final backlight partition target brightness map and the final pixel end tone mapping parameter map, and issue them under unified timing conditions.

[0250] A pixel rendering engine is used to receive the instructions and parameters and align the backlight emission time window and the pixel switching time window within the same frame to achieve HDR enhancement and suppress halo.

[0251] The beneficial effects of this invention are:

[0252] 1. Significantly suppress edge halos and cross-boundary overshoot: By generating a halo risk heatmap through spatially variable, anisotropic PSF light leakage modeling and boundary coupled attention, combined with a double-layer buffer band of inner ring limiting backlight and outer ring limiting pixels, and differentiable joint optimization projecting to the boundary safety set defined by the cross-boundary brightness gradient and boundary hue offset threshold, edge brightness and hue are controlled, protecting the readability of sensitive categories such as subtitles, faces, and text;

[0253] 2. While improving HDR brightness and color, maintain zonal continuity and control stability: By using semantically weighted brightness fidelity terms, spatial smoothing regularization and amplitude limiting constraints, the target brightness of the backlight zone and the pixel-end tone mapping parameters are jointly solved, and the backlight and pixels are aligned under a unified time sequence to reduce the "competition" between the backlight and pixels and reduce artifacts, flicker and zone boundary discontinuity.

[0254] 3. Adaptability and robustness: The PSF parameters are conditionalized according to the panel state conditions such as partition location, driving current, temperature and aging; the buffer bandwidth is adaptive according to the risk intensity and anisotropy ratio; and the constraint projection intensity is adjusted according to the recognition confidence. This enables the present invention to stably output consistent enhancement effects under different panel states and diverse scenarios. Attached Figure Description

[0255] 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:

[0256] Figure 1 This is a flowchart of a method and system for dynamic color enhancement of LCD screens based on image recognition, as proposed in this invention.

[0257] Figure 2 The diagrams show the display effects under a high-contrast point light source scene (moon). The left image shows the display results without the present invention. Due to the mutual interference between backlight partitioning and pixel enhancement, as well as the influence of the isotropic light leakage model, a significant halo is generated around the moon's outline, which also raises the dark background. The right image shows the display results of the embodiment of the present invention. After adopting spatially variable anisotropic PSF modeling, halo risk heatmap generated by boundary coupled attention, double-layer buffer zone along the diffusion axis, and differentiable joint optimization and projection constraints, the lunar surface boundary is clear, the brightness overshoot across the boundary is controlled, the background remains low-brightness and there is no significant halo, while the lunar surface details and peak brightness are preserved.

[0258] Figure 3 This diagram illustrates the display effects of complex boundaries and semantically sensitive scenes involving candle flames, figures, and subtitles. The left image shows the results of the comparison scheme, where the halo around the candle flame spreads to the figures and subtitle areas, causing skin tone shift, edge bloom, and decreased contrast in subtitle strokes. The right image shows the results of an embodiment of the present invention. By assessing risk through boundary coupling attention, limiting the target brightness of the backlight zone in the inner ring buffer zone, limiting the brightness / saturation gain at the pixel end in the outer ring buffer zone, and applying boundary safety set projection to the hue shift, the candle flame edges converge, the figure's skin tone becomes more stable, the subtitle edges are clearer, and readability is improved. Detailed Implementation

[0259] 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.

[0260] refer to Figure 1 A method for dynamic HDR color enhancement of LCD screens based on image recognition, characterized by comprising:

[0261] S1. Perform semantic segmentation and linearization on the input image frame to generate a semantic mask map, boundary map, confidence map, and brightness map;

[0262] S2. The brightness map is weighted and aggregated according to the backlight partition mapping table, and combined with the semantic mask map, boundary map and confidence map for weighting to generate backlight partition brightness candidate map;

[0263] S3. Based on the candidate image of backlight zone brightness and panel state parameters, conditional modeling is performed to generate a spatially variable, anisotropic point spread function kernel map.

[0264] S4. Convolve and superimpose the candidate backlight partition brightness maps using the point spread function kernel map to generate a light leakage distribution map;

[0265] S5. Based on the leakage distribution map, boundary map, confidence map and semantic mask map, a halo risk heat map is generated by weighted fusion through a boundary coupling attention mechanism.

[0266] S6. Generate a double-layer buffer zone map based on the halo risk heat map, wherein the inner ring buffer zone is used to limit the target brightness of the backlight zone, and the outer ring buffer zone is used to limit the pixel-end tone enhancement gain.

[0267] S7. Using the semantic mask map as a condition, construct a differentiable joint optimization objective function based on the candidate backlight partition brightness map, halo risk heat map and double-layer buffer zone map, and jointly solve the target brightness of the backlight partition and the pixel tone mapping parameters to obtain the initial backlight partition target brightness map and the initial pixel tone mapping parameter map.

[0268] S8. Under the constraints of the halo risk heat map and the double-layer buffer zone map, perform differentiable joint optimization iteration on the initial backlight partition target brightness map and the initial pixel end tone mapping parameter map, and adaptively adjust the constraint strength according to the confidence map. After each iteration, perform differentiable constraint projection to the boundary safety set defined by the cross-boundary brightness gradient and the boundary hue offset threshold to obtain the final backlight partition target brightness map and the final pixel end tone mapping parameter map.

