Mini LED direct display lighting defect intelligent detection system based on image recognition

Abstract

This invention relates to the fields of machine vision and optoelectronic display inspection technology, specifically to an intelligent detection system for MiniLED direct-light display lighting defects based on image recognition. The system includes: an image acquisition unit for acquiring surface-emitting images of a MiniLED direct-light display device as initial image data; a light field potential energy initialization unit for receiving the initial image data, extracting the pixel brightness matrix, and converting the pixel brightness matrix into a floating-point light field potential energy map; a light field divergence decoupling unit for generating a light field divergence distribution map; an intrinsic divergence stripping unit for loading a pre-calibrated ideal point spread function convolution kernel of a defect-free emitting unit, simultaneously generating a stripping residual distribution map containing sub-pixel level distortion data; and a defect classification and compensation closed-loop unit for performing conditional classification based on the intrinsic divergence peak matrix and outputting a structured defect correction mask map and the optical compensation reference matrix. This invention effectively avoids the defect coordinate drift problem in traditional solutions.

Description

Technical Field

[0001] This invention relates to the fields of machine vision and optoelectronic display inspection technology, specifically to an intelligent detection system for MiniLED direct-display lighting defects based on image recognition. Background Technology

[0002] In the current MiniLED direct-view module production line testing environment, a high-density light-emitting array will generate a complex surface light-emitting image after being powered on, and there is serious light field overlap and crosstalk between adjacent light-emitting units.

[0003] To detect illumination defects, existing image processing solutions generally employ methods such as extracting two-dimensional brightness thresholds, connected components, or texture edge features for morphological recognition. While this approach has some discrimination capability in low-density scenes, it only focuses on the two-dimensional brightness representation of discrete pixels and fails to establish a structured mapping from image space to device physical space. Consequently, when faced with strong light flow crosstalk from neighboring areas, it is highly prone to misjudging illuminated dead light areas as normal units, or misjudging the halo diffusion of normal units as leakage bright spots. Furthermore, such solutions struggle to extract sub-pixel-level distortion data within the same processing chain to support subsequent image correction.

[0004] Therefore, how to eliminate local optical crosstalk to accurately distinguish between real self-luminous anomalies and pseudo-defects, and simultaneously extract optical compensation reference data to form a closed loop, has become an urgent technical problem to be solved. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent detection system for MiniLED direct-display lighting defects based on image recognition, and to solve the following technical problems:

[0006] The original point detection, which focused on two-dimensional brightness, is transformed into an image processing process based on the dynamic properties of the light field. By calculating the light field divergence and performing nonlinear inverse filtering, the influence of optical flow crosstalk caused by adjacent light-emitting units is effectively removed, thereby avoiding misjudgment in the case of strong crosstalk in the neighborhood halo. It can accurately distinguish between real self-illumination anomalies and false bright spots or false dark spots formed by the influence of the neighborhood. It can also simultaneously output a structured defect correction mask map and an optical compensation reference matrix, realizing closed-loop linkage between defect identification, rework location, and image uniformity compensation.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] The image recognition-based intelligent detection system for MiniLED direct-display lighting defects includes:

[0009] Image acquisition unit: used to acquire surface-emitting images of the MiniLED direct-view display device as initial image data;

[0010] Light field potential energy initialization unit: receives the initial image data, establishes a topological mesh model of the luminescent array; performs geometric distortion correction on the initial image data based on the topological mesh model of the luminescent array, and extracts the pixel brightness matrix; converts the pixel brightness matrix into a floating-point light field potential energy map;

[0011] Optical field divergence decoupling unit: Analyzes the floating-point optical field potential energy map and calculates the gradient vector field; calculates the divergence matrix based on the gradient vector field and generates an optical field divergence distribution map;

[0012] Intrinsic divergence stripping unit: Load the ideal point spread function convolution kernel of the pre-calibrated defect-free luminescent unit, and perform nonlinear inverse filtering calculation with the light field divergence distribution map; strip neighborhood optical flow crosstalk based on the nonlinear inverse filtering calculation result, and extract the intrinsic divergence peak matrix; simultaneously generate a stripping residual distribution map containing sub-pixel level distortion data;

[0013] Defect classification and compensation closed-loop unit: Performs conditional classification based on the intrinsic divergence peak matrix to generate defect category labels; maps the defect category labels to the light-emitting array topology mesh model to generate a structured defect correction mask; generates an optical compensation reference matrix based on the stripped residual distribution map, and outputs the structured defect correction mask and the optical compensation reference matrix to the display control terminal, so that it can shield defective units based on the structured defect correction mask and perform brightness compensation on normal light-emitting units based on the optical compensation reference matrix.

