Mura quantitative evaluation method for display panel based on space-time multi-dimensional response modeling

Abstract

The application relates to a display panel Mura quantitative evaluation method based on space-time multi-dimensional response modeling. The method comprises the following steps: controlling a display panel to be tested to display a target test image, and collecting images according to a time sampling sequence; performing dark field correction, flat field correction, geometric registration and exposure normalization processing on the original collected images; performing background modeling on the pretreated images, calculating a normalized residual error, and constructing a Mura response tensor covering the space, time, gray scale, color and state dimensions; performing time domain decomposition on the response tensor to obtain static and dynamic defect components, and calculating a time fluctuation graph; performing multi-dimensional response feature extraction including time persistence, dynamic fluctuation and recovery lag, and inputting the normalized and homodirectional sub-indicators into a comprehensive scoring model to obtain a comprehensive Mura score. The application improves the Mura from two-dimensional static defect characterization to multi-dimensional response modeling, and can effectively identify lag-type, fluctuation-type and working condition trigger-type defects.

Description

Technical Field

[0001] This invention relates to the field of display panel defect detection and image quality evaluation technology, specifically to a display panel Mura quantitative evaluation method based on spatiotemporal multidimensional response modeling. Background Technology

[0002] In the field of display panel quality inspection, Mura defects mainly manifest as visual defects such as uneven brightness, uneven color, localized cloudiness, blocky spots, streaks, and abnormal edge transitions, which are important factors affecting panel quality. Existing display panel Mura detection methods mostly employ a single-frame image analysis mode. This involves controlling the control panel to display a uniform grayscale or solid color image, acquiring a single image using a luminance meter, industrial camera, or imaging colorimeter, and then obtaining a static quality score using methods such as background fitting, low-frequency separation, local contrast analysis, or neural network scoring. This type of method treats Mura defects as a static two-dimensional distribution problem, which has significant drawbacks: actual Mura defects exhibit significant time dependence; some defects only appear in the initial lighting stage, after high-brightness excitation, or during temperature rise. Single-frame spatial residual analysis cannot fully reflect the perceived defects and risks under real-world usage conditions. Furthermore, it is difficult to quantify features such as defect duration, recovery hysteresis, and grayscale dependence, leading to discrepancies between the evaluation results and actual visual experience. Additionally, Mura responses at different grayscale levels and colors are treated in isolation, resulting in fragmented information and an inability to achieve a joint representation of spatial and temporal behavior.

[0003] Furthermore, existing Mura detection methods suffer from significant shortcomings in system scalability and engineering compatibility. Current detection processes are mostly designed for specific test conditions. Adding analytical dimensions such as temperature dependence, multi-view consistency, and historical correlation requires substantial system modifications, making smooth expansion within the existing data framework difficult, resulting in poor system reusability and high upgrade costs. In terms of output format, existing methods only provide a single quality score or a two-dimensional defect map, which can only meet the needs of production line grading and cannot be directly used for subsequent engineering scenarios such as De-Mura compensation, panel aging trend tracking, and online health monitoring. These scenarios require the development of independent data interfaces and conversion logic, leading to significant repetitive development workload and low system integration efficiency. Therefore, the industry urgently needs a quantitative Mura evaluation method that can simultaneously characterize the spatial distribution, temporal evolution, and multi-condition response characteristics of Mura, and output structured results with a unified multi-dimensional data framework to meet the needs of multi-scenario reuse and reduce system integration and R&D costs. Summary of the Invention

[0004] To address the technical problem that existing display panel Mura evaluation methods rely only on single-frame images and cannot characterize time dependence and differences in multiple operating conditions, this invention provides a display panel Mura quantification evaluation method based on spatiotemporal multidimensional response modeling. This method elevates Mura from a traditional two-dimensional static defect characterization to a multidimensional response modeling problem covering spatial, temporal, grayscale, color, and state dimensions.

[0005] This invention provides a method for quantitative evaluation of Mura response of display panels based on spatiotemporal multidimensional response modeling, comprising: controlling the display panel under test to sequentially display target test images; acquiring images of the display panel under test according to a preset time sampling sequence to obtain original acquired images; sequentially performing dark field correction, flat field correction, geometric registration, and exposure normalization processing on the original acquired images to obtain preprocessed images; performing background modeling on the preprocessed images to obtain a reference background; calculating normalized residuals based on the preprocessed images and the reference background to construct a Mura response tensor covering spatial, temporal, grayscale, color, and state dimensions; performing temporal decomposition on the Mura response tensor to obtain static defect components and dynamic defect components, and calculating a time fluctuation map based on the dynamic defect components; extracting multidimensional response features based on the Mura response tensor, the static defect components, the dynamic defect components, and the time fluctuation map to obtain multidimensional response feature indices; and inputting the multidimensional response feature indices into a preset comprehensive scoring model to obtain a comprehensive Mura score.

[0006] Furthermore, the control panel under test sequentially displays the target test image, and the panel under test is image-acquired according to a preset time sampling sequence to obtain the original acquired image. This includes: traversing all combinations of grayscale set, color set, and state set; controlling the panel under test to continuously display the target test image under each fixed grayscale, color, and state condition to obtain a test condition combination; and for each test condition combination, the panel under test is image-acquired for multiple frames according to a preset time sampling sequence to obtain the original acquired image.

[0007] Further, for each of the test condition combinations, the process of acquiring multiple frames of images from the display panel under test according to a preset time sampling sequence to obtain the original acquired image includes: controlling the display panel under test to continuously display a preset cold start duration in a first grayscale state to establish a cold start acquisition state, wherein the grayscale value of the first grayscale is less than a preset excitation grayscale threshold; switching the display panel under test from the cold start acquisition state to a high-brightness pre-excitation state; in the high-brightness pre-excitation state, sequentially acquiring an initial stable frame and a recovery frame sequence according to the preset time sampling sequence; and using the initial stable frame and the recovery frame sequence together as the original acquired image.

[0008] Further, the step of sequentially performing dark field correction, flat field correction, geometric registration, and exposure normalization on the original acquired image to obtain a preprocessed image includes: subtracting the dark field template obtained by averaging multiple frames acquired under shading from the original acquired image to obtain a dark field corrected image; dividing the dark field corrected image by the flat field template obtained by calibration with a standard uniform light source to obtain a flat field corrected image; mapping the camera coordinate system to the panel coordinate system using the geometric mapping relationship obtained by fitting the corner points of the calibration plate to the flat field corrected image to obtain a registered image; and using the registered image as input to perform exposure normalization processing to obtain the preprocessed image.

[0009] Further, the step of performing background modeling on the preprocessed image to obtain a reference background, and calculating a normalized residual based on the preprocessed image and the reference background to construct a Mura response tensor covering spatial, temporal, grayscale, color, and state dimensions, includes: performing global polynomial fitting and low-pass filtering on multiple frames of preprocessed images acquired under the same grayscale and color conditions to obtain a polynomial background and a smooth background; performing a weighted combination of the polynomial background and the smooth background to obtain a reference background; calculating the difference between the preprocessed image and the reference background to obtain the original residual; dividing the original residual by the sum of the reference background and a preset stability constant to obtain a normalized residual; and using the normalized residual as the Mura response tensor.

[0010] Further, the step of performing time-domain decomposition on the Mura response tensor to obtain static defect components and dynamic defect components, and calculating a time fluctuation map based on the dynamic defect components, includes: averaging the Mura response tensor along the time dimension to obtain the static defect component; calculating the difference between the Mura response tensor and the static defect component to obtain the dynamic defect component; and calculating the standard deviation of the dynamic defect component along the time dimension to obtain the time fluctuation map.

[0011] Furthermore, the multidimensional response feature indicators include temporal persistence indicators, spatial intensity indicators, dynamic fluctuation indicators, grayscale sensitivity indicators, color sensitivity indicators, state sensitivity indicators, and recovery hysteresis indicators. The step of extracting multidimensional response features based on the Mura response tensor, the static defect component, the dynamic defect component, and the temporal fluctuation map to obtain the multidimensional response feature indicators includes: determining whether each pixel position is in a visible defect state at each sampling time based on the absolute response value of the Mura response tensor at each sampling time and a preset visibility threshold, and statistically analyzing the performance of each pixel position at a given grayscale, color, and... The time duration index is obtained by calculating the proportion of time the static defect component is in a visible defect state under the given conditions; the spatial intensity index is obtained by calculating the defect intensity of the static defect component in the spatial dimension; the dynamic fluctuation index is obtained by calculating the global mean of the time fluctuation map; the grayscale sensitivity index, the color sensitivity index, and the state sensitivity index are obtained by calculating the degree of difference of the static defect component under different gray levels, different color channels, and different display states; and the recovery hysteresis index is obtained by fitting the recovery process of the defect response over time based on the Mura response tensor in the recovery frame sequence under the bright pre-excitation state.

[0012] Furthermore, the recovery hysteresis index is obtained as follows: After highlighting pre-excitation, the defect response amplitude at each sampling time is calculated based on the Mura response tensor in the recovery frame sequence, and the recovery sequence is obtained by arranging them in chronological order; the recovery sequence is fitted with an exponential recovery model to obtain the recovery time constant; an abnormal region is determined according to a preset abnormal region judgment threshold; the recovery time constant within the abnormal region is statistically analyzed to obtain the recovery hysteresis index; wherein, the recovery hysteresis index is positively correlated with the recovery time constant, and the larger the recovery hysteresis index, the slower the defect recovery and the higher the degree of hysteresis.

[0013] Further, the step of inputting the multidimensional response feature indicators into a preset comprehensive scoring model to obtain a comprehensive Mura score includes: normalizing and homogenizing the time persistence indicator, spatial intensity indicator, dynamic fluctuation indicator, recovery hysteresis indicator, grayscale sensitivity indicator, color sensitivity indicator, and state sensitivity indicator among the multidimensional response feature indicators, so that each indicator satisfies the condition that the larger the value, the more severe the defect, thus obtaining normalized feature indicators; weighting and summing the normalized feature indicators with their corresponding preset weight coefficients to obtain a weighted summation result; inputting the weighted summation result into a preset nonlinear mapping function to obtain a mapped score; using the mapped score as a benchmark, determining the corresponding level of the comprehensive score according to a preset scoring level threshold, and outputting the comprehensive Mura score result.

[0014] The present invention has the following beneficial effects:

[0015] First, in response to the problem that existing methods in the background technology treat Mura as a static two-dimensional defect distribution and calculate spatial residuals only through a single frame image, this invention defines Mura as a multidimensional response tensor covering five dimensions: space, time, grayscale, color, and state. Through multi-frame acquisition and temporal decomposition, the time-dependent behavior of defects can be characterized under a unified mathematical framework, breaking through the limitation of traditional methods that can only handle single-frame spatial inhomogeneity.