[0269] S9. Based on the final backlight partition target brightness map and the final pixel end tone mapping parameter map, generate backlight partition control instructions and pixel rendering parameters, and send them to the backlight driver and pixel rendering engine under unified timing conditions to achieve HDR color dynamic enhancement and suppress edge halo on the input image frame.

[0270] In this specific embodiment, S1 specifically refers to:

[0271] The input image frame is subjected to color space normalization and inverse electro-optical transfer function transformation to obtain a linear luminance representation. Specifically, the normalized three-channel code values ​​are aggregated into a luminance map according to the luminance weights of the target color gamut. Linear brightness calculation can be achieved by the following formula:

[0272] ;

[0273] in, These are the discrete coordinates of the panel pixel grid. These are the red, green, and blue channel code values ​​in the target color gamut after color space standardization. The inverse transform of the electro-optic transfer function is used to map nonlinear code values ​​to linear brightness. The luminance weighting coefficients for the target color gamut are used to converge the three-channel linear luminance into a single-channel luminance map. ;

[0274] After obtaining the brightness map Then, semantic segmentation is performed on the input image frame to generate a semantic mask map. Each pixel is assigned a category label or category probability vector according to a predefined category set, while maintaining the same spatial resolution as the panel pixel grid. A confidence map is generated based on the category probability of each pixel obtained from semantic segmentation. The maximum component or equivalent confidence value of the category probability vector for each pixel is taken for subsequent weighting and constraint adaptation, and further based on the semantic mask image. Category boundaries and brightness map Gradient response generates boundary map At locations near semantic boundaries and with significant brightness gradients, higher boundary probabilities are assigned, and stability and continuity are improved through smoothing and normalization.

[0275] The final output is a semantic mask image. Boundary map Confidence plot and brightness diagram .

[0276] In this specific embodiment, S2 specifically refers to:

[0277] Brightness diagram based on backlight zone mapping table Each pixel specifies a backlight partition number Based on this, semantic weights are constructed respectively. (from semantic mask diagram) Prioritize according to predefined categories to improve readability protection in categories such as captions, faces, and text interfaces; and use boundary proximity attenuation weights. (from boundary map) (Boundary proximity generation to reduce cross-boundary brightness contribution) and confidence weights (From the confidence plot) (Generate, and reduce weights when confidence is below a threshold), multiply the three to obtain pixel weighting coefficients. Subsequently, weighted partition aggregation is performed on each backlight partition, and robust clipping is applied to suppress the impact of abnormally bright or abnormally low bright pixels on the results. The candidate brightness can be calculated by the following formula:

[0278] ;

[0279] in, These are the discrete coordinates of the panel pixel grid. This is a function that maps pixel coordinates to backlight zone numbers in a backlight zone mapping table. Numbering the backlight zones The linear brightness map obtained in step S1, The above pixel weighting coefficients, and In the partitions respectively The robust amplitude limiting threshold is determined based on the lower and upper quantiles of the brightness distribution. The limiting operator is used to crop the partition-weighted average brightness to an allowable range;

[0280] After obtaining the candidate brightness for each zone, the backlight zone numbers are organized into a backlight zone brightness candidate map, which can be used for subsequent processing as needed. Will Mapped to the panel pixel grid, while maintaining the semantic mask map from step S1 throughout the weighted and robust clipping process. Boundary map With confidence plot Consistency in terminology and numerical values ​​is ensured to guarantee the stability and spatial continuity of the partitioning results.

[0281] In this specific embodiment, S3 specifically refers to:

[0282] Number each backlight zone First, use the candidate backlight zone brightness values ​​obtained in step S2. As the weighted average brightness of this partition and in the partition's pixel set From the brightness diagram Extracting peak brightness Statistics on brightness gradient within the partition (Used to characterize the intensity and direction distribution of brightness changes), and then constructing a partition condition vector based on the panel state parameters. ,in Including partition location (based on partition center coordinates) Representation), partition adjacency relationship (Used to constrain the continuity of parameters in adjacent partitions), current drive current (Reflecting energy and diffusion trends), panel temperature (Light leakage characteristics reflecting temperature changes) and aging parameters (Used to correct for diffusion changes caused by long-term decay), followed by brightness statistical features With partition condition vector Input parameter mapping function The set of parameters for the point spread function used to generate this partition includes the principal axis spread radius. Secondary axis diffusion radius (The ratio of these ratios is used as a directional coefficient to characterize the degree of anisotropy), diffusion principal axis orientation angle (Used to describe the direction of the diffusion principal axis) and energy attenuation coefficient (Used to control the weight deceleration rate with distance), and the combined weights of the basic kernel set. (used to weight and combine different forms of anisotropic nuclei in the nucleus dictionary), thereby utilizing the basic nucleus set. The anisotropic point diffusion function kernel for this partition is constructed, and energy normalization, nonnegativity constraints, and continuity constraints on adjacent partition parameters are applied to avoid abrupt changes across partitions. Boundary clipping is also applied to partitions located at the panel edges to prevent out-of-bounds access. The kernel construction can be given by the following equation:

[0283] ;

[0284] in, For partitioning The point spread function kernel in local coordinates The weight at each location is used to represent the spatial distribution of light leakage from that partition to the panel pixel grid. Relative to the partition center The local pixel coordinates are used to represent the discrete sampling of the kernel on the panel pixel grid. The energy normalization constant is used to ensure that the sum of the kernel weights satisfies energy conservation. The number of kernel bases in the basic kernel set is used to limit the number of basis functions in the combination. For partitioning In the The non-negative combination weights on each basic kernel are used to adjust the contribution ratio of each basic kernel. For the first The kernel function of a basic kernel is used to generate anisotropic kernel shapes given parameters. and The diffusion radii along the primary and secondary axes, respectively, are used to determine the diffusion scale. The diffusion principal axis orientation angle is used to control the orientation of the nucleus. The energy decay coefficient is used to control the rate at which the weight decays with distance;

[0285] Finally, all partitions will be... Organize the data into spatially variable point diffusion function kernel maps according to the partition number and output them for subsequent light leakage distribution calculations.