[0014] Furthermore, the method for receiving the initial image data and establishing the topological mesh model of the luminescent array includes:

[0015] Obtain the initial image data;

[0016] Extract the theoretical arrangement parameters of the MiniLED direct-view device;

[0017] A two-dimensional coordinate system is constructed based on the theoretical arrangement parameters to generate the topological mesh model of the light-emitting array.

[0018] Furthermore, geometric distortion correction is performed on the initial image data based on the light-emitting array topological mesh model, and the pixel brightness matrix is ​​extracted; the method of converting the pixel brightness matrix into a floating-point light field potential energy map includes:

[0019] Extract edge localization features from the initial image data;

[0020] The edge localization features are spatially aligned with the topological mesh model of the light-emitting array, and the affine transformation parameters are calculated.

[0021] Based on the affine transformation parameters, the initial image data is spatially resampled to generate corrected image data;

[0022] Extract the grayscale value of each pixel in the corrected image data to construct the pixel brightness matrix;

[0023] The discrete grayscale values ​​in the pixel brightness matrix are mapped to continuous potential energy height values ​​to generate the floating-point optical field potential energy map.

[0024] Further, the floating-point optical field potential energy map is analyzed, and the gradient vector field is calculated; based on the gradient vector field, the divergence matrix is ​​calculated, and the optical field divergence distribution map is generated in the following ways:

[0025] Spatial differentiation is performed on the floating-point optical potential energy map to calculate the partial derivatives of each pixel in the horizontal and vertical directions;

[0026] The gradient vector field is generated by combining the partial derivatives in the horizontal and vertical directions.

[0027] The second-order spatial derivative of the gradient vector field is calculated by applying the Laplacian operator to obtain the divergence matrix; wherein the Laplacian operator adopts a four-neighborhood difference template to quantify the energy source and sink intensity of the current pixel relative to its neighborhood.

[0028] The divergence matrix is ​​mapped to a two-dimensional image format to generate the light field divergence distribution map.

[0029] Furthermore, the process of applying the Laplacian operator to the gradient vector field to calculate the second spatial derivative and obtain the divergence matrix includes the following steps:

[0030] Perform a Fast Fourier Transform on the gradient vector field to convert the spatial domain data to the frequency domain;

[0031] Perform divergence matrix calculation in the frequency domain to obtain the frequency domain divergence matrix;

[0032] Perform an inverse fast Fourier transform on the frequency domain divergence matrix to restore it to the spatial domain and obtain the divergence matrix.

[0033] Further, a pre-calibrated ideal point spread function convolution kernel of a defect-free emitting unit is loaded, and nonlinear inverse filtering calculation is performed with the light field divergence distribution map; based on the nonlinear inverse filtering calculation result, neighborhood optical flow crosstalk is removed, and the intrinsic divergence peak matrix is ​​extracted; the method for simultaneously generating a stripped residual distribution map containing sub-pixel level distortion data includes:

[0034] Retrieve the ideal point spread function convolution kernel of the pre-calibrated defect-free light-emitting unit;

[0035] The light field divergence distribution map is deconvolved with the ideal point spread function kernel to obtain the nonlinear inverse filtering calculation result.

[0036] Based on the nonlinear inverse filtering calculation results, the overflow optical flow data between the light-emitting units is deducted by matrix subtraction operation, the local intrinsic divergence peak value of each light-emitting unit is extracted, and the intrinsic divergence peak value matrix is ​​constructed.

[0037] Record the overflow optical flow data that is subtracted during the matrix subtraction operation to generate the stripping residual distribution map containing subpixel-level distortion data.

[0038] Furthermore, the defect category labels are generated by performing conditional classification based on the intrinsic divergence peak matrix in the following ways:

[0039] Extract the eigendivergence peak value of each luminescent unit in the eigendivergence peak matrix;

[0040] The intrinsic divergence peak value is compared with a preset critical divergence threshold and a preset extreme divergence threshold, wherein the preset extreme divergence threshold is greater than the preset critical divergence threshold.