[0016] Second, addressing the difficulty of existing methods in answering questions such as "how long does the defect last," "does the defect have a recovery lag," and "is the defect only significant at certain gray levels," this invention uses the time dimension as the core evaluation dimension. By separating static and dynamic defect components and constructing sub-features such as time persistence index, dynamic fluctuation index, and recovery lag index, it can distinguish different categories such as steady-state Mura, instantaneous fluctuation type Mura, and recovery lag type Mura. This ensures that the evaluation results not only include "how strong is the defect," but also "how long does the defect last" and "how slow is the defect recovery."

[0017] Third, addressing the issues of inconsistent Mura responses under different grayscale and color conditions in the background technology, and the information fragmentation caused by the isolated processing of different test images in traditional methods, this invention integrates grayscale, color, and state responses into a unified multi-dimensional evaluation framework. It quantifies cross-condition differences through grayscale sensitivity, color sensitivity, and state sensitivity indices, avoiding the information fragmentation problem caused by the isolated processing of different test conditions in traditional methods. Simultaneously, it can identify latent and triggered Muras that traditional single-condition detection cannot detect. Due to the use of multi-component feature extraction and a comprehensive scoring mechanism, the results can be used for both quality level determination and process cause analysis and compensation strategy design.

[0018] Fourth, this invention employs a tensor modeling framework, possessing excellent scalability and engineering compatibility. In terms of tensor dimensions, in addition to space, time, grayscale, color, and operating condition, additional dimensions such as temperature, viewing angle, drive history, refresh rate, and PWM duty cycle can be introduced according to application requirements, thereby forming a structurally unified but more representative Mura evaluation system. Regarding application scenarios, the tensor results of this invention can be directly used for production line grading and inspection, and can also provide a data foundation for subsequent applications such as compensation parameter generation, panel aging trend analysis, and online health monitoring. Specifically, based on the defect responses at different grayscale levels and spatial locations in the tensor, regional correction coefficients or lookup tables can be directly generated to guide panel display compensation; based on the recovery hysteresis index and the state sensitivity index, the potential device aging risk of the panel can be assessed, providing a quantitative basis for lifetime prediction. Attached Figure Description

[0019] Figure 1 This is an overall flowchart of the Mura quantification evaluation method for display panels based on spatiotemporal multidimensional response modeling according to the present invention.

[0020] Figure 2 This is a schematic diagram of the image preprocessing process of the present invention;

[0021] Figure 3 This is a schematic diagram illustrating the process of reference background modeling and Mura response tensor construction for this invention.

[0022] Figure 4 This is a schematic diagram of the multidimensional response feature extraction process of the present invention;

[0023] Figure 5 This is a schematic diagram of the process for calculating the comprehensive Mura score according to the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0026] Example 1

[0027] This embodiment describes the basic process of a display panel Mura quantitative evaluation method based on spatiotemporal multidimensional response modeling, covering a complete technical solution from image acquisition to comprehensive scoring.

[0028] like Figure 1 As shown in the figure, this embodiment provides a Mura quantization evaluation method for display panels based on spatiotemporal multidimensional response modeling, which includes the following steps:

[0029] S1: Control the display panel under test to display the target test image sequentially, and perform image acquisition on the display panel under test according to the preset time sampling sequence to obtain the original acquired image.

[0030] In this invention, Mura refers to the uneven brightness or chromaticity response of a display panel under the combined effects of spatial, temporal, grayscale, color, and state conditions, rather than a traditional single-frame two-dimensional static defect. Specifically, this invention expands the research object from two-dimensional images to multi-dimensional tensors, enabling the dynamic behavior, hysteresis behavior, and multi-condition differences of Mura to be characterized within a unified mathematical framework. Traditional methods only acquire single-frame images and calculate spatial residuals, failing to capture the changing patterns of defects over time. This invention, by acquiring multiple frames in a time-series manner under multiple grayscale, multiple color, and multiple state conditions, expands the representation dimension of Mura from spatial two-dimensional to spatiotemporal multi-dimensional, thereby comprehensively describing the dynamic behavioral characteristics of defects.

[0031] In actual testing, the test plan is first designed based on the type of display panel under test and the type of target defect. The test plan should include at least multiple grayscale conditions, multiple color conditions, and multiple sampling times, and can also include different operating conditions such as cold start state and high-brightness pre-excitation state. The selection of grayscale conditions should cover the commonly used display brightness range of the panel under test, ensuring that both low and high grayscale areas are fully tested, because the backlight utilization, pixel driving voltage, and luminous efficiency of the panel differ significantly under different grayscale conditions, causing the same physical defect to exhibit different visual severity at different grayscale levels. The selection of color conditions should cover the white field and each primary color sub-pixel channel, because the generation mechanisms of luminance-type Mura and chrominance-type Mura are different. The former is mainly caused by backlight non-uniformity or panel thickness differences, while the latter is mainly caused by differences in luminous efficiency between sub-pixels, and they need to be tested separately. The design of the state conditions aims to simulate the different operating conditions that the panel may experience in actual use, because some Mura defects only appear under specific operating conditions. For example, the initial overshoot Mura in the cold start state may completely disappear after the panel enters a steady state.

[0032] To uniformly represent the acquisition results, the original acquired image is defined as:

[0033]

[0034] in, and Represents panel space coordinates, Indicates the sampling time. This represents the test grayscale levels traversed in the current combination of test conditions. Indicates color conditions. This indicates additional state variables.

[0035] The preferred test grayscale set is G={32,64,96,128,160,192,224,255}, covering the range from low gray to high gray; the preferred color set is C={W,R,G,B}, representing four conditions: white point, red point, green point, and blue point; the preferred state set is... The conditions are cold start state and high-brightness pre-excitation state; the preferred time sampling set is T={0.1s,0.5s,1.0s,3.0s}, which correspond to the initial lighting stage, short-term transition stage and near steady state stage, respectively.

[0036] For each set of fixed grayscale, color, and state conditions in the above test condition combinations, the display panel under test is controlled to continuously display the target test image, and multiple frames of images are acquired from the display panel under test according to a preset time sampling sequence. Under fixed grayscale, fixed color, and fixed state conditions, the original image sequence can be represented as follows: .

[0037] During the data acquisition process, sufficient time should be allowed for the panel display to stabilize after each switch in test conditions to avoid transient interference affecting the acquisition results. Simultaneously, ambient lighting and temperature conditions should be kept constant to eliminate the influence of external interference factors on the acquired data. When the grayscale set contains 8 grayscale points, the color set contains 4 conditions, the state set contains 2 conditions, and the time sampling set contains 4 time points, the total number of original images acquired is... Frames. All the original acquired images together constitute the raw data set required for subsequent analysis. Their multi-dimensional coverage characteristics are the basis for subsequent time-domain decomposition and multi-condition sensitivity analysis.

[0038] In practical implementation, the image acquisition module can be implemented using an industrial camera, luminance camera, imaging luminance meter, or imaging colorimeter. The equipment used should have sufficient spatial resolution and dynamic range to ensure effective capture of low-contrast Mura signals. For color Mura analysis, the image acquisition module also needs to have color channel separation capability or colorimetric measurement capability. A stable timing synchronization relationship needs to be established between the display driver module and the image acquisition module, which can be achieved by using hardware trigger lines, synchronization clocks, software timestamp alignment, or a combination of both. The purpose of the timing synchronization relationship is to ensure that each sampling time t accurately corresponds to the corresponding display conditions (g,c,s), thereby ensuring that the original acquired image can truly reflect the spatial response distribution and temporal variation characteristics of the display panel under test under the corresponding conditions, and avoiding the introduction of spurious time-varying signals due to the mismatch between the acquisition time and the display state.

[0039] It is worth noting that during the data acquisition process, the system can also simultaneously record auxiliary information such as the current test grayscale, color conditions, operating status, exposure parameters, ambient temperature, and panel surface temperature. This accompanying information can be used in subsequent analysis to establish a clear correlation between defect changes and test conditions, thereby providing a data foundation for process diagnosis and failure attribution.

[0040] S2: Perform dark field correction, flat field correction, geometric registration and exposure normalization on the original acquired image in sequence to obtain a preprocessed image.

[0041] After obtaining the raw acquired images, it is necessary to eliminate noise and distortion introduced by the imaging system itself, so that images from different times, gray levels, and states can be compared under a unified standard. In actual inspection scenarios, the raw images acquired by industrial cameras are affected by a variety of non-panel factors: sensor dark current and readout circuits can produce fixed offsets under no-light conditions; the spatial response inhomogeneity of lenses and optical components can cause non-uniform images from uniform light sources; small positional deviations between the camera and the panel can lead to spatial misalignment between different frames; and differences in exposure parameters under different gray levels can cause changes in the overall brightness scale. If these factors are not corrected, they will directly introduce false defects or mask the true Mura signal, seriously affecting the accuracy of subsequent background modeling and defect quantification results. Therefore, the quality of image preprocessing directly determines the reliability of the final Mura evaluation results.

[0042] First, perform dark field correction by subtracting the dark field template from the original acquired image. The dark field correction result is defined as follows:

[0043]

[0044] Among them, dark field template The image is obtained by averaging 32 consecutive frames acquired under dark conditions. This average is used to characterize the camera sensor's inherent noise and dark current bias, thereby eliminating the fixed offset of the imaging system under no-light conditions. During the multi-frame averaging process, random noise is effectively suppressed, and the signal-to-noise ratio of the dark field template is significantly improved, preventing dark field noise from being introduced into the target image after correction. Through dark field correction, the obtained dark field corrected image only reflects the brightness information of the panel itself, eliminating the interference of the sensor's fixed offset.

[0045] Next, flat field correction is performed. The result of the flat field correction is defined as follows:

[0046]

[0047] Among them, the flat template Calibrated using a standard uniform light source, this term characterizes the spatial non-uniform response of the lens, sensor, and optical link; stability term. The value is positive to avoid numerical divergence caused by an excessively small denominator. The purpose of planar correction is to eliminate spatial brightness non-uniformity caused by optical systems (such as lens vignetting, sensor edge response attenuation, etc.). Under the condition of uniform light source illuminating an ideal panel, the corrected image should exhibit a uniform brightness distribution; any residual non-uniformity can be attributed to Mura defects in the panel itself. Planar correction weakens the spurious defects introduced by imaging system non-uniformity, resulting in a planar corrected image.