[0286] In this specific embodiment, S4 specifically refers to:

[0287] Anisotropic convolution is performed on the candidate backlight zone brightness maps on the panel pixel grid, and then pixel-by-pixel superposition is performed to generate a light leakage distribution map. Specifically, for each backlight zone number, the corresponding kernel is selected according to the spatially variable point spread function kernel map, and the candidate brightness is used as the convolution signal amplitude for superposition. At the same time, energy normalization, non-negativity constraints, and boundary clipping are followed, and the dynamic range is limited on the result to be consistent with the range of the brightness map. The calculation of the light leakage distribution map can be given by the following formula:

[0288] ;

[0289] in, The values ​​at the panel pixel grid coordinates of the light leakage distribution map are used to describe the light leakage intensity after convolution and superposition of each partition. Discrete coordinates of the panel pixel grid are used to locate pixel positions. The backlight zone number comes from the backlight zone mapping table. The set of partition indexes, The candidate backlight partition brightness values ​​obtained in step S2 are used as the amplitude of the convolution signal for that partition. The partition constructed in step S3 The weights of the anisotropic point spread function kernel at local coordinates with the partition center as the origin are used to represent the contribution of light leakage from that partition to the relative pixel offset position. For partitioning The center coordinates are used to determine the local coordinate system of the kernel. The range limiting operator is used to restrict the result to the dynamic range of the brightness map. Linear brightness diagrams The minimum and maximum values ​​can represent brightness to ensure and The scope is consistent;

[0290] In the above convolution and superposition process, due to Energy normalization and nonnegativity constraints have been implemented in step S3, and boundary clipping has been performed at the panel edges. The values ​​are non-negative and do not include pixels outside the panel. The final output is a light leakage distribution map for subsequent risk assessment and buffer band generation.

[0291] In this specific embodiment, S5 specifically includes:

[0292] Based on the light leakage distribution diagram Boundary map Confidence plot With semantic mask graph Constructing boundary-coupled attention weights involves: first, processing the boundary graph... Perform a distance transformation to obtain the boundary proximity and map it to the boundary proximity weight. A distance-decreasing function is used to give higher weights to pixels closer to semantic boundaries;

[0293] Next, calculate the gradient vector of the boundary map. Gradient vector of the light leakage distribution map directional consistency and mapping to boundary direction weights The weight increases when the main diffusion direction of the light leakage is consistent with the boundary normal direction;

[0294] Again from the confidence plot Generate confidence weights To improve robustness, the weight of pixels with confidence levels below the threshold is reduced.

[0295] And based on the semantic mask diagram Category priority generates category priority weights Increase the weight at the edges of predefined high-priority categories such as subtitles, faces, and text interfaces;

[0296] The above sub-weights are combined by scaling and normalizing to form a boundary coupling attention weight graph. ,by Light leakage distribution diagram The initial risk map is obtained by pixel-wise weighted fusion. Noise suppression and numerical range normalization are then applied to the initial risk map to improve spatial continuity and stability. The calculation of the halo risk heatmap can be expressed by the following formula:

[0297] ;

[0298] in, Discrete coordinates of the panel pixel grid are used to locate pixel positions. The light leakage distribution map obtained in step S4 is used to characterize the light leakage intensity after convolution and superposition of each backlight partition. For the reason and The boundary-coupled attention weight map obtained by normalization combination is used to measure the halo correlation of pixels. The noise suppression operator is used to denoise and smooth the initial risk map to improve its continuity and stability. The numerical range normalization operator is used to map the results to a preset range for use by downstream modules. The halo risk heatmap is used to characterize the risk intensity of halo artifacts in each pixel and is used for the generation of the double-layer buffer band in step S6 and subsequent joint optimization.

[0299] In this specific embodiment, S6 specifically refers to:

[0300] In step S6, a halo risk heatmap is used. The pixel value is used as the risk intensity, and the backlight zone number of each pixel is selected on the panel pixel grid according to the spatially variable point diffusion function kernel map. The diffusion principal axis orientation angle of the partition is obtained from step S3. With the diffusion radius of the primary and secondary axes Calculate the anisotropy ratio Based on this, along the orientation angle Main axis direction Anisotropic expansion is applied to the perimeter of high-value regions, and clipping is performed at the panel edges to generate an inner ring buffer band mask. Its bandwidth is adaptively determined by a preset function of risk intensity and anisotropy ratio;

[0301] exist Anisotropic expansion is continued along the principal axis on the outer side, and non-overlapping constraints are applied to generate an outer ring buffer band mask. Its bandwidth is adaptively determined by a function of decreasing risk intensity and a preset function of anisotropy ratio;