[0041] The configuration is as follows: if the intrinsic divergence peak value is greater than the preset critical divergence threshold and less than the preset extreme divergence threshold, the light-emitting unit is determined to be normal, and a normal category label is generated; if the intrinsic divergence peak value is less than or equal to the preset critical divergence threshold, the light-emitting unit is determined to have a non-luminous defect, and a dead light dark spot category label is generated; the above thresholds are obtained based on statistical sampling of the intrinsic divergence distribution of defect-free benchmark samples; if the intrinsic divergence peak value is greater than or equal to the preset extreme divergence threshold, the light-emitting unit is determined to have a leakage defect, and a puncture leakage category label is generated.

[0042] The normal category label, the dead light dark spot category label, and the puncture leakage category label are integrated to generate the defect category label.

[0043] Further, the defect category labels are mapped to the luminescent array topology mesh model to generate a structured defect correction mask; the optical compensation reference matrix is ​​generated based on the stripping residual distribution map, and the structured defect correction mask and the optical compensation reference matrix are output in the following ways:

[0044] Extract the two-dimensional coordinates of the light-emitting unit corresponding to the defect category label;

[0045] Spatially bind the two-dimensional coordinates of the light-emitting unit to the topological mesh model of the light-emitting array to generate the structured defect correction mask map;

[0046] The subpixel-level distortion data is the residual value in the stripped residual distribution map that reflects the offset of the luminescence center or the asymmetric diffusion characteristics; the subpixel-level distortion data in the stripped residual distribution map is extracted.

[0047] The inverse compensation parameters are calculated based on the subpixel-level distortion data to generate the optical compensation reference matrix;

[0048] The structured defect correction mask and the optical compensation reference matrix are output synchronously and sent to the display control terminal for image compensation.

[0049] The beneficial effects of this invention are:

[0050] 1. This invention establishes a topological mesh model of the light-emitting array and performs geometric distortion correction on the initial image to extract the pixel brightness matrix. This mechanism eliminates the spatial misalignment between the imaging coordinates and the device coordinates, provides a unified and stable structured coordinate basis for subsequent calculations, and successfully achieves a consistent mapping from the image space to the device physical space, effectively avoiding the defect coordinate drift problem in traditional schemes.

[0051] 2. This invention converts discrete brightness into a floating-point optical field potential energy map and calculates the gradient vector field and divergence matrix to generate a distribution map. This process transforms the illumination detection into the calculation of the dynamic properties of the optical field. By quantifying the intensity of local energy sources and sinks, it is possible to directly distinguish the real source from the passively illuminated area from the structure of image intensity changes, breaking through the limitations of traditional morphological recognition and brightness threshold.

[0052] 3. This invention utilizes the ideal point spread function convolution kernel to perform nonlinear inverse filtering calculations, accurately deducting the overflow optical flow data between light-emitting units to extract the intrinsic divergence peak value; this completely separates the neighborhood coupling component from the unit's own light-emitting component, accurately eliminating pseudo-defect interference in high-density scenes with severe halo crosstalk, and achieving objective and accurate classification of dead lamps and leakage current by combining thresholds.

[0053] 4. This invention simultaneously records the subtracted optical flow data during crosstalk stripping to generate a stripping residual distribution map, and calculates the optical compensation reference matrix and structured mask map accordingly. This not only provides a direct output of structured abnormal coordinates for in-plant rework positioning, but also transforms the difficult-to-detect subpixel level distortion into reverse compensation parameters, achieving a perfect closed loop of defect identification and image uniformity compensation in the same processing link. Attached Figure Description

[0054] The invention will now be further described with reference to the accompanying drawings.

[0055] Figure 1 This is a module architecture diagram of the MiniLED direct-display lighting defect intelligent detection system based on image recognition in the embodiments of this application. Detailed Implementation

[0056] 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.

[0057] Please see Figure 1 The image recognition-based intelligent detection system for MiniLED direct-display lighting defects includes: an image acquisition unit, used to acquire surface luminescent images of the MiniLED direct-display device as initial image data;

[0058] The light field potential energy initialization unit receives the initial image data and establishes a topological mesh model of the luminescent array; performs geometric distortion correction on the initial image data based on the topological mesh model of the luminescent array and extracts the pixel brightness matrix; and converts the pixel brightness matrix into a floating-point light field potential energy map.

[0059] The optical field divergence decoupling unit analyzes the floating-point optical field potential energy map and calculates the gradient vector field; based on the gradient vector field, it calculates the divergence matrix and generates an optical field divergence distribution map.

[0060] The intrinsic divergence stripping unit is loaded with a pre-calibrated ideal point spread function convolution kernel of a defect-free emitting unit and subjected to nonlinear inverse filtering calculation with the light field divergence distribution map. Based on the nonlinear inverse filtering calculation result, neighborhood optical flow crosstalk is stripped and the intrinsic divergence peak matrix is ​​extracted. Simultaneously, a stripping residual distribution map containing sub-pixel level distortion data is generated.