[0048] Geometric registration is then performed, and the result of the geometric registration is defined as follows:

[0049]

[0050] in, This represents the inverse mapping function from panel coordinates to camera image coordinates, determined using a quadratic polynomial mapping obtained by fitting the calibration plate corner points. In actual acquisition, due to mechanical vibration, temperature changes, or panel handling, the relative position between the camera and the panel may shift slightly between different acquisition batches. Geometric registration establishes a mapping relationship from the camera coordinate system to the panel coordinate system, uniformly mapping images under different sampling conditions to the same panel coordinate system, ensuring that the same spatial coordinate always corresponds to the same physical location. Without geometric registration, even sub-pixel-level positional deviations can lead to spurious signals in multi-frame averaging and temporal analysis, affecting the accuracy of temporal decomposition. Geometric registration yields the registered image.

[0051] Finally, exposure normalization is performed, and the result of exposure normalization is defined as follows:

[0052]

[0053] Among them, the normalization factor was exposed. The ratio of the current frame's global median brightness to the reference exposure scale is used to eliminate overall scale variations caused by differences in exposure time, gain, or acquisition configuration. Since the absolute brightness levels of a panel can differ by several orders of magnitude under different grayscale conditions (for example, the brightness of a low grayscale 32 may only be one percent of that of a high grayscale 255), exposure normalization unifies all images to a relative brightness scale, laying the foundation for subsequent cross-grayscale comparisons. Through all the above preprocessing steps, a preprocessed image can be obtained.

[0054] S3: Perform background modeling on the preprocessed image to obtain a reference background. Based on the preprocessed image and the reference background, calculate the normalized residual and construct a Mura response tensor covering the spatial dimension, temporal dimension, grayscale dimension, color dimension and state dimension.

[0055] After obtaining the preprocessed image, it is necessary to separate local Mura defects from the overall brightness trend and construct a response tensor of uniform scale for cross-condition comparison. In actual panel images, even in an ideal panel without Mura defects, its brightness distribution is not perfectly uniform—the central region of the panel is usually brighter than the edge region, exhibiting a gradually varying brightness gradient determined by factors such as backlight structure, light guide plate design, and liquid crystal molecule arrangement. The magnitude of this global brightness trend is much larger than the magnitude of the Mura defects themselves (usually one to several orders of magnitude higher). If it is not removed from the original image, the signal of the Mura defects will be completely submerged by the background gradient. Therefore, the core objective of background modeling is to accurately estimate the brightness distribution of the defect-free ideal panel, i.e., the reference background, and then extract the brightness deviation caused only by defects through residual calculation. Specifically, it includes the following four steps:

[0056] Step 1: Perform global polynomial fitting and low-pass filtering on multiple preprocessed images acquired under the same grayscale and color conditions to obtain polynomial background and smooth background. The reference background modeling adopts a combination of global gradual variation term and local smoothing term. The polynomial background is used to fit the large-scale brightness trend, and the smooth background is used to characterize the locally gradual variation background components.

[0057] Specifically, a global polynomial fitting is performed on multiple preprocessed images to obtain a polynomial background:

[0058]

[0059] in, Represents the polynomial coefficients. and They represent in direction and Polynomial order in direction and Indicates the order of fitting; for global polynomial fitting, a cubic polynomial background model is preferred (i.e., Simultaneously, low-pass filtering is performed to obtain a smooth background. Where * denotes convolution operation, and a Gaussian low-pass kernel is preferred for low-pass filtering. Scale parameters Preferably 25 pixels.

[0060] Step two: Perform a weighted combination of the polynomial background and the smooth background to obtain the reference background. The reference background is represented as:

[0061]

[0062] in, This represents the reference background, used to characterize the ideal brightness distribution of the panel under the current test conditions when there are no local anomalies. and These represent the combined weights of the polynomial background and the smooth background, respectively, and the combined weight of the polynomial background. The preferred value is 0.6, a combined weight for smoothing the background. The preferred value is 0.4.

[0063] Step 3: Calculate the difference between the preprocessed image and the reference background to obtain the original residual, defined as:

[0064]

[0065] The original residual represents the original residual of the preprocessed image relative to the reference background. A positive value indicates that the brightness is too high, and a negative value indicates that the brightness is too low.

[0066] Step four: Divide the original residual by the sum of the reference background and the preset stability constant to obtain the normalized residual. Use the normalized residual as the Mura response tensor. Specifically, the Mura response tensor is defined as:

[0067]

[0068] Among them, the preset stability constant The value is positive to avoid numerical instability when the background value is small. The Mura response tensor covers five dimensions: spatial, temporal, grayscale, color, and state. All preprocessed images are uniformly mapped through this normalization operation to enable comparisons across grayscale, color, and state.

[0069] S4: Perform time-domain decomposition on the Mura response tensor to obtain static defect components and dynamic defect components, and calculate the time fluctuation map based on the dynamic defect components.

[0070] After obtaining the Mura response tensor, it is necessary to distinguish between long-term stable defects and time-varying defects in order to perform feature extraction separately. The temporal behavior of Mura defects can be roughly divided into three categories: the first category is steady-state Mura, which exists with similar amplitudes at all sampling times and is mainly caused by structural factors of the panel (such as backlight non-uniformity, pixel circuit differences, panel thickness variations, etc.), manifesting as static spatial brightness deviations; the second category is transient fluctuation-type Mura, which appears only at specific moments or whose amplitude changes rapidly over time, possibly caused by transient responses of the driving circuit or fluctuations in ambient temperature; the third category is recovery hysteresis-type Mura, which is significantly enhanced at the initial moment after high-brightness pre-excitation and then gradually recovers to a steady-state level, possibly related to physical processes such as pixel charge storage or liquid crystal molecule relaxation. The above three types of defects require different feature extraction strategies and evaluation criteria. Through temporal decomposition, the steady-state component and the temporal component in the Mura response tensor are separated, laying the foundation for subsequent multidimensional response feature extraction.

[0071] The static defect component is obtained by averaging the Mura response tensor along the time dimension, and is defined as follows:

[0072]

[0073] in This represents the Mura response tensor. This represents the number of sampling moments in the time sampling set. Under fixed grayscale, color, and state conditions, the Mura response at four sampling moments is averaged to obtain a spatial distribution of the near-steady-state Mura.

[0074] The dynamic defect component is obtained by calculating the difference between the Mura response tensor and the static defect component, and is defined as follows:

[0075]

[0076] The dynamic defect component is used to remove the time-varying components remaining after removing long-term stable components, highlighting flickering, drifting, fluctuation, and short-term unstable behavior.

[0077] The standard deviation of the dynamic defect component is calculated along the time dimension to obtain the time fluctuation plot, which is defined as:

[0078]

[0079] When the time sampling set contains four sampling times, the root mean square (RMS) of the deviations at these four times is calculated to quantify the temporal instability of each spatial location under the current grayscale, color, and state conditions. A larger value in the time fluctuation map indicates a more unstable defect response and more pronounced time fluctuations at that location.

[0080] S5: Based on the Mura response tensor, the static defect component, the dynamic defect component, and the time fluctuation map, multidimensional response features are extracted to obtain multidimensional response feature indices.

[0081] After obtaining the static defect components, dynamic defect components, and temporal fluctuation maps, it is necessary to extract feature indicators that can comprehensively characterize Mura defects from multiple dimensions. Traditional methods typically only calculate the peak or mean of the spatial residuals as a measure of defect severity, but this approach loses the rich defect behavior information contained in the time, grayscale, and state dimensions. This invention designs seven sub-indices to capture the key characteristics of Mura defects from different evaluation dimensions: The temporal persistence index evaluates the stability of the defect over time, answering the question "How long does the defect last?"; the spatial intensity index evaluates the spatial distribution and severity of the defect, answering the question "How strong is the defect?"; the dynamic fluctuation index evaluates the temporal instability of the defect, answering the question "Does the defect fluctuate?"; the grayscale sensitivity index evaluates the dependence of the defect on different grayscale conditions, answering the question "At which grayscales is the defect more severe?"; the color sensitivity index evaluates the dependence of the defect on different color channels, answering the question "Is the defect a chromatic Mura of a specific color channel?"; the state sensitivity index evaluates the dependence of the defect on different operating conditions, answering the question "Does the defect only appear under specific operating conditions?"; and the recovery hysteresis index evaluates the speed of defect recovery after highlight pre-excitation, answering the question "How slow is the defect recovery?". These seven sub-indices together constitute a multi-dimensional response characteristic index system.

[0082] Specifically, the multidimensional response characteristic indicators include time persistence indicators, spatial intensity indicators, dynamic fluctuation indicators, grayscale sensitivity indicators, color sensitivity indicators, state sensitivity indicators, and recovery hysteresis indicators.

[0083] When calculating the time persistence index, visibility is determined based on the absolute response values ​​of the Mura response tensor at each sampling time, and the proportion of persistent visibility of the defect response over time is statistically analyzed. Specifically, a visibility threshold determination function is first defined:

[0084]

[0085] Among them, the visible threshold The optimal value is 0.020. Based on this, the duration of a single pixel under given conditions is calculated. When the time sampling set contains four time points, the value of this proportion can be 0, 0.25, 0.5, 0.75, or 1. Finally, the average proportion of persistence under all spatial, all grayscale, all color, and all state conditions is calculated, and the global temporal persistence index is defined as follows: This yields the overall time-duration index of the panel. Among them, |A| represents the effective pixel area of ​​the panel, and |A| represents the total number of pixels in the effective pixel area of ​​the panel.

[0086] When calculating the spatial intensity index, under fixed grayscale, color, and state conditions, the proportion of pixels whose absolute value of static defect components exceeds a preset intensity threshold is calculated out of the total number of pixels in the effective pixel area of ​​the panel. This proportion of spatial defect pixels under that condition is obtained. Then, the maximum value among all test conditions is taken as the spatial intensity index, defined as follows: This characterizes the spatial defect strength under worst-case conditions. A preset strength threshold is used. The preferred value is 0.015.

[0087] When calculating the dynamic volatility index, the global mean of the time volatility chart is statistically analyzed and defined as follows: The intensity of time fluctuations is averaged across the entire space, grayscale, color, and state range to obtain the overall dynamic instability.