[0302] Then based on the risk intensity Limiting coefficient maps are generated for the inner and outer ring buffer zones, respectively. The inner ring corresponds to the target brightness limiting coefficient map of the backlight zone, and the outer ring corresponds to the pixel-end tone enhancement gain limiting coefficient map. Both tighten as the risk intensity increases and maintain a smooth transition at the junction of the panel edge and the buffer zone. The generation of the limiting coefficient can be expressed by the following formula:

[0303] ;

[0304] in, Discrete coordinates of the panel pixel grid are used to locate pixel positions. This is a function that maps pixel coordinates to backlight zone numbers in a backlight zone mapping table. For the reason The specified backlight partition number is used to index the partition corresponding to this pixel. The halo risk heatmap generated in step S5 is used to measure the risk intensity of halo artifacts occurring in pixels. The anisotropy ratio of the region to which the pixel belongs is used to quantify the degree of diffusion imbalance. The binary function for the inner loop buffer mask takes the value 1 in the band and 1 / 2 out of the band. The binary function of the outer loop buffer band mask does not overlap with the inner loop and takes the value 1 within the band and 1 / 2 outside the band. The output range of a monotonically non-increasing function that generates the target brightness limiting coefficient for backlight zones based on the risk intensity and anisotropy ratio is as follows: The output range of a monotonically non-increasing function that generates pixel-level tone enhancement gain limiting coefficients based on the risk intensity and anisotropy ratio is as follows: , The target brightness limiting factor map for the backlight zone is used to limit the upper limit of the target brightness of the backlight zone associated with that pixel. The pixel-level tone enhancement gain limiting coefficient diagram is used to limit the allowable upper limit of pixel-level brightness gain and saturation gain;

[0305] Ultimately and and and The merged organization is plotted as a two-layer buffer zone diagram for subsequent joint optimization and control.

[0306] In this specific embodiment, S7 specifically refers to:

[0307] semantic masking graph The optimization variable for joint optimization is set as the target brightness of the backlight zone. Pixel-side tone mapping parameters and ,in For backlight partition mapping table The specified backlight zone number, For brightness gain parameters, For saturation gain parameters, This refers to the hue shift parameter;

[0308] Constructing a brightness fidelity item Approximating the candidate values ​​of backlight zone brightness in step S2 The weights are aggregated for each partition based on semantic category priority. Weighted, the Based on semantic weight according to The pixel set is obtained by spatial aggregation;

[0309] Construct a halo suppression penalty term, where the inner ring buffer mask is used. The barrier function suppresses the target brightness of the backlight partition. Through the spatially variable point diffusion function kernel Predicted backlight leakage pattern Cross-boundary brightness gradient The The gradient magnitude in the boundary normal direction, and the boundary normal are determined by the boundary map. unit gradient vector Give For Replace the convolution magnitude in step S4 and use The predicted light leakage distribution obtained by superposition;

[0310] Outer ring buffer mask The synthesis gain metric of pixel-end tone mapping parameters is suppressed by the barrier function. The For brightness gain With saturation gain By weight and Combined gain intensity metric and The coefficients are non-negative.

[0311] Based on the limiting coefficient diagram in the double-layer buffer zone diagram and Generate the corresponding threshold function and This is used to set the upper limit of allowable values ​​in the barrier function and to apply a spatial smoothing regularization term to the target brightness of the backlight zones to constrain the continuity of adjacent zones. The adjacency relationship is determined by the adjacency set of the zones. Adjacency weight Characterization;

[0312] The differentiable objective function for joint optimization can be expressed as:

[0313] ;

[0314] in, To jointly optimize the objective function to simultaneously measure the cost of fidelity, smoothing, and halo suppression, Semantic-weighted fidelity weights are used to improve the fidelity of predefined categories within their respective partitions. The weighting coefficients for spatial smoothing regularization are used to adjust the continuity strength between adjacent partitions. For partitioning The adjacency set is used to enumerate the adjacency sets with Adjacent partition indexes, The smoothing weights for adjacent partition pairs are used to quantize the continuity constraints of the partition intervals. and The weights of the inner and outer loop penalty terms are used to control the strength of the barrier function's effect in different buffer bands. and The binary functions representing the inner and outer ring buffer band masks are used to gating the corresponding spatial ranges. The halo risk heatmap from step S5 is used to provide risk intensity weighting for the penalty item. A smooth barrier function is used to rapidly increase the penalty when the metric exceeds a threshold and maintain a weak penalty below the threshold. For the reason pass The predicted brightness gradient magnitude of the light leakage distribution along the boundary normal direction is used to measure the risk of overshoot across the boundary. The cross-boundary brightness gradient threshold in the inner ring buffer band is determined by... The mapping yields the upper limit for the allowed brightness gradient. For the reason and The combined pixel-side gain metric is used to measure the pixel-side enhancement intensity. The pixel gain threshold in the outer ring buffer band is determined by The mapping yields the upper limit of allowed enhancements;

[0315] Based on the above objective function, an initial solution is generated using a joint solution strategy of weighted least squares and a barrier function, wherein the amplitude limiting constraint in the inner loop buffer band is applied to... Risk-weighted shrinkage is performed to obtain the initial target brightness of the backlight partition. Based on the amplitude limiting constraints in the outer ring buffer zone and combined with Category priority and Initial pixel-level tone mapping parameters are obtained by performing risk weighting. and ;

[0316] Finally, the initial backlight zone target brightness map and the initial pixel end tone mapping parameter map are organized and output according to the zone number and pixel coordinates respectively.