[0061] The defect classification and compensation closed-loop unit performs conditional classification based on the intrinsic divergence peak matrix to generate defect category labels; maps the defect category labels to the light-emitting array topology mesh model to generate a structured defect correction mask; generates an optical compensation reference matrix based on the stripped residual distribution map, and outputs the structured defect correction mask and the optical compensation reference matrix to the display control terminal, so that it can shield defective units based on the structured defect correction mask and perform brightness compensation on normal light-emitting units based on the optical compensation reference matrix.

[0062] This embodiment provides an intelligent detection mechanism for MiniLED direct-display lighting defects based on image recognition. Specifically, the mechanism is deployed at the power-on lighting detection station of the MiniLED direct-display module production line. After the entire board is powered on, it takes an image of the display panel under test, performs a calculation, and simultaneously outputs the defect location result and the data benchmark required for subsequent compensation.

[0063] The entire main scenario takes a MiniLED direct-display module used in a conference display terminal as the object. After the module is packaged, mounted and initially lit, it enters the online testing stage. Without relying on a deep neural network classifier, the system identifies dead light dark spots, leakage bright spots and local optical unevenness based solely on the spatial processing results of image data.

[0064] Specifically, the image acquisition unit acquires the surface luminous image of the panel under test under a preset illumination pattern; the preset illumination pattern can be a full white, full gray, or a pattern of uniform illumination in different areas. In this embodiment, a full-screen illumination mode with a uniform grayscale is preferred to keep each luminous unit in a nearly consistent driving state; the acquired original image can be a RAW format image or a grayscale image after linear response correction, which essentially records the light intensity distribution at each position on the panel surface.

[0065] The light field potential energy initialization unit maps the initial image data to the luminescent array topological mesh model; here, the topological mesh model is not an abstract graph structure, but a two-dimensional array coordinate frame consistent with the theoretical arrangement of the panel.

[0066] The two-dimensional array coordinate framework not only includes the logical row and column indexes of each light-emitting unit, but also, based on the actual physical engineering design drawings of the MiniLED direct display device, is bound with physical measurement attributes including the absolute physical coordinates of the center point of each light-emitting unit and the theoretical light-emitting area of ​​the light-emitting unit, so as to support the subsequent sub-pixel level spatial physical quantization of the distortion data.

[0067] For a regularly arranged MiniLED module, each theoretical light-emitting unit has a unique coordinate in the grid; the system uses this coordinate framework as a unified benchmark for subsequent pixel assignment, defect location, and compensation write-back.

[0068] After completing the grid anchoring, the system performs geometric distortion correction on the initial image and extracts the pixel brightness matrix from the correction result. For ease of explanation, a local detection matrix example can be constructed: assume that the corrected local area corresponds to a 3×3 pixel brightness matrix, and the matrix values ​​are ([12,18,13;17,80,20;11,19,12]); here the brightness of the center position is higher, and there is secondary brightness in the surrounding position affected by the diffusion of the neighborhood.

[0069] The system does not directly judge that it is normal based on 80 being higher than the threshold. Instead, it maps each discrete gray value to a continuous floating-point value to form a potential energy map of the light field.

[0070] This can be understood as treating each pixel as a sampling point on the potential energy surface; the higher the brightness, the higher the corresponding potential energy height. After this transformation, the above matrix can be represented as a floating-point potential energy map, for example ([0.12,0.18,0.13;0.17,0.80,0.20;0.11,0.19,0.12]). The reason for this processing is that traditional shape recognition is no longer performed subsequently; instead, the gradient and divergence calculation process of the continuous field is entered.

[0071] The optical field divergence decoupling unit receives the floating-point optical field potential energy map and performs spatial differentiation on it; first, it calculates the variation trend of each pixel in the horizontal and vertical directions to obtain the gradient vector field, and then generates the divergence matrix according to the local divergence degree of the gradient vector field.

[0072] Taking the aforementioned local 3×3 region as an example, if the central pixel is a real light source, its surrounding brightness should exhibit a diffusion pattern that decreases from the center outwards. After gradient and divergence calculations, the central position will show a significant positive divergence peak.

[0073] If the location itself does not emit light, but is only slightly bright due to the superposition of halos from surrounding units, then its local variation relationship is closer to a smooth transition or even a convergence mode, and the resulting divergence value may be close to zero or negative; thus, the system does not generate a visual diagram of whether it is bright or not, but rather a light field divergence distribution diagram of whether it constitutes a light source.