[0088] When calculating the grayscale sensitivity index, the maximum difference in the mean absolute values ​​of static defect components is calculated for different grayscale levels. The physical significance of the grayscale sensitivity index lies in the fact that mura defects in a panel are not equally visible under all grayscale conditions. Some defects may only be observable under low grayscale conditions (such as low grayscale cloudiness caused by pixel leakage current), some defects may only be significant under high grayscale conditions (such as thermally induced non-uniformity caused by backlight module thermal effects), and some defects may appear or disappear near specific grayscale levels (called the grayscale critical effect). Therefore, the grayscale sensitivity index not only reflects the overall severity of defects but also provides auxiliary information for defect type identification. Specifically, the mean static defect response under all grayscale conditions is first calculated at a fixed color and state as a reference center for grayscale changes. Then, the dispersion of the response at each location under all grayscale conditions is calculated. Finally, the average is taken over all space, all color, and all state conditions to obtain the overall grayscale sensitivity. The larger the value of this index, the stronger the dependence of the panel's mura defects on grayscale changes, requiring additional inspection under specific grayscale conditions during production line quality control.

[0089] When calculating the color sensitivity index, the maximum difference in the mean absolute values ​​of static defect components under different color channels is calculated. The physical meaning of the color sensitivity index is that luminance-type mura (such as backlight unevenness) is usually most significant under white field conditions, while chromaticity-type mura (such as red-green sub-pixel luminance deviation) may be more prominent under specific color channel conditions. By quantifying the dependence of defects on different color channels, the generation mechanism of mura can be distinguished—if the defect is significant under both white field and primary color conditions, it tends to be luminance-type mura, which may be related to the backlight module or the overall structure of the panel; if the defect is significant only under a certain primary color condition, it tends to be chromaticity-type mura, which may be related to the driving circuit or luminescent material of a specific sub-pixel. The calculation logic is similar to that of the grayscale sensitivity index. First, a reference baseline for color dimension analysis is established, then the dispersion of the response at each location under white field, red field, green field, and blue field conditions is calculated, and finally, the average value across the entire space is taken.

[0090] When calculating the state sensitivity index, the maximum difference in the mean absolute value of the static defect component is calculated under different display states. This is used to quantify the degree of response difference at the same location between the cold start state and the state after high-brightness pre-excitation. The physical meaning of the state sensitivity index is that Mura defects in the panel may only appear under specific operating conditions. For example, in the cold start state, the panel temperature is low, and the characteristics of the driving circuit are different from those in the steady state. Some defects that are not obvious in the steady state may be exposed during the cold start. After high-brightness pre-excitation, the pixel charge state and liquid crystal molecule arrangement of the panel change, which may cause Mura defects that were not originally present to appear temporarily. The higher the state sensitivity index value, the greater the influence of the operating conditions on the panel's Mura defects. In production line inspection, it is necessary to conduct sufficient testing under various operating conditions to avoid missed detections.

[0091] When calculating the recovery hysteresis index, the defect response amplitude at each sampling time is calculated based on the Mura response tensor in the recovery frame sequence under the highlighted pre-excitation state, and the recovery sequence is obtained by arranging them in chronological order. An exponential recovery model is fitted to the recovery sequence to obtain the recovery time constant. The larger the recovery time constant, the slower the defect recovery and the higher the degree of hysteresis. To avoid diluting the hysteresis analysis results with normal background areas, abnormal areas can be determined according to a preset abnormal area judgment threshold, and the statistical value of the recovery time constant within the abnormal area is used as the recovery hysteresis index. Specifically, the defect response amplitude at time t in the recovery stage is first defined as:

[0092]

[0093] in, This indicates the state after pre-excitation.

[0094] Then, the recovery sequence is fitted with an exponential recovery model to obtain the recovery time constant; the exponential recovery model is:

[0095] in, Indicates the recovery phase time. The defect response amplitude, Indicates the initial recovery amplitude. This represents the recovered steady-state bias term. Indicates the start time of the recovery modeling. The recovery time constant is represented by the value of the recovery time constant. An abnormal region is determined according to a preset abnormal region judgment threshold. The recovery time constant within the abnormal region is statistically analyzed, and the statistical value is used as the recovery hysteresis index. The recovery hysteresis index is positively correlated with the recovery time constant. The larger the recovery hysteresis index, the slower the defect recovery and the higher the degree of hysteresis.

[0096] The above method can be used to obtain a multidimensional response characteristic index containing 7 sub-indicators.

[0097] S6: Input the multidimensional response feature index into the preset comprehensive scoring model to obtain the comprehensive Mura score.

[0098] After obtaining the multidimensional response characteristic indicators, multiple sub-features need to be integrated into a single comprehensive score to quantify the overall defect level of the entire panel. Within the multidimensional evaluation framework, each of the seven sub-indicators reflects the Mura defect characteristics of the panel from different perspectives. However, production line quality inspection and quality control typically require a concise overall evaluation result to support rapid judgment. The design of the comprehensive score needs to consider two levels of issues: first, weight allocation, i.e., how to determine the relative importance of each sub-indicator in the comprehensive evaluation; and second, score mapping, i.e., how to map the weighted results of multiple indicators to an intuitively usable score range. The weight allocation should be determined based on the correlation between each indicator and subjective perception, as well as its correlation with the long-term reliability of the panel.

[0099] Before weighted fusion, the multidimensional response feature indicators are first normalized and oriented to ensure that the time persistence index, spatial intensity index, dynamic fluctuation index, recovery hysteresis index, grayscale sensitivity index, color sensitivity index, and state sensitivity index all satisfy the condition that a larger value indicates a more severe Mura defect, thus obtaining the corresponding normalized feature indicators. Subsequently, each normalized feature indicator is weighted and summed with its corresponding preset weight coefficient to obtain the comprehensive Mura score. The comprehensive Mura score is defined as follows:

[0100]

[0101] Among them, the weighting coefficient of the spatial intensity index The preferred value is 0.25, which is the weighting coefficient for the recovery hysteresis index. The preferred value is 0.20, which is the weighting coefficient for the time persistence index. The preferred value is 0.15, which is the weighting coefficient for the dynamic fluctuation index, grayscale sensitivity index, color sensitivity index, and state sensitivity index. , , , The preferred values ​​are 0.10 for each. The space strength index and recovery hysteresis index are given higher weights to highlight their association with subjective perception and potential device risks.

[0102] The weighted summation result is input into a preset nonlinear mapping function to obtain the mapped score. The nonlinear mapping function can be a piecewise linear mapping or a sigmoid curve mapping, used to map the weighted summation result to a predetermined scoring interval. Based on the mapped score, the corresponding level of the comprehensive score is determined according to a preset scoring level threshold, and the comprehensive Mura score result is output. This comprehensive Mura score result can be used for production line classification, yield control, and subsequent process diagnostics and compensation strategy formulation.

[0103] Example 2

[0104] This embodiment describes in detail the design method of combining test conditions for image acquisition, based on embodiment 1.

[0105] The control panel under test sequentially displays the target test image, and the panel under test is sampled according to a preset time sampling sequence to obtain the original sampled image. This includes the following steps:

[0106] S1.1: Traverse all combinations of the grayscale set, color set, and state set, and control the display panel under test to continuously display the target test image under each fixed grayscale, color, and state condition to obtain the test condition combination.

[0107] As described in Example 1, the grayscale set G contains 8 grayscale points, the color set C contains 4 color conditions, and the state set S contains 2 operating conditions. The dimensional design of the above three sets is based on the following: the 8 grayscale points in the grayscale set are distributed at approximately equal intervals (the difference between adjacent grayscale levels is approximately 28 to 32 levels). Sampling is appropriately densified in the low grayscale range (three low grayscale points: 32, 64, and 96), because the human eye is more sensitive to uneven brightness under low grayscale conditions, and Mura defects are more easily perceived; sampling is appropriately sparse in the mid-to-high grayscale range (128 to 255), balancing sampling efficiency and coverage integrity. In the color set, the white field condition is used to detect luminance-type Mura, and the red, green, and blue field conditions are used to detect chromaticity-type Mura in each sub-pixel channel, respectively. Among them, defects under the white field condition are usually the most significant, so it is used as the benchmark reference condition. The state set includes a cold start state simulating the first lighting of the panel after a long period of inactivity, and a state simulating the recovery process of the panel after a long period of high-brightness display following a high-brightness pre-excitation. These two states cover the two most representative extreme conditions in actual production line use. Exhaustive combination of the above three sets yields a total of... A set of test conditions is used. Under fixed grayscale, color, and status conditions in each set, the display panel under test is controlled to continuously display the target test image, providing stable display conditions for subsequent multi-frame acquisition.

[0108] S1.2: For each combination of test conditions, the display panel under test is sampled for multiple frames according to a preset time sampling sequence to obtain the original sampled image.

[0109] As described in Example 1, the time sampling sequence comprises four time points. The design of these four time points is based on the following: In the initial lighting phase (approximately 0.1 seconds), the panel's response is primarily determined by the transient characteristics of the driving circuit. At this time, the panel brightness has not yet reached a steady state, reflecting the drive-related Mura characteristics. During the short transition phase (approximately 0.5 and 1.0 seconds), the panel brightness gradually stabilizes, capturing the temporal evolution of recovery hysteresis and gradual Mura. In the near-steady-state phase (approximately 3.0 seconds), the panel essentially enters a steady state, reflecting the true distribution of steady-state Mura. The time intervals of the four time points are arranged in a denser-to-sparser manner (0.1s→0.5s→1.0s→3.0s) to provide higher temporal resolution in the rapidly changing initial phase. Under each set of fixed grayscale, color, and state conditions, the display panel under test is controlled to display the corresponding test image for 3 seconds, and images are acquired sequentially according to the aforementioned time sampling sequence, resulting in four original images. The above acquisition operation is performed on 64 sets of test condition combinations, yielding a total of 64×4=256 original acquired images.

[0110] Example 3

[0111] This embodiment, based on Embodiment 2, describes in detail a multi-frame acquisition method including cold start and high-brightness pre-excitation state switching. For each test condition combination, the display panel under test is subjected to multi-frame image acquisition according to a preset time sampling sequence to obtain the original acquired image, including:

[0112] S1.2.1: Control the display panel under test to continuously display a preset cold start duration in the first grayscale state to establish a cold start acquisition state, wherein the grayscale value of the first grayscale is less than a preset excitation grayscale threshold.

[0113] The display panel under test is controlled to continuously display a preset cold start duration in the first grayscale state to establish a cold start acquisition state, wherein the grayscale value of the first grayscale is less than a preset excitation grayscale threshold. In each test condition, establishing the cold start state requires ensuring the panel is sufficiently still before acquisition. Specifically, the display panel under test is powered off and left to stand for 10 minutes to allow the panel temperature to return to ambient temperature, eliminating any residual influence of past display states on the current test. After standing, the display panel under test is controlled to continuously display the preset cold start duration in the first grayscale state. The grayscale value of the first grayscale is less than the preset excitation grayscale threshold, i.e., the first grayscale is a grayscale value outside the test grayscale set G, preferably grayscale value 0 (i.e., black field state), and the preset excitation grayscale threshold is preferably grayscale value 32. The cold start duration is set according to the panel type and size to ensure the panel enters a stable cold working condition.