[0317] In this specific embodiment, S8 specifically refers to:

[0318] Initial backlight zone target brightness Initial pixel-side tone mapping parameters As the initial state for joint optimization, an iterative process for constructing differentiable joint optimization is established;

[0319] In the In this iteration, firstly based on the objective function The gradient is used to perform a risk-adaptive update on the variables, with the update step size varying spatially with the halo risk heatmap. Scaling, where the scaling factor at the partition level is determined by the scaling factor within the partition. The statistics were obtained and applied to The scaling factor at the pixel level is directly determined by Map and act on respectively and ;

[0320] Subsequently, under the constraint of the double-layer buffer band diagram, the inner ring buffer band mask was... Related to China Implement smoothing and limiting, and mask the outer ring buffer band. Related to China and Implement smoothing limiting, and for Introduce rate suppression to avoid mutations;

[0321] Based on the halo risk heatmap after each update With limiting coefficient diagram Determine the cross-boundary brightness gradient threshold Boundary hue offset threshold Based on this, a boundary security set is constructed that simultaneously satisfies and The parameter set is used, and the updated variables are projected onto this set using an approximately differentiable projection operator, with the projection intensity varying with the confidence plot. Adaptive, reducing projection in high-confidence regions and enhancing projection in low-confidence regions;

[0322] The gradient step of risk scaling and the confidence-weighted constrained projection described above can be summarized by the following formula:

[0323] ;

[0324] in, and For the first The joint variable vectors before and after each iteration are stacked in order of their components. and Used to uniformly represent the optimization state. For the first The basic step size vector for each iteration has positive components used to control the update scale. The risk step-size scaling map is located at the backlight partition component, which is divided by the partition within the partition. The statistics show that at the pixel component, Pixel-by-pixel mapping is used to reduce the update magnitude at high-risk locations. For the objective function Compared to The gradient is used to indicate the descent direction, and ○ indicates element-wise multiplication for scaling the gradient by its components. For projection onto the boundary safe set The approximate differentiable operator has its second parameter as the projection intensity weight. Used based on confidence plot Adaptive adjustment of constraint strength;

[0325] Iterations until the maximum number of steps is reached Or the parameter update magnitude is less than the convergence threshold The process terminates at a certain time, and the final backlight zone target brightness map is output. The final pixel-level tone mapping parameter map is obtained by organizing the data according to the partition number. and It is obtained by combining on the panel pixel grid.

[0326] In this specific embodiment, S9 specifically refers to:

[0327] Based on the final target brightness of the backlight zones With final pixel-level tone mapping parameters Backlight zone control commands and pixel rendering parameters are generated separately. First, the calibration response curve of the backlight driver is used to perform inverse solving on each zone to obtain the current, duty cycle, and current change slope that meet the target brightness and are constrained by the driving limit. Then, bit depth quantization is performed. Subsequently, the pixel-level brightness gain map, saturation gain map, and hue shift map are quantized and their numerical ranges are corrected according to the bit depth of the pixel rendering engine. Alignment and synchronous distribution under a unified timing are then completed. The calculation of the backlight driver settings can be given by the following formula:

[0328] ;

[0329] in, For backlight partition mapping table The specified backlight partition number is used for indexing partitions. The final backlight zone target brightness for this partition is used to set the driving target. The drive current setting is used to determine the reference for the partitioned light output. The duty cycle is used to adjust the average light output. The slope of the current change is used to constrain the rate of current change per unit time to suppress overshoot and flicker. For partitioning The inverse response mapping function is based on the calibration response curve of the backlight driver. Compare the target brightness with the current of the previous frame. The mapping is nominally set as a triplet to balance brightness and slope control. For partitioning The current from the previous frame is used to calculate the slope and smooth transition. The limiting operator applies upper and lower limits to the current, duty cycle, and slope to ensure they do not exceed the driving limit range. For the upper and lower limits of current, For duty cycle upper and lower limits, These are the upper and lower limits of the slope. The backlight driver quantization operator quantizes the continuous settings into discrete code values ​​based on the backlight driver bit depth;

[0330] On the pixel side, step S8 and Quantization is performed according to the bit depth of the rendering engine and numerical range correction is used to ensure that the brightness gain map, saturation gain map and hue shift map are formed within the legal range. At the same time, under the unified timing condition, the same frame synchronization identifier is assigned to the backlight partition control command and pixel rendering parameters, and phase alignment and delay compensation are set according to the panel row scanning order to make the backlight emission time window and pixel switching time window aligned in the same frame.

[0331] Finally, the backlight zoning control command is sent to the LCD screen backlight driver and the pixel rendering parameters are sent to the pixel rendering engine to complete HDR color dynamic enhancement and suppress edge halos on the input image frame.

[0332] 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.