[0074] The intrinsic divergence stripping unit introduces a pre-calibrated ideal point spread function convolution kernel of the defect-free emitting unit to perform nonlinear inverse filtering on the divergence distribution map; its function is to strip away the crosstalk effect caused by adjacent emitting units.

[0075] For ease of explanation, assume that the observed values ​​of three consecutive emitting units in a certain row on the divergence plot are ([0.20, 0.65, 0.22]). According to the ideal diffusion kernel obtained by the calibration of defect-free samples, if the middle unit is a real normal source, it will theoretically leak 0.10 divergence contribution to the left and right sides respectively. Then, after inverse filtering, the system will deduct the corresponding crosstalk components in the left and right positions.

[0076] For example, in the observation value of 0.20 at the left position, if 0.09 overflows from the middle unit, its remaining eigenvalue is about 0.11; after subtracting the weak back coupling from the left and right positions from the 0.65 at the middle position, about 0.60 can be retained as the eigenvalue divergence peak; the same applies to the right side; thus, the eigenvalue divergence peak matrix is ​​constructed; at the same time, the part stripped off during the subtraction process is not discarded, but is recorded as the stripping residual distribution map, which reflects the local subpixel level optical distortion and crosstalk residue of the screen;

[0077] The defect classification and compensation closed-loop unit performs conditional classification based on the intrinsic divergence peak matrix; for example, for four light-emitting units in a certain local array, the intrinsic divergence peaks are 0.58, 0.05, 0.62 and 0.95, respectively, and the system can set the critical divergence threshold to 0.10 and the extreme divergence threshold to 0.85.

[0078] 0.58 and 0.62 fall within the normal range and correspond to the normal category; 0.05 is not higher than the critical threshold and corresponds to the dead light dark spot category; 0.95 is not lower than the extreme value threshold and corresponds to the puncture leakage category; then, the system remaps the category labels back to the luminescent array topology mesh model and generates a structured defect correction mask with mesh coordinate index.

[0079] If the coordinates of the aforementioned four units in the grid are (10,20), (10,21), (10,22), and (10,23) respectively, then (10,21) can be marked as a dead light dark spot and (10,23) as a puncture leakage in the output mask image; at the same time, the system calculates the optical compensation reference matrix based on the stripping residual distribution map, which is used for the subsequent display control terminal to perform Demura-type compensation processing;

[0080] As an anomaly handling mechanism, if local overexposure, underexposure, occlusion, or missing frames occur during image acquisition, the system can first mark the abnormal area as valid and temporarily suspend the intrinsic divergence classification in that area, only outputting the label to be re-examined.

[0081] If, after geometric correction, a local area is found to be unable to stably align with the grid, the original image index of that area is retained and repeated shooting is triggered; if abnormal spikes appear in the deconvolution results at certain locations during the inverse filtering process, exceeding the preset upper limit, truncation or neighborhood smoothing can be used to prevent single-point anomalies from spreading to the entire classification result.

[0082] For example, in the testing of sub-millimeter LED direct display modules in a conference display terminal, the entire module under test is powered on and placed at the testing station, and an industrial camera captures a full-gray point illumination image. After aligning the image to the theoretical grid, the system obtains a full-screen floating-point optical field potential energy map. After divergence decoupling, it is found that although the brightness of a certain area in the upper right region of the module is not significantly lower than the surrounding area, its divergence value is close to zero. After inverse filtering and stripping, the intrinsic divergence peak value further decreases and is finally marked as a dead LED dark spot. In addition, a local abnormally sharp divergence peak is found in the middle region. Although the original gray level is only slightly higher, the intrinsic peak value exceeds the extreme value threshold and is marked as puncture leakage. The system synchronously outputs the error correction mask map of the coordinates of the two locations, as well as the optical compensation reference matrix corresponding to the whole board residual map, and sends it to the display control terminal for trial display compensation.

[0083] The purpose of this step is to transform the original point detection, which was based on the two-dimensional brightness appearance, into an image data processing process based on the dynamic properties of the light field. In this way, even in the presence of neighborhood halo crosstalk, it is still possible to distinguish between real self-illumination anomalies and false bright spots or false dark spots formed by neighborhood influences, and further link the detection results and compensation results in the same processing link for output.