[0114] It should be understood that the design basis for the 10-minute power-off rest period during the establishment of the cold start acquisition state is that after the display panel has been working for a long time, the orientation of liquid crystal molecules, the exciton state distribution of organic light-emitting materials, and the charge storage state in the pixel driving circuit are all affected by the historical display state. Sufficient power-off rest allows these physical processes to fully relax to thermal equilibrium, eliminating the residual influence of the historical display state on the current test results, thus ensuring that the cold start acquisition state truly represents the response characteristics of the panel from its "initial static" state to its working state. For different types of display panels, the power-off rest time can be adjusted according to the panel type: for OLED panels, due to the longer relaxation time of organic materials, it is recommended to appropriately extend the rest time; for LCD panels, the relaxation time of liquid crystal molecules is relatively short, and 10 minutes is usually sufficient.

[0115] S1.2.2: Switch the display panel under test from the cold start acquisition state to the high brightness pre-excitation state. In the high brightness pre-excitation state, acquire the initial stable frame and the recovery frame sequence in sequence according to the preset time sampling sequence, and use the initial stable frame and the recovery frame sequence together as the original acquired image.

[0116] The display panel under test is switched from the cold start acquisition state to the high-brightness pre-excitation state. In the high-brightness pre-excitation state, initial stable frames and recovery frame sequences are acquired sequentially according to the preset time sampling sequence. The initial stable frames and the recovery frame sequences are used together as the original acquired image. After establishing the cold start acquisition state, the response characteristics of the panel need to be stimulated by high-brightness pre-excitation, and images need to be acquired during the recovery process after the excitation ends to capture hysteresis and recovery Mura.

[0117] Specifically, the display panel under test is controlled to display a full-screen white grayscale of 255 for 60 seconds to complete the highlight pre-excitation. After the pre-excitation, the display panel under test is immediately switched to the target test conditions (i.e., target grayscale and target color), and images are acquired sequentially according to the preset time sampling sequence. The image acquired at 0.1 seconds is the initial stable frame, and the images acquired at 0.5 seconds, 1.0 seconds, and 3.0 seconds constitute the recovery frame sequence. The initial stable frame and the recovery frame sequence are used together as the original acquired images. Through this acquisition method of excitation followed by recovery, the Mura change characteristics of the panel during the gradual recovery process after the highlight pre-excitation can be captured, including behaviors such as initial overshoot, slow decay, and steady-state convergence.

[0118] The design basis for the high-brightness pre-excitation parameters (full-screen white level grayscale 255, duration 60 seconds) is as follows: High brightness represents the state of maximum excitation applied to the panel driving circuit and luminescent material. Under this condition, the organic luminescent material of the OLED panel experiences maximum power density, and the pixel driving transistors are at maximum gate-source voltage, making it most likely to excite hysteresis defects related to material aging, charge accumulation, and thermal effects. For LCD panels, high-brightness pre-excitation brings the backlight module to thermal stability, and the liquid crystal molecules are aligned for a long time under maximum driving voltage, making it most likely to expose defects related to liquid crystal relaxation and threshold drift of the driving circuit. The 60-second excitation duration is a balance between testing efficiency and sufficient excitation: too short a time (e.g., less than 10 seconds) may not allow the panel to reach a fully excitation state; too long a time (e.g., more than 5 minutes) will significantly increase the overall testing time, which is unfriendly to production line applications. In specific applications, the excitation duration can be adjusted according to the panel type and the type of target defect.

[0119] The sequential excitation protocol of "cold start → high brightness pre-excitation → recovery acquisition" can effectively excite three types of defects that are difficult to detect in traditional single-frame testing: The first type is cold start sensitive Mura, which only appears in the initial stage when the panel is first lit from a cold state and gradually disappears after the panel enters a thermal steady state. Traditional methods cannot detect it at all when testing after the panel is preheated. The second type is high brightness pre-excitation induced Mura, which appears during the recovery process after high brightness excitation and is manifested as a persistent brightness deviation in the recovery frame sequence. Its generation mechanism is related to charge residue in the pixel driving circuit or hysteretic relaxation of liquid crystal molecules. The third type is hysteretic recovery Mura, which does not disappear immediately after high brightness excitation but recovers gradually at a slower rate. In the recovery frame sequence, it is manifested as a defect response that decays monotonically with time and can be quantitatively characterized by the recovery hysteresis index.

[0120] Example 4

[0121] like Figure 2 As shown, this embodiment, based on Embodiment 1, further describes the implementation method of sequentially performing dark field correction, flat field correction, geometric registration, and exposure normalization on the original acquired image to obtain a preprocessed image, including:

[0122] S2.1: Subtract the dark field template obtained by averaging multiple frames acquired under shading conditions from the original acquired image to obtain the dark field corrected image.

[0123] The dark-field corrected image is obtained by subtracting the dark-field template (obtained by averaging multiple frames acquired under shading) from the original acquired image. Dark-field correction eliminates the sensor's fixed offset by subtracting the dark-field template from the original acquired image. In practice, the acquisition of the dark-field template must be performed in a completely shading environment to ensure that no external light enters the imaging optical path. Averaging 32 frames can reduce random noise by approximately [missing information]. This ensures that the residual noise level in the dark field template is significantly lower than the typical signal amplitude of panel Mura defects, preventing noise from the dark field template from being introduced into the corrected image. Once calibrated, the dark field template can be reused with the camera configuration unchanged, but it is recommended to re-acquire it periodically (e.g., daily or before each batch of tests) to eliminate the effect of sensor temperature drift on dark current bias. For camera sensor types where dark current varies significantly with temperature, dark field templates can also be acquired separately under different temperature conditions, and interpolation can be performed based on the sensor temperature at the time of acquisition.

[0124] S2.2: Divide the dark field corrected image by the flat field template obtained by standard uniform light source calibration to obtain the flat field corrected image.

[0125] The dark-field corrected image is divided by the flat-field template calibrated by a standard uniform light source to obtain the flat-field corrected image. Flat-field correction eliminates the spatial non-uniform response of the imaging system by dividing the dark-field corrected image by the flat-field template. In specific implementations, the calibration quality of the flat-field template directly determines the correction effect: the standard uniform light source should be as close as possible to an ideal Lambertian surface light source, and the light source's luminous uniformity and color temperature stability must meet preset requirements. For large-area panel inspection scenarios, integrating sphere light sources or multi-point LED array light sources can be used to improve uniformity. When calibrating the flat-field template, it should be performed under the same focal length, aperture, and object distance conditions as the actual acquisition to maintain the consistency of the imaging system's optical state. Stability Items The value of should match the minimum value of the flat template to avoid numerical amplification in areas where the local value is too small in the flat template.

[0126] S2.3: The camera coordinate system is mapped to the panel coordinate system using the geometric mapping relationship obtained by fitting the corner points of the calibration plate to the flat field correction image, so as to obtain the registration image.

[0127] The flat-field corrected image is mapped from the camera coordinate system to the panel coordinate system using a geometric mapping relationship obtained by fitting the corner points of the calibration board, resulting in a registered image. Geometric registration transforms the flat-field corrected image from the camera coordinate system to the panel coordinate system using an inverse mapping function. In specific implementations, the geometric calibration board should cover the effective display area of ​​the panel, and the corner point distribution density on the calibration board should meet the fitting accuracy requirements. The quadratic polynomial mapping relationship is suitable for approximately rigid geometric relationships between the camera and the panel. When the camera shooting angle is large or there is significant lens distortion, a higher-order polynomial mapping or a flexible registration method based on thin plate splines can be considered. Registration accuracy can be evaluated by reprojection error, preferably controlled within 0.5 pixels to ensure the accuracy of alignment of the same physical location in different frame images.

[0128] S2.4: Using the registered image as input, perform exposure normalization processing to obtain the preprocessed image.

[0129] The registered image is used as input, and exposure normalization is performed. Exposure normalization eliminates overall scale differences between different acquisition conditions by dividing by an exposure normalization factor. In specific implementations, the exposure normalization factor is taken as the ratio of the global median brightness of the current frame to the reference exposure scale, where the median is more robust than the mean and is not affected by abnormal brightness values ​​in local Mura defect areas. The reference exposure scale is preferably taken as the global median brightness under white point conditions (grayscale 255), so that all images are unified to a relative brightness scale based on the white point. In actual calculations, for extreme cases where the median brightness is close to zero under extremely low grayscale conditions, a lower limit threshold for the median brightness can be set to avoid the exposure normalization factor being too large and causing numerical overflow. After performing the above exposure normalization processing on the registered image, the preprocessed image is obtained.

[0130] Example 5

[0131] like Figure 3 As shown, this embodiment, based on embodiment 1, describes in detail the specific implementation of performing background modeling on the preprocessed image to obtain a reference background, calculating the normalized residual based on the preprocessed image and the reference background, and constructing a Mura response tensor covering the spatial dimension, temporal dimension, grayscale dimension, color dimension, and state dimension, including:

[0132] S3.1: Perform global polynomial fitting and low-pass filtering on multiple preprocessed images acquired under the same grayscale and color conditions to obtain polynomial background and smooth background.

[0133] The reference background modeling employs a combination of polynomial background and smooth background methods. In practice, the input data for polynomial fitting requires special handling: because panel brightness exhibits slight temporal fluctuations over a short period, directly fitting a single frame image is easily affected by random noise. Therefore, it is preferable to use the temporal mean of multiple preprocessed images under the same grayscale and color conditions as the fitting input; that is, first calculate... And then Perform polynomial fitting. Cubic polynomial background model ( The model contains 16 coefficients, selected based on the following criteria: Low-order coefficients (such as first or second order) cannot fit the non-linear brightness decay at the panel edges, while high-order coefficients (such as fourth order and above) easily absorb the brightness distortion of local Mura anomaly areas into the background, leading to an underestimation of defect intensity in subsequent residuals. The coefficients are solved using the least squares method during the fitting process. For regions with obvious local anomalies, median filtering can be used for preprocessing before fitting to remove the influence of abnormal pixels and improve the robustness of the polynomial background fitting to the overall brightness trend.