[0333] This invention transforms the mutual interference between backlight and pixels into synergy by integrating recognition, optical modeling, and optimization control through a joint link between the two. Semantic segmentation and linearization provide categories, boundaries, confidence levels, and brightness metrics. The brightness of candidate partitions is convolved with spatially variable, anisotropic point spread functions to obtain the light leakage distribution, and attention coupled with the boundaries generates a halo risk heatmap, enabling the localization and quantification of cross-boundary light leakage. Based on the risk heatmap, a double-layer buffer band limits the backlight target brightness and pixel hue gain, respectively. Joint optimization simultaneously considers brightness fidelity, halo penalty, and spatial smoothing, and after iteration, projects it onto a boundary safety set defined by the cross-boundary brightness gradient and hue shift threshold. Finally, control is issued under a unified timing, enabling HDR enhancement while controlling edge brightness gradients and stabilizing hue, significantly reducing halos and overshoot, and maintaining partition continuity and readability.

[0334] To specifically address edge halos, this invention incorporates four improvements to the algorithm structure: First, a spatially variable anisotropic point diffusion function conditionalized to the panel state adapts to partition location, adjacency, driving force, temperature, and aging, and is trimmed at the panel edges to improve light leakage prediction accuracy; second, an attention mechanism coupled with the point diffusion function and boundaries, combining directional consistency, distance, and semantic priority to form risk weights and focus on high-risk boundaries; third, a dual-layer buffer band with an inner ring limiting backlight and an outer ring limiting pixels, anisotropically expanding along the diffusion axis and adaptively adjusting bandwidth to suppress overshoot through layered localization; and fourth, differentiable constraints projected onto a boundary safety set, combined with adaptive confidence in step size and amplitude limiting, ensuring iterative convergence and stable control. These synergistic structures transform the parallel, independent control of backlight and pixels into risk-guided collaborative optimization, more efficiently achieving the technical effects of halos suppression and improved HDR display quality.

Claims

1. A method for dynamic color enhancement of LCD screens based on image recognition, characterized in that, include: S1. Perform semantic segmentation and linearization on the input image frame to generate a semantic mask map, boundary map, confidence map, and brightness map; S2. The brightness map is weighted and aggregated according to the backlight partition mapping table, and combined with the semantic mask map, boundary map and confidence map for weighting to generate backlight partition brightness candidate map; S3. Based on the candidate image of backlight zone brightness and panel state parameters, conditional modeling is performed to generate a spatially variable, anisotropic point spread function kernel map. S4. Convolve and superimpose the candidate backlight partition brightness maps using the point spread function kernel map to generate a light leakage distribution map. S5. Based on the leakage distribution map, boundary map, confidence map and semantic mask map, a halo risk heat map is generated by weighted fusion through a boundary coupling attention mechanism. S6. Generate a double-layer buffer zone map based on the halo risk heat map, wherein the inner ring buffer zone is used to limit the target brightness of the backlight zone, and the outer ring buffer zone is used to limit the pixel-end tone enhancement gain. S7. Using the semantic mask map as a condition, construct a differentiable joint optimization objective function based on the candidate backlight partition brightness map, halo risk heat map and double-layer buffer zone map, and jointly solve the target brightness of the backlight partition and the pixel tone mapping parameters to obtain the initial backlight partition target brightness map and the initial pixel tone mapping parameter map. S8. Under the constraints of the halo risk heat map and the double-layer buffer zone map, perform differentiable joint optimization iteration on the initial backlight partition target brightness map and the initial pixel end tone mapping parameter map, and adaptively adjust the constraint strength according to the confidence map. After each iteration, perform differentiable constraint projection to the boundary safety set defined by the cross-boundary brightness gradient and the boundary hue offset threshold to obtain the final backlight partition target brightness map and the final pixel end tone mapping parameter map. S9. Based on the final backlight partition target brightness map and the final pixel end tone mapping parameter map, generate backlight partition control instructions and pixel rendering parameters, and send them to the backlight driver and pixel rendering engine under unified timing conditions to achieve HDR color dynamic enhancement and suppress edge halo on the input image frame.

2. The method for dynamic color enhancement of a liquid crystal screen based on image recognition according to claim 1, characterized in that, S2 specifically refers to: Assign a backlight zone number to each pixel of the brightness map based on the backlight zone mapping table; A semantic weight is assigned to each pixel based on the semantic mask image. The semantic weight is set according to a predefined category priority to improve readability protection in categories such as subtitles, faces and text interfaces. Based on the boundary map, a boundary proximity attenuation weight is applied to pixels located in the region near the boundary of a semantic object to reduce the cross-boundary brightness contribution. Pixels with confidence scores below a threshold are weighted based on the confidence map. The semantic weights, boundary proximity attenuation weights, and confidence weights are multiplied to obtain pixel weighting coefficients, and the luminance map is weighted and partitioned to obtain the weighted average luminance value of each backlight partition. A robust limiting is performed on the weighted average brightness value of each backlight zone. The robust limiting is set based on the high quantile threshold and low quantile threshold of the zone to suppress the influence of abnormally bright or abnormally dark pixels on the zone result. Generate candidate images of backlight zone brightness.

3. The method for dynamic color enhancement of a liquid crystal screen based on image recognition according to claim 1, characterized in that, S3 specifically refers to: Based on the candidate backlight brightness map, calculate the brightness statistics of each backlight zone, including weighted average brightness, peak brightness, and brightness gradient within the zone; A partition condition vector is constructed based on the partition location, adjacency relationship, drive current, temperature, and aging parameters in the panel status parameters; The point spread function parameter set for the backlight zone is generated by using the partition brightness statistical characteristics and partition condition vector through a parameter mapping function, including the spread radius, directionality coefficient and energy attenuation coefficient. The point spread function parameters are weighted and combined using a basic kernel set to generate an anisotropic point spread function kernel; Energy normalization and nonnegativity constraints are applied to the generated anisotropic point spread function kernel, and continuity constraints are applied to the point spread function parameters between adjacent backlight partitions to avoid abrupt changes across partitions. Boundary clipping is applied to the point spread function kernel of the backlight zone located at the edge of the panel to suppress light leakage to the outside of the panel; The point spread function kernels of all backlight zones are organized into a spatially variable point spread function kernel diagram according to the zone number; Output spatially variable point spread function kernel diagram.