[0084] In a preferred embodiment of the present invention, the method of receiving the initial image data and establishing the topological mesh model of the light-emitting array includes: acquiring the initial image data; extracting the theoretical arrangement parameters of the MiniLED direct display device; constructing a two-dimensional coordinate system based on the theoretical arrangement parameters; and generating the topological mesh model of the light-emitting array.

[0085] This embodiment provides a mechanism for establishing a topological mesh model of a light-emitting array. Specifically, in the above-mentioned whole-board lighting detection scenario, if the subsequent analysis relies solely on the acquired image itself without first establishing a theoretical coordinate framework consistent with the actual device arrangement, the system can find abnormal bright and dark areas in the image space, but it is difficult to stably map defects to specific lamp positions. Especially when the panel has slight rotation, perspective tilt, or boundary trimming, the defect results are prone to coordinate drift. Therefore, this embodiment establishes a topological mesh model of the light-emitting array in combination with device design parameters before the image enters the potential energy mapping.

[0086] Specifically, the system acquires initial image data and simultaneously reads the theoretical arrangement parameters of the MiniLED direct-view display device under test. The theoretical arrangement parameters can be derived from product engineering documents, dot matrix layout lists, or control drive configuration tables, and at least include the number of rows and columns of the light-emitting units, row spacing, column spacing, starting reference point position, and arrangement direction. If it is a standard rectangular array, the system can directly use a regular grid. If it is an irregular array with seams, chamfers, or local gaps, a disabled coordinate position is set on the basis of the regular grid.

[0087] For ease of explanation, assume that a certain area to be tested theoretically contains 4 rows and 5 columns of light-emitting units, and the row spacing and column spacing are both normalized to 1 grid unit. Then the two-dimensional coordinate system generated by the system can be represented as a set of discrete coordinate points from (1,1) to (4,5). If the third row and fifth column is a structural clearance area and there are no actual lamp beads, then this coordinate can be pre-marked as an invalid node. In this way, no matter whether there is optical interference in the image, the system will not misjudge it as a real defective lamp position.

[0088] Furthermore, the system transforms the theoretical arrangement parameters into a topological mesh model that can be used for image processing. This model records at least two types of information: one is the two-dimensional coordinate index of each theoretical luminescent unit; the other is the adjacency relationship between adjacent units. For example, in a 5-column arrangement, the direct neighbors of coordinate (2,3) can be defined as (2,2), (2,4), (1,3), and (3,3). If stronger crosstalk needs to be considered, it can also be extended to diagonal neighbors. This adjacency relationship is not used for abstract graph reasoning, but rather to limit the effective influence range for subsequent crosstalk stripping.

[0089] As a fallback, if the theoretical arrangement parameters are missing or there are version inconsistencies, the system can first make a rough array estimation based on the global bright spot distribution in the image to form a temporary grid model, and mark the detection result as low confidence.

[0090] If the theoretical number of rows and columns read is significantly different from the effective light-emitting area that can be detected in the image, for example, the theoretical value is 100×100 but the image only covers 80×80, the system should output a message indicating insufficient coverage and limit the establishment of a local grid within the visible area; if the device is a splicing screen and there are different offsets between modules, a local coordinate system can be established for each sub-module and then attached to the overall coordinate frame of the whole screen.

[0091] For example, in the aforementioned testing station of the terminal module, the engineering document indicates that the module contains 64 rows and 96 columns of light-emitting units, with the starting reference point located at the upper left corner and the standard pitch fixed; after reading this information, the system establishes a 64×96 two-dimensional coordinate system and shields a small section of invalid coordinates in the reserved drive via area on the right.

[0092] In this way, when the intrinsic divergence anomaly in row 32 and column 48 is subsequently identified, the system can directly locate it as the corresponding lamp position in the actual module, instead of just outputting the pixel position in the image;

[0093] The purpose of this step is to provide a unified and stable structured coordinate basis for subsequent geometric correction, crosstalk stripping, defect write-back, and compensation output, thereby achieving a consistent mapping from image space to device physical space.

[0094] In a preferred embodiment of the present invention, geometric distortion correction is performed on the initial image data based on the light-emitting array topological mesh model, and a pixel brightness matrix is ​​extracted; the method of converting the pixel brightness matrix into a floating-point light field potential energy map includes: extracting edge localization features in the initial image data; spatially aligning the edge localization features with the light-emitting array topological mesh model, and calculating affine transformation parameters;

[0095] Spatial resampling is performed on the initial image data based on the affine transformation parameters to generate corrected image data; the gray value of each pixel in the corrected image data is extracted to construct the pixel brightness matrix; the discrete gray values ​​in the pixel brightness matrix are mapped to continuous potential energy height values ​​to generate the floating-point optical field potential energy map.