[0134] In low-pass filtering, the scaling parameter of the Gaussian low-pass kernel The pixel selection is based on the following: this scale corresponds to a spatial range of approximately 1 to 2 centimeters on the panel, effectively filtering out pixel-level noise and sub-pixel-level brightness fluctuations, while preserving the true brightness gradient trend on larger scales (several centimeters or more) of the panel. For a smooth background after low-pass filtering, its spatial frequency cutoff frequency is... Inversely proportional, The larger the resolution, the fewer low-frequency components are retained. In practice, adjustments can be made based on panel resolution and defect size distribution. For high-resolution panels (such as 4K and above), the value can be appropriately increased. For low-resolution panels, the size can be appropriately reduced. .

[0135] S3.2: The polynomial background and the smooth background are weighted and combined to obtain a reference background.

[0136] In this embodiment, the polynomial background and the smooth background are weighted and combined to obtain a reference background, wherein the combination weight is preferably... , In practical applications, the specific value of the weight can be adjusted according to the panel characteristics: for panel types with good brightness uniformity and a dominant global trend, the weight can be appropriately increased. To enhance the fitting of global trends; for panel types with many localized gradual inhomogeneities, the value can be appropriately increased. To improve local adaptability. The quality of the reference background directly affects the detection sensitivity of defects in the subsequent Mura response tensor—if the reference background overfits the local anomaly, the defect signal in the residual is weakened; if the reference background underfits the local trend, the residual contains too much residual background, affecting the accurate localization and quantification of defects.

[0137] It should be understood that the aforementioned reference background modeling method combining polynomial background and low-pass smoothing background is a preferred embodiment of the present invention, and other background modeling methods are not excluded. In an optional embodiment, a low-rank plus sparse decomposition method can also be used to model the reference background: the Mura response tensor is decomposed into the sum of a low-rank part and a sparse part. The low-rank part is considered as the system background of the panel under the current test conditions, and the sparse part is considered as the local abnormal defect response, thereby adaptively separating the background from the defect without explicitly fitting polynomial coefficients. This scheme is particularly suitable for scenarios with complex background distributions or where there is a need for joint analysis of multiple operating conditions. In another optional embodiment, a background template based on a good product reference panel can also be used as the reference background. That is, images are acquired using a known good product panel of the same model as the panel under test under the same test conditions, and the temporal mean of the good product panel image is directly used as the reference background. All of the above background modeling methods can be flexibly used within the technical framework of the present invention, and those skilled in the art can choose according to the specific detection scenario and panel characteristics.

[0138] S3.3: Calculate the difference between the preprocessed image and the reference background to obtain the original residual.

[0139] The raw residual is used to characterize the degree to which each spatial location deviates from the ideal background. In actual calculations, the numerical distribution of the raw residual usually approximates a zero-mean distribution. In areas where there are no Mura defects, the residual is close to zero, while in defective areas, the residual deviates significantly from zero. A positive residual indicates that the brightness at that location is higher than the reference background (i.e., bright Mura), and a negative residual indicates that the brightness at that location is lower than the reference background (i.e., dark Mura). Because the absolute brightness levels of the panel vary greatly under different grayscale conditions (for example, the brightness of a low grayscale 32 may only be one percent of that of a high grayscale 255), the absolute value of the raw residual for the same physical defect at different grayscales may differ by several orders of magnitude. Therefore, the raw residual itself is not suitable for direct comparison across grayscale levels.

[0140] S3.4: Divide the original residual by the sum of the reference background and the preset stability constant to obtain the normalized residual, and use the normalized residual as the Mura response tensor.

[0141] In this embodiment, the defect responses under different grayscale and color conditions are uniformly mapped to a relative scale space by dividing by the reference background and normalizing the residuals to obtain the Mura response tensor. In specific implementations, the normalized values ​​are typically in the range of [missing information - likely a specific value]. Fluctuations within a certain range, with areas where the absolute value exceeds 0.02 typically correspond to Mura defects that are visually perceptible. Preset stable constant. The value of should match the level of background brightness under the lowest grayscale condition of the panel, preferably 0.1% to 1% of the lowest grayscale background brightness, in order to achieve a balance between numerical stability and normalization accuracy. All 256 preprocessed images are uniformly mapped to the relative residual space through this normalization operation, covering five dimensions: spatial dimension, temporal dimension, grayscale dimension, color dimension, and state dimension.

[0142] Example 6

[0143] This embodiment, based on Embodiment 1, describes in detail the implementation method of performing time-domain decomposition on the Mura response tensor to obtain static and dynamic defect components, and calculating a time fluctuation map based on the dynamic defect components, including:

[0144] S4.1: Average the Mura response tensor along the time dimension to obtain the static defect component.

[0145] The static defect component is obtained by averaging the Mura response tensor along the time dimension. In practice, the time-domain averaging operation is equivalent to low-pass filtering at each spatial location. Its physical significance lies in the fact that long-term stable Mura defects on the panel (such as fixed brightness non-uniformity caused by differences in backplane structure or pixel circuitry) exist at all sampling times. After time-domain averaging, the signal is preserved and the signal-to-noise ratio is improved; while random noise and short-term flicker tend to cancel each other out after time-domain averaging. When the time sampling set contains four moments, time-domain averaging can reduce the standard deviation of random noise by about half. The static defect component reflects the Mura spatial distribution of the panel closer to a steady state under the current grayscale, color, and state conditions, and is one of the main inputs for subsequent feature extraction.

[0146] S4.2: Calculate the difference between the Mura response tensor and the static defect component to obtain the dynamic defect component.

[0147] The dynamic defect component is the difference between the Mura response tensor and the static defect component. In practice, the dynamic defect component is essentially a high-pass component in the time dimension, capable of revealing time-varying defects that cannot be detected by static analysis alone. For example, some panels exhibit weak flickering under low grayscale conditions; the mean value of its static component at each moment may be close to zero (because the positive and negative deviations of flicker cancel each other out), but it appears as a significant non-zero value in the dynamic component. Furthermore, the sign information of the dynamic defect component also has diagnostic value: if the dynamic component is positive in the early stage of illumination and tends to zero in the later stage, it indicates the presence of initial overshoot type Mura; if the dynamic component remains positive at all moments, it indicates that the panel brightness is continuously increasing, possibly indicating a drive-related gradient drift problem.

[0148] S4.3: Calculate the standard deviation of the dynamic defect component along the time dimension to obtain the time fluctuation diagram.

[0149] The time fluctuation map is obtained by calculating the standard deviation of the dynamic defect component along the time dimension. In practice, the time fluctuation map compresses the multi-frame time variation of each spatial location into a single scalar value, facilitating the rapid location of time-varying unstable regions. The time fluctuation map can be used in conjunction with the static defect component to form a two-dimensional joint analysis view: the static defect component is used as the brightness channel of one image, and the time fluctuation map is used as the brightness channel of another image, which can intuitively distinguish between steady-state Mura (high static component, low time fluctuation), transient fluctuation Mura (low static component, high time fluctuation), and recovery hysteresis Mura (medium static component, medium time fluctuation). In actual production line applications, the time fluctuation map can also be used to set dynamic thresholds—when the time fluctuation value of a certain area exceeds a preset threshold, an alarm is triggered, indicating that there is abnormal time-varying behavior in that area.

[0150] Example 7

[0151] like Figure 4 As shown, this embodiment, based on embodiment 6, further describes the multidimensional response feature indicators, including time persistence indicators, spatial intensity indicators, dynamic fluctuation indicators, grayscale sensitivity indicators, color sensitivity indicators, state sensitivity indicators, and recovery hysteresis indicators. The implementation method for extracting multidimensional response features based on the Mura response tensor, the static defect component, the dynamic defect component, and the time fluctuation map to obtain the multidimensional response feature indicators includes:

[0152] S5.1: Based on the absolute response value of the Mura response tensor at each sampling time and the preset visibility threshold, determine whether each pixel position is in a visible defect state at each sampling time, and calculate the proportion of time that each pixel position is in a visible defect state under a given gray level, color and state condition to obtain the time persistence index.

[0153] In this embodiment, the absolute response value of the Mura response tensor at each sampling time is first compared with a preset visibility threshold. If the absolute response value is greater than or equal to the visibility threshold, the pixel position is determined to be in a visible defect state at that time and marked as 1; if the absolute response value is less than the visibility threshold, it is determined to be invisible and marked as 0. The visibility threshold is preferably set to 0.020, which corresponds to the normalized residual amplitude that the human eye can just distinguish under standard observation distance and brightness conditions. After extensive subjective experimental verification, when the absolute value of the normalized residual of the Mura response tensor exceeds this threshold, the proportion of "visible" or "clearly visible" in the human eye rating increases significantly.

[0154] Based on this, for each pixel location, given grayscale, color, and state conditions, the proportion of sampling times in the visible defect state to the total number of sampling times is calculated to obtain the time duration proportion. When the time sampling set contains 4 times, this proportion can take the value of 0, 0.25, 0.5, 0.75, or 1.

[0155] Finally, the arithmetic mean of the time duration percentages of all pixel positions, all grayscale conditions, all color conditions, and all state conditions within the effective pixel area of ​​the panel is taken to obtain the global time duration index of the panel.

[0156] In practical applications, the requirements for continuity can be adjusted according to the panel application scenario: for consumer electronics panels, the weight of the time-duration indicator can be appropriately reduced, because users have a higher tolerance for momentary mura; for application scenarios such as automotive displays or medical displays, which have high requirements for long-term reliability, the time-duration indicator should be given higher weight.

[0157] S5.2: Calculate the defect intensity of the static defect component in the spatial dimension to obtain the spatial intensity index.

[0158] In this embodiment, the defect intensity of the static defect component in the spatial dimension is calculated to obtain the spatial intensity index, and the specific implementation is as follows:

[0159] The spatial intensity index is obtained by statistically analyzing the proportion of pixels whose static defect components exceed a preset intensity threshold in the spatial dimension to the total number of pixels in the effective pixel area of ​​the panel. Under fixed grayscale, color, and state conditions, the number of pixels whose absolute value of the static defect component exceeds the preset intensity threshold is counted, and the proportion of this number to the total number of pixels in the effective pixel area of ​​the panel is calculated to obtain the spatial defect pixel ratio under that condition. Then, the maximum value among all test conditions is taken as the spatial intensity index, defined as:

[0160]

[0161] To characterize the spatial defect strength under worst-case conditions.

[0162] In practical implementation, a preset intensity threshold is used. The preferred value is 0.015, which is slightly lower than the visibility threshold. This is because the spatial intensity index targets the static defect component (the result after time-domain averaging), whose random noise has been suppressed by the averaging operation. Therefore, a lower detection threshold can be used to capture weak but spatially widespread Mura defects. The maximum value (rather than the mean) is used as the spatial intensity index because Mura defects on the same panel are usually not equally significant across all grayscale levels, colors, and states. Instead, they reach maximum exposure under specific conditions, and the maximum value index ensures that the most severe defects are not underestimated. In production line applications, the combination of grayscale, color, and state conditions corresponding to when the spatial intensity index reaches its maximum value can also be recorded, providing reference information for subsequent process localization and defect tracing.