4. The method for dynamic color enhancement of a liquid crystal screen based on image recognition according to claim 1, characterized in that, S4 specifically refers to: Anisotropic convolution is performed on the backlight partition brightness candidate map on the panel pixel grid. For each backlight partition, the corresponding point spread function kernel is selected according to the partition number of the backlight partition in the spatially variable point spread function kernel map, and the candidate brightness value of the backlight partition is used as the amplitude of the convolution signal. The convolution results of each backlight zone are superimposed pixel by pixel on the panel pixel grid to form a light leakage distribution map; During convolution and stacking, the energy normalization, non-negativity constraint and boundary clipping implemented in the spatially variable point diffusion function kernel map are followed, so that the value of the light leakage distribution map is non-negative and does not include pixels outside the panel. The light leakage distribution map is numerically limited to match the dynamic range of the brightness map; Generate a light leakage distribution map.

5. The method for dynamic color enhancement of a liquid crystal screen based on image recognition according to claim 1, characterized in that, S5 specifically refers to: Boundary coupling attention weights are constructed based on the light leakage distribution map, boundary map, confidence map, and semantic mask map. The boundary coupling attention weights are obtained by combining the following weights at each pixel: boundary proximity weights are generated based on the distance transformation of the boundary map, and pixels closer to the boundary are given higher weights using a distance decreasing function. Boundary direction weights are generated based on the consistency of the gradient vectors of the boundary map and the gradient vectors of the light leakage distribution map. The weights are increased when the main light leakage diffusion direction is consistent with the boundary normal direction. Confidence weights are generated based on the confidence map, and the weights of pixels with confidence scores below the threshold are reduced. Generate category priority weights based on the semantic mask image, giving higher weights to the edges of predefined categories; The above weights are normalized and combined to form a boundary coupling attention weight graph; The leakage distribution map is weighted and fused pixel by pixel using the boundary coupling attention weight map to generate an initial risk map; Noise suppression and numerical range normalization are performed on the initial risk map to improve continuity and stability; Generate and output a halo risk heatmap.

6. The method for dynamic color enhancement of a liquid crystal screen based on image recognition according to claim 1, characterized in that, S6 specifically refers to: The pixel values ​​of the halo risk heatmap are used as the risk intensity, and the ratio of the diffusion principal axis direction to the anisotropy is calculated on the panel pixel grid based on the spatially variable point diffusion function kernel map. Anisotropic expansion is performed on the perimeter of the high-value region of risk intensity according to the diffusion principal axis direction and anisotropy ratio to generate an inner ring buffer band mask. The bandwidth of the inner ring buffer band mask is adaptively determined according to a preset function of risk intensity and anisotropy ratio, and is clipped at the edge of the panel to avoid exceeding the boundary. Anisotropic expansion is performed along the diffusion principal axis on the outer side of the inner ring buffer band mask, and a non-overlapping constraint is applied to generate an outer ring buffer band mask. The bandwidth of the outer ring buffer band mask is adaptively determined based on a risk intensity decreasing function and a preset function of anisotropy ratio. Based on the risk intensity of the halo risk heat map, the backlight zone target brightness limiting coefficient map is generated using the inner ring buffer band mask, so that the limiting coefficient of the backlight zone target brightness tightens as the risk intensity increases; Based on the risk intensity of the halo risk heatmap, a pixel-end tone enhancement gain limiting coefficient map is generated using the outer ring buffer band mask, so that the limiting coefficient of pixel-end tone enhancement gain tightens as the risk intensity increases; The inner ring buffer band mask and its backlight zone target brightness limiting coefficient map are combined with the outer ring buffer band mask and its pixel end tone enhancement gain limiting coefficient map to form a double-layer buffer band map.

7. The method for dynamic color enhancement of a liquid crystal screen based on image recognition according to claim 1, characterized in that, S7 specifically refers to: The optimization variables for joint optimization are set as the target brightness of the backlight zone and the pixel-end tone mapping parameters, where the pixel-end tone mapping parameters include brightness gain parameters, saturation gain parameters and hue shift parameters; A brightness fidelity term is constructed based on the backlight partition brightness candidate map, so that the target brightness of the backlight partition approximates the backlight partition brightness candidate map in each backlight partition, and the fidelity of the predefined category is improved by using the category priority of the semantic mask map as the weight. A halo suppression penalty term is constructed based on the halo risk heat map and the double-layer buffer zone map. In the inner ring buffer zone, a barrier function for the cross-boundary brightness gradient is set, and in the outer ring buffer zone, a barrier function for the pixel-end tone mapping parameter gain is set, so that the penalty coefficient is larger at pixels with higher risk intensity. Based on the target brightness limiting coefficient map of the backlight zone and the color enhancement gain limiting coefficient map of the pixel end in the double-layer buffer zone map, a limiting constraint is set so that the target brightness of the backlight zone and the color mapping parameters of the pixel end do not exceed the corresponding limiting upper limit within their respective buffer zones. A spatial smoothing regularization term is constructed based on the adjacency relationship between backlight zones to constrain the target brightness difference between adjacent backlight zones to maintain continuity. The brightness fidelity term, halo suppression penalty term, amplitude limiting constraint and spatial smoothing regularization term are combined into a differentiable joint optimization objective function, and the initial solution is generated by the joint solution of weighted least squares and barrier function. Based on the amplitude limiting constraint in the inner ring buffer zone, the candidate brightness map of the backlight partition is risk-weighted and shrunken to generate the initial target brightness map of the backlight partition. Based on the amplitude limiting constraints in the outer ring buffer band and combined with the category priority of the semantic mask map, the pixel-end tone mapping parameters are risk-weighted to generate an initial pixel-end tone mapping parameter map. Output the initial backlight zone target brightness map and the initial pixel-level tone mapping parameter map.