[0096] This embodiment provides a processing step for geometric distortion correction and optical field potential energy initialization. Specifically, after the theoretical grid has been established in the previous stage, if the viewing angle deviation, slight lens tilt and panel placement error in the original image are not corrected, the energy distribution of the same light-emitting unit will span multiple theoretical grids, resulting in misalignment in the subsequent divergence calculation.

[0097] Especially in high-density MiniLED scenarios, the imaging area corresponding to a single lamp position is already small, and even a small deviation may cause the crosstalk stripping result to be distorted. Therefore, this embodiment introduces edge localization features and affine transformation parameter calculation mechanism to resample and correct the initial image, and convert the discrete grayscale into a floating-point optical field potential energy map.

[0098] Specifically, the system first extracts edge localization features from the initial image; the edges here are not limited to the edges of a single LED bead, but more commonly the outer boundary of the panel frame, the seam line of the module, the periodic brightness transition zone of the regular dot matrix, or manually set positioning markers; the extraction method can adopt conventional image processing methods such as grayscale gradient enhancement, edge detection or corner point extraction; the extracted edge localization features are used as the reference structure on the image side;

[0099] The edge localization features are spatially aligned with the theoretical mesh model, and the affine transformation parameters are calculated. This can be understood as the system finding a set of parameters that make the reference boundary in the image coincide with the theoretical coordinate frame as much as possible.

[0100] Specifically, the process of solving the affine transformation parameters adopts the least squares method or random sample consensus algorithm. Based on the image coordinates of the extracted edge localization features and their corresponding theoretical coordinates in the light-emitting array topology mesh model, an error equation is constructed. The optimal affine transformation parameter matrix containing translation, scaling, rotation and slight shear components is solved by minimizing the reprojection error.

[0101] For ease of explanation, assume the coordinates of the four corner points of the theoretical grid are (0,0), (0,100), (200,0), and (200,100), while the corresponding four points detected in the image are (5,8), (3,108), (206,4), and (202,106). The system can then calculate the affine parameters, which include translation, scaling, rotation, and slight shearing. After spatial resampling of the entire image using these parameters, the positions of the light array in the image will be pulled back onto the theoretical grid.

[0102] After generating the corrected image data, the system extracts the gray value of each pixel and constructs a pixel brightness matrix. In order to ensure the stability of subsequent continuous differentiation, the gray value is preferably subjected to black level subtraction and quantization normalization processing.

[0103] For example, in an 8-bit grayscale image, if the original value of a pixel is 135 and the black level bias is 7, then the normalized value after correction can be expressed as ((135-7) / 248); thus, the entire image forms a pixel brightness matrix with continuous values ​​that can directly participate in differential operations;

[0104] The system maps the pixel brightness matrix to a floating-point optical field potential energy map; here, the potential energy does not require the introduction of additional physical measurements, but rather the brightness matrix is ​​regarded as the height value on a two-dimensional continuous surface;

[0105] Taking a local 3×3 region as an example, if the corrected grayscale matrix is ​​([110,125,118;123,180,130;115,126,119]), after normalization we can obtain ([0.44,0.50,0.47;0.49,0.72,0.52;0.46,0.51,0.48]). The system will no longer use 180 as the main basis, but will use the local change trend of this continuous potential energy surface as the analysis object.

[0106] As a fallback, if the image edge localization feature extraction fails, for example, due to partial occlusion of the panel outline or image blurring that makes it impossible to stably extract more than four effective reference points, the system can degenerate into a grid fitting method based on the center of the periodic point array.

[0107] If the affine solution residual is greater than the preset threshold, it indicates that the current image may have more complex perspective distortion or warping, which can trigger a higher-order geometric correction process, or only correct the local sub-region in blocks; if some pixels fall into the invalid region after resampling, neighborhood interpolation can be used to fill in the gaps and the confidence of that region can be lowered.

[0108] For example, during the detection process of the aforementioned conference display terminal module, due to the slight rotation of about 1 degree when the module was placed by the operating robot, the dot matrix in the lower right corner of the original image was slightly stretched.

[0109] After extracting positioning features from the module outline and edge light-emitting strip, the system aligns them with a 64×96 theoretical grid, calculates the affine transformation parameters, and completes resampling. In the resampled image, the light spots around each theoretical lamp position are realigned to the fixed grid area and transformed into a floating-point light field potential energy map, which serves as the input for the next stage of divergence calculation.