[0163] S5.3: Calculate the global mean of the time fluctuation graph to obtain the dynamic fluctuation index.

[0164] The dynamic fluctuation index is obtained by taking the global average of the temporal fluctuation map across the entire space, grayscale, color, and state range. In practice, the value of the dynamic fluctuation index is usually much smaller than that of the spatial intensity index because the time-varying component of most panel muras only accounts for a small portion of the total defect response. However, the dynamic fluctuation index is important for identifying specific types of defects: for example, flickering muras may appear weak in the static component (because the positive and negative deviations of flickering are partially canceled out in the time-domain averaging), but they appear significantly abnormal in the dynamic fluctuation index. In addition, the dynamic fluctuation index can also be used to distinguish between time-varying noise associated with the drive circuit and genuine panel mura defects—the former are usually spatially randomly distributed with high global fluctuation values, while the latter are spatially locally clustered with fluctuations concentrated in specific areas.

[0165] S5.4: Calculate the degree of difference of the static defect component under different gray levels, different color channels, and different display states to obtain the gray level sensitivity index, the color sensitivity index, and the state sensitivity index.

[0166] After obtaining the static defect components, it is necessary to analyze the dependence of the defects on different gray levels, colors, and operating conditions. This multi-condition sensitivity analysis is one of the key advantages of this invention compared to traditional single-frame detection methods. Traditional methods typically perform detection only under single gray level and single color conditions, failing to detect latent and triggered muras that only appear under other operating conditions. This invention, by extracting and comparing the differences in static defect components under all gray levels, colors, and conditions, can identify and quantify multi-condition sensitive defects within a unified framework.

[0167] When calculating the grayscale sensitivity index, the root mean square dispersion is used as the metric. First, the average grayscale response under fixed color and state is defined. As a reference center for grayscale changes, this reference center reflects the average Mura level of the panel across grayscale levels under the current color and state; then, the local grayscale sensitivity intensity is defined. This value measures the degree to which the Mura response at each spatial location deviates from the reference center under all grayscale conditions. Using the root mean square instead of the maximum difference provides more robust resistance to the effects of a single anomalous measurement. Finally, a global grayscale sensitivity index is defined. The average value is calculated across all space, color, and state conditions. A higher grayscale sensitivity index indicates a stronger dependence of the panel's Mura on the displayed grayscale, meaning that relying solely on a single grayscale level for quality inspection on the production line may lead to missed detections. In practical applications, if the grayscale sensitivity index is abnormally high, it is recommended to perform independent testing under multiple grayscale conditions.

[0168] When calculating the color sensitivity index, the calculation logic is consistent with that of the grayscale sensitivity index, and the root mean square dispersion measure is also used. First, the average color response under a fixed grayscale and state is defined. Establish a reference baseline for color dimension analysis, and then define the local color sensitivity intensity. Finally, a global color sensitivity index is defined. The color sensitivity index is used to evaluate whether panel defects have obvious color-dependent characteristics. The higher the value, the greater the difference between defects in different color channels, which may involve problems with the driving or luminescent materials of specific sub-pixel channels.

[0169] When calculating the state sensitivity index, since the state set only includes two states—cold start and highlighted pre-excitation—the root mean square dispersion degenerates to half the absolute value of the difference, but its physical meaning remains clear. First, we define the average state response under fixed grayscale and color. Then define the local state sensitivity intensity. Finally, a global state sensitivity index is defined. State sensitivity metrics are used to evaluate whether a panel exhibits condition-triggered murmurs such as cold start sensitivity or high-brightness pre-excitation sensitivity. In practical applications, panels with abnormally high state sensitivity metrics may exhibit temperature-related device characteristic changes or charge storage-related hysteresis effects, requiring comprehensive judgment in conjunction with recovery hysteresis metrics.

[0170] S5.5: Based on the Mura response tensor in the recovered frame sequence under the bright pre-excitation state, fit the recovery process of the defect response over time to obtain the recovery hysteresis index.

[0171] After obtaining the Mura response tensor, it is necessary to further characterize the defect recovery behavior with hysteresis. Specifically, after highlighting pre-excitation, the Mura defect response of the panel undergoes a process from initial overshoot to gradual recovery to a steady state. For this recovery process, the defect response amplitude at each sampling time is calculated based on the Mura response tensor in the recovery frame sequence, and the recovery sequence is obtained by arranging them in chronological order.

[0172] In a preferred embodiment, the recovery sequence can be fitted using an exponential recovery model:

[0173]

[0174] in, Indicates the recovery phase at time [time]. The defect response amplitude, Indicates the initial recovery amplitude. This represents the recovered steady-state bias term. Indicates the start time of the recovery modeling. This represents the recovery time constant.

[0175] The recovery time constant It is used to characterize how quickly a defect response recovers from a highlighted pre-excitation state to a steady-state level. The larger the value, the slower the defect disappears and the higher the degree of recovery hysteresis; The smaller the value, the faster the defect disappears and the lower the recovery hysteresis.

[0176] To avoid diluting the hysteresis analysis results in normal background areas, abnormal regions can be identified based on a preset abnormal region judgment threshold. Recovery time constant statistics are then performed only on pixel positions within these abnormal regions. Finally, the average, median, quantile, or weighted average of the recovery time constants within the abnormal regions are used as the recovery hysteresis index.

[0177] Example 8

[0178] This embodiment, based on Embodiment 7, describes in detail the implementation method of fitting the recovery time constant based on the Mura response tensor in the recovered frame sequence under the bright pre-excitation state, and obtaining the recovery hysteresis index based on the recovery time constant, specifically including:

[0179] S5.5.1: Calculate the defect response amplitude at each sampling time based on the Mura response tensor in the recovered frame sequence, and arrange them in chronological order to obtain the recovered sequence.

[0180] In the recovery frame sequence after the pre-excitation of the highlight, the Mura defect response of the panel undergoes a process from initial overshoot to gradual recovery to a steady state. For this recovery frame sequence, the defect response amplitude at each sampling time is calculated based on the Mura response tensor. The defect response amplitude can be represented by the mean absolute value, root mean square value, or quantile value of the Mura response within the abnormal region.

[0181] Time-sampled set For example, the defect response amplitude was calculated at four time points and arranged in chronological order to form a recovery sequence. This sequence reflects the trend of the defect intensity gradually recovering over time after the highlight pre-excitation.

[0182] S5.5.2: Fit the recovery sequence to an exponential recovery model to obtain the recovery time constant.

[0183] After obtaining the recovery sequence, an exponential recovery model is applied to fit the recovery sequence. The exponential recovery model can be expressed as:

[0184] in, Indicates the recovery phase time. The defect response amplitude, Indicates the initial recovery amplitude. This represents the recovered steady-state bias term. Indicates the start time of the recovery modeling. This represents the recovery time constant.

[0185] S5.5.3: Calculate the recovery hysteresis index based on the recovery time constant.

[0186] To avoid diluting the recovery hysteresis analysis results due to normal background areas, abnormal regions can be identified based on a preset abnormal region threshold, and only the recovery time constant within these abnormal regions will be statistically analyzed. The recovery hysteresis index can be defined as:

[0187] in, Indicates a lag in recovery indicators. Indicates an abnormal region. Indicates the number of pixels in the abnormal area. Represents a grayscale set. Represents a set of colors. Indicates spatial location In grayscale ,color Recovery time constant under certain conditions.

[0188] The recovery hysteresis index It is positively correlated with the recovery time constant. The larger the value, the longer it takes for the defect to recover from the highlighted pre-excitation state to the steady state, and the higher the degree of hysteresis. The smaller the value, the faster the defect recovery and the lower the hysteresis.

[0189] Example 9

[0190] like Figure 5 As shown, this embodiment is a further refinement of step S6 in embodiment 1. The calculation process of inputting the multidimensional response feature index into a preset comprehensive scoring model to obtain the comprehensive Mura score includes:

[0191] S6.1: The time persistence index, spatial intensity index, dynamic fluctuation index, recovery hysteresis index, grayscale sensitivity index, color sensitivity index, and state sensitivity index in the multidimensional response characteristic index are normalized and homogenized so that each index satisfies the condition that the larger the value, the more severe the defect, thus obtaining normalized characteristic indexes; the normalized characteristic indexes are weighted and summed with the corresponding preset weight coefficients to obtain the weighted summation result.

[0192] The time persistence index, spatial intensity index, dynamic fluctuation index, recovery hysteresis index, grayscale sensitivity index, color sensitivity index, and state sensitivity index among the multidimensional response characteristic indicators are normalized and oriented to ensure that each index satisfies the condition that a larger value indicates a more severe defect, thus obtaining normalized characteristic indicators. As described in Example 1, the comprehensive Mura score is obtained by weighted summation of the seven sub-indicators. In specific implementation, since the dimensions and value ranges of each sub-indicator are different, direct weighted summation will lead to the larger-scale index dominating the comprehensive score result. Therefore, before weighted summation, it is necessary to normalize and oriented each sub-indicator. The purpose of normalization is to ensure that the numerical direction of each sub-indicator is consistent, that is, a larger value indicates a more severe defect. For spatial intensity, temporal persistence, dynamic fluctuation, grayscale sensitivity, color sensitivity, and state sensitivity indices, larger values ​​indicate more severe defects, requiring no directional adjustment. For recovery hysteresis indices, larger recovery time constants indicate slower defect recovery and higher hysteresis, conforming to the principle of "larger values ​​indicate more severe defects," thus also requiring no directional adjustment. Normalization employs a linear normalization method, mapping each sub-indicator to the [0,1] interval, defined as... ,in Indicates the first The original values ​​of each sub-indicator and These represent the lower and upper limits of the indicator in historical data or preset specifications, respectively. Normalized feature indicator. The larger the value, the more severe the defect.