8. The method for dynamic color enhancement of a liquid crystal screen based on image recognition according to claim 1, characterized in that, S8 specifically refers to: Using the initial backlight partition target brightness map and the initial pixel-end tone mapping parameter map as the initial state for joint optimization, a differentiable joint optimization iterative process is constructed. In each iteration, the target brightness of the backlight partition and the pixel tone mapping parameters are updated according to the gradient of the joint optimization objective function, and the update step size is scaled in high-risk areas according to the halo risk heatmap, so that the update step size is smaller for pixel positions with higher risk intensity. During the update process, the target brightness of the backlight zone in the inner ring buffer zone is smoothed and limited according to the double-layer buffer zone diagram, the brightness gain parameter and saturation gain parameter in the outer ring buffer zone are smoothed and limited, and the rate of change of the hue shift parameter is suppressed, so that the parameter changes in the buffer zone are limited and continuous. After each iteration, based on the risk intensity of the halo risk heatmap and the limiting coefficient in the double-layer buffer zone map, the cross-boundary brightness gradient threshold and the boundary hue offset threshold are determined. The boundary safety set is constructed as a set of parameters that satisfy the cross-boundary brightness gradient not exceeding the cross-boundary brightness gradient threshold and the boundary hue offset not exceeding the boundary hue offset threshold. Perform approximate and differentiable constraint projection to project the iteratively updated backlight partition target brightness and pixel-end tone mapping parameters onto the boundary safety set; Based on the confidence map, the projection intensity is reduced in high-confidence regions and increased in low-confidence regions, so that the constraint intensity adapts to the recognition confidence. The iteration is terminated when the upper limit of the number of iterations is met or the parameter update magnitude is lower than the convergence threshold. The final backlight partition target brightness map and the final pixel end tone mapping parameter map are generated and output.

9. The method for dynamic color enhancement of a liquid crystal screen based on image recognition according to claim 1, characterized in that, S9 specifically refers to: Based on the final target brightness map of the backlight zones, the drive settings for each backlight zone are calculated using the response curve and drive limit of the backlight drive. The drive settings include the drive current setting value, duty cycle and current change slope. The drive settings are then quantized and limited to generate backlight zone control commands. Based on the final pixel tone mapping parameter map, calculate the brightness gain map, saturation gain map and hue shift map required for pixel rendering, and quantize and correct the numerical range according to the bit depth of the pixel rendering engine to generate pixel rendering parameters. Based on the unified timing conditions, the same frame synchronization identifier is assigned to the backlight partition control command and the pixel rendering parameters, and phase alignment and delay compensation are set according to the panel row scanning order so that the backlight emission time window and the pixel switching time window are aligned in the same frame. The backlight zoning control command is sent to the LCD backlight driver, and the pixel rendering parameters are sent to the pixel rendering engine to complete HDR color dynamic enhancement and suppress edge halos on the input image frame, and output the backlight zoning control command and pixel rendering parameters.

10. An image recognition-based HDR color dynamic enhancement system for LCD screens, used to execute the image recognition-based HDR color dynamic enhancement method for LCD screens according to any one of claims 1 to 9, comprising: The recognition module is used to generate brightness maps, semantic mask maps, boundary maps, and confidence maps; The partition aggregation module is used to weight and aggregate the brightness map according to the backlight partition mapping table to obtain the backlight partition brightness candidate map. The light leakage module is used to generate a spatially variable, anisotropic point spread function kernel map based on the candidate map and panel state parameters, and to generate a light leakage distribution map. The buffer zone module is used to generate halo risk heatmaps based on boundary-coupled attention and to generate a two-layer buffer zone map. The joint optimization module is used to construct a differentiable joint optimization, generate an initial backlight partition target brightness map and an initial pixel-end tone mapping parameter map, iterate and project to a boundary safety set defined by the cross-boundary brightness gradient and the boundary hue offset threshold, and generate the final backlight partition target brightness map and the final pixel-end tone mapping parameter map. The control timing module is used to generate backlight partition control commands and pixel rendering parameters based on the final backlight partition target brightness map and the final pixel end tone mapping parameter map, and issue them under unified timing conditions. A pixel rendering engine is used to receive the instructions and parameters and align the backlight emission time window and the pixel switching time window within the same frame to achieve HDR enhancement and suppress halo.

Images