[0193] The normalized feature indices are weighted and summed with their corresponding preset weight coefficients to obtain the weighted sum result. The weight allocation is based on the following principle: the spatial intensity index is assigned the highest weight. Because spatial unevenness has the strongest correlation with subjective perception, it is the dimension of greatest concern in production line quality inspection and user evaluation; the recovery hysteresis index is given the second highest weight. Because the potential device risks associated with hysteresis-type Mura-related panels (such as charge leakage or threshold voltage drift in pixel drive circuits) have predictive value for the long-term reliability of the panel; the weight of time-duration indicators This reflects the degree to which persistent defects affect user experience. The dynamic fluctuation index, grayscale sensitivity index, color sensitivity index, and state sensitivity index each have a weight of 0.10, providing supplementary dimensions to offer multi-condition variation information. In specific applications, the weighting coefficients can be adjusted according to panel type and application scenario: for OLED panels, the weight of the recovery hysteresis index can be appropriately increased because the decay characteristics of OLED's organic materials make recovery hysteresis more significant; for LCD panels, the weight of grayscale sensitivity can be appropriately increased because the response differences of LCDs at different grayscale levels are usually more pronounced. The weighted sum is obtained by multiplying each normalized sub-index by its weighting coefficient.

[0194] S6.2: Input the weighted summation result into a preset nonlinear mapping function to obtain the mapped score.

[0195] After obtaining the weighted summation result, it needs to be mapped to a predetermined scoring interval for grade determination. The preset non-linear mapping function can be either piecewise linear mapping or a sigmoid curve mapping. Piecewise linear mapping maps the weighted summation result to a scoring interval of 0 to 100 points according to preset segmentation points and slopes. Sigmoid curve mapping uses logarithmic or arctangent functions, providing higher resolution in the intermediate score intervals, which is beneficial for distinguishing panels of medium quality. The weighted summation result is input into the non-linear mapping function, and the mapped score is output.

[0196] Furthermore, the aforementioned weighted summation and nonlinear mapping combined scoring scheme is a preferred embodiment of the present invention. It should be understood that the calculation method for the comprehensive Mura score is not limited to linear weighted fusion. In optional embodiments, a parametric nonlinear mapping model can also be used to fuse the multidimensional response feature indicators, i.e.:

[0197]

[0198] in, For parameterized nonlinear mapping models, The model parameters are represented. The nonlinear mapping model can take the form of a quadratic model, a gradient boosting tree model, or a multilayer perceptron, and is learned on a dataset labeled with subjective quality scores or process judgment results. Compared with linear weighted models, parametric nonlinear mapping models can more flexibly fit the interaction relationships between various sub-indicators, thereby improving the consistency between comprehensive scores and human evaluations in specific panel types or application scenarios. When using the above-mentioned optional scoring schemes, the extraction method of the multidimensional response feature indicators is consistent with other embodiments of the present invention, only the fusion strategy is different, so it can be flexibly combined with other technical solutions of the present invention.

[0199] S6.3: Based on the mapped score, determine the corresponding level of the comprehensive score according to the preset scoring level threshold, and output the comprehensive Mura score result.

[0200] After obtaining the mapped score, it needs to be converted into a quality level judgment result that can be directly used by the production line. Preset scoring level thresholds divide the scoring range into several levels, for example, four levels: Excellent, Good, Average, and Poor, each corresponding to a different threshold range. Based on the mapped score, it is determined which level range it falls into, and a comprehensive Mura score result is output. The comprehensive Mura score result includes the score value and the corresponding quality level label, which can be used for production line classification, yield control, and subsequent process diagnostics. In addition, the system can also output auxiliary results such as defect heatmaps, time-varying curves, and cross-grayscale response curves for mechanism analysis and compensation strategy formulation during the R&D phase.

[0201] In summary, this invention elevates the traditional static Mura evaluation method based on a single-frame image to a dynamic quantitative evaluation method based on spatiotemporal multidimensional response modeling by constructing a multidimensional response tensor of the display panel across spatial, temporal, grayscale, color, and state dimensions. Through joint analysis of static and dynamic features, it can effectively identify hysteresis-type, fluctuation-type, and condition-triggered defects that are difficult to detect using traditional single-frame methods, improving the accuracy and consistency of the evaluation results. The technical solution of this invention can be flexibly extended by adjusting parameters such as tensor dimensions, background modeling methods, feature indicators, and scoring models, making it suitable for various application scenarios such as production line inspection, laboratory evaluation, and subsequent display compensation optimization.

Claims

1. A method for quantitatively evaluating Mura of a display panel based on modeling of spatio-temporal multi-dimensional responses, characterized in that, include: The test panel is controlled to display the target test images sequentially, and the test panel is image acquired according to a preset time sampling sequence to obtain the original acquired images. The original acquired image is sequentially subjected to dark field correction, flat field correction, geometric registration and exposure normalization to obtain a preprocessed image; Background modeling is performed on the preprocessed image to obtain a reference background. Based on the preprocessed image and the reference background, the normalized residual is calculated, and a Mura response tensor covering the spatial dimension, temporal dimension, grayscale dimension, color dimension and state dimension is constructed. The Mura response tensor is decomposed in the time domain to obtain static defect components and dynamic defect components, and a time fluctuation diagram is calculated based on the dynamic defect components. Multidimensional response features are extracted based on the Mura response tensor, the static defect component, the dynamic defect component, and the time fluctuation diagram to obtain multidimensional response feature indices. The multidimensional response feature indexes are input into a preset comprehensive scoring model to obtain a comprehensive Mura score.

2. The method of claim 1, wherein, The control panel under test sequentially displays the target test image, and the control panel under test acquires images according to a preset time sampling sequence to obtain the original acquired image, including: Traverse all combinations of the grayscale set, color set, and state set, and under each fixed grayscale, color, and state condition, control the display panel under test to continuously display the target test image to obtain the test condition combination; For each of the test condition combinations, multiple frames of images are acquired from the display panel under test according to a preset time sampling sequence to obtain the original acquired image.

3. The method of claim 2, wherein, For each of the test condition combinations, the display panel under test is sampled for multiple frames according to a preset time sampling sequence to obtain the original sampled image, including: The display panel under test is controlled to continuously display a preset cold start duration in the first grayscale state to establish a cold start acquisition state, wherein the grayscale value of the first grayscale is less than a preset excitation grayscale threshold. The display panel under test is switched from the cold start acquisition state to the high brightness pre-excitation state. In the high brightness pre-excitation state, the initial stable frame and the recovery frame sequence are acquired sequentially according to the preset time sampling sequence. The initial stable frame and the recovery frame sequence are used together as the original acquired image.

4. The method of claim 1, wherein, The process of sequentially performing dark field correction, flat field correction, geometric registration, and exposure normalization on the original acquired image to obtain a preprocessed image includes: Subtract the dark field template obtained by averaging multiple frames acquired under shading conditions from the original acquired image to obtain the dark field corrected image; Divide the dark field corrected image by the flat field template obtained by standard uniform light source calibration to obtain the flat field corrected image; The camera coordinate system is mapped to the panel coordinate system using the geometric mapping relationship obtained by fitting the corner points of the calibration plate to the flat field corrected image, thus obtaining the registered image; The registered image is used as input, and exposure normalization processing is performed to obtain the preprocessed image.

5. The method of claim 1, wherein, The process involves performing background modeling on the preprocessed image to obtain a reference background, calculating a normalized residual based on the preprocessed image and the reference background, and constructing a Mura response tensor covering spatial, temporal, grayscale, color, and state dimensions, including: Global polynomial fitting and low-pass filtering were performed on multiple preprocessed images acquired under the same grayscale and color conditions to obtain polynomial background and smooth background. A reference background is obtained by weighted combination of the polynomial background and the smooth background. The difference between the preprocessed image and the reference background is calculated to obtain the original residual; The original residual is divided by the sum of the reference background and the preset stability constant to obtain the normalized residual, which is then used as the Mura response tensor.

6. The method of claim 1, wherein, The step of performing time-domain decomposition on the Mura response tensor to obtain static and dynamic defect components, and calculating a time fluctuation map based on the dynamic defect components, includes: The static defect component is obtained by averaging the Mura response tensor along the time dimension. The dynamic defect component is obtained by calculating the difference between the Mura response tensor and the static defect component. The standard deviation of the dynamic defect component is calculated along the time dimension to obtain the time fluctuation diagram.

7. The method of claim 6, wherein, The multidimensional response feature indicators include time persistence indicators, spatial intensity indicators, dynamic fluctuation indicators, grayscale sensitivity indicators, color sensitivity indicators, state sensitivity indicators, and recovery hysteresis indicators. The multidimensional response feature extraction based on the Mura response tensor, the static defect component, the dynamic defect component, and the time fluctuation map yields the multidimensional response feature indicators, including: Based on the absolute response value of the Mura response tensor at each sampling time and the preset visibility threshold, it is determined whether each pixel position is in a visible defect state at each sampling time, and the proportion of time that each pixel position is in a visible defect state under a given gray level, color and state condition is calculated to obtain the time persistence index. Calculate the defect intensity of the static defect component in the spatial dimension to obtain the spatial intensity index; The global mean of the time fluctuation graph is calculated to obtain the dynamic fluctuation index; The degree of difference of the static defect component under different gray levels, different color channels, and different display states is calculated respectively to obtain the gray level sensitivity index, the color sensitivity index, and the state sensitivity index. Based on the Mura response tensor in the recovered frame sequence under the bright pre-excitation state, the recovery process of the defect response over time is fitted to obtain the recovery hysteresis index.

8. The method of claim 7, wherein, The process involves constructing a recovery sequence of defect response amplitude over time based on the Mura response tensor in the recovered frame sequence after highlight pre-excitation, fitting the recovery sequence with an exponential recovery model to obtain a recovery time constant, and calculating the recovery hysteresis index based on the recovery time constant. Specifically, this includes: After the highlight pre-excitation state, the defect response amplitude at each sampling time is calculated based on the Mura response tensor in the recovery frame sequence, and the recovery sequence is obtained by arranging them in chronological order; The recovery sequence was fitted with an exponential recovery model to obtain the recovery time constant; The statistical value of the recovery time constant or the recovery time constant within the abnormal region is used as the recovery hysteresis index. The larger the recovery time constant, the slower the defect recovery and the higher the degree of recovery hysteresis.

9. The method of claim 7, wherein, The step of inputting the multidimensional response feature indexes into a preset comprehensive scoring model to obtain a comprehensive Mura score includes: The time persistence index, spatial intensity index, dynamic fluctuation index, recovery hysteresis index, grayscale sensitivity index, color sensitivity index, and state sensitivity index in the multidimensional response feature index are normalized and homogenized so that each index satisfies the condition that the larger the value, the more severe the defect, thus obtaining normalized feature indexes; the normalized feature indexes are then weighted and summed with the corresponding preset weight coefficients to obtain the weighted summation result. The weighted summation result is input into a preset nonlinear mapping function to obtain the mapped score; Based on the mapped score, the corresponding level of the comprehensive score is determined according to the preset scoring level threshold, and the comprehensive Mura score result is output.

Images