[0110] The purpose of this step is to eliminate the spatial misalignment between the imaging coordinates and the device coordinates, and to transform the discrete grayscale data affected by imaging quantization into a potential energy expression suitable for continuous field analysis, thereby ensuring that the subsequent divergence calculation has a stable spatial basis.

[0111] In a preferred embodiment of the present invention, the method of parsing the floating-point optical potential energy map and calculating the gradient vector field, and generating the optical field divergence distribution map based on the gradient vector field includes: performing spatial differentiation on the floating-point optical potential energy map to calculate the partial derivatives of each pixel in the horizontal and vertical directions; combining the partial derivatives in the horizontal and vertical directions to generate the gradient vector field; applying the Laplacian operator to the gradient vector field to calculate the second-order spatial derivative to obtain the divergence matrix; wherein the Laplacian operator uses a four-neighborhood difference template to quantify the energy source and sink intensity of the current pixel relative to its neighborhood; and mapping the divergence matrix into a two-dimensional image format to generate the optical field divergence distribution map.

[0112] This embodiment provides a divergence decoupling processing step based on a floating-point optical field potential energy map. Specifically, after completing image correction and potential energy initialization, if the traditional brightness threshold, connected component, or texture edge method is still used for defect identification, the system is prone to misidentifying illuminated dead light areas as normal light-emitting units or misjudging the halo expansion of normal units as leakage bright spots when faced with strong crosstalk in the neighborhood.

[0113] To address this issue, this embodiment does not directly classify brightness values, but instead performs gradient and divergence calculations on the potential energy map to identify the characteristics of the real light source from the local energy change structure of the light field.

[0114] Specifically, the system performs spatial differentiation on the floating-point optical field potential energy map, and calculates the partial derivatives of each pixel in the horizontal and vertical directions respectively;

[0115] This can be achieved using a differential method. For example, for the partial derivative in the horizontal direction at a certain position, the potential energy of the right neighbor pixel is subtracted from the potential energy of the left neighbor pixel; for the partial derivative in the vertical direction, the potential energy of the lower neighbor pixel is subtracted from the upper neighbor pixel. To illustrate the data flow, a simplified 3×3 potential energy region can still be used: ([0.44,0.50,0.47;0.49,0.72,0.52;0.46,0.51,0.48]).

[0116] For a center position of 0.72, if the left and right neighbors are 0.49 and 0.52, the horizontal change can be expressed as a decrease from the center to both sides; if the upper and lower neighbors are 0.50 and 0.51, the vertical change also shows a decreasing trend; based on this, the system generates a gradient vector representing the intensity and direction of local change for the center position.

[0117] The partial derivatives of the horizontal and vertical directions are combined to form a gradient vector field; in other words, each pixel in the entire image no longer corresponds to only a brightness value, but to a two-dimensional directional quantity; for real light sources, the gradient vectors of their surrounding pixels usually exhibit outward divergence; while for flat areas illuminated only by their neighbors, the gradient directions often cancel each other out or tend to be flat; the system applies the Laplacian operator to the gradient vector field to perform second-order spatial derivative operations to obtain the divergence matrix.

[0118] Specifically, the calculation of the second-order spatial derivative refers to differentiating the horizontal partial derivative component along the horizontal direction and the vertical partial derivative component along the vertical direction, and then summing these two second-order partial derivative results. Since calculating the divergence of the gradient vector field is mathematically equivalent to applying the Laplace operator to the original potential energy distribution map, the divergence matrix can be obtained through the above component summation operation. In the specific business logic of the discrete pixel grid, the divergence value at the center pixel position is based on the formula:

[0119]

[0120] In this formula, the divergence value at coordinate ((x,y)) is denoted as (D(x,y)), and the potential energy value at the corresponding position is denoted as (P(x,y)). This formula is equal to the sum of the potential energy values ​​of the four directly adjacent pixels above, below, left, and right, minus four times the potential energy value of the center pixel itself. This is used to quantify the energy source and sink intensity of the center node relative to its neighborhood, thereby avoiding opaque calculation processes. This matrix essentially reflects whether a point is a local source or sink.

[0121] To illustrate this more intuitively, a local comparison of three units can be constructed. In the first case, the central unit emits light normally, and the brightness decreases outward from the periphery. In this case, the central divergence value may be positive and relatively high, such as 0.56. In the second case, the central unit has failed, but the halo of the normal units on both sides covers its area, resulting in a medium gray level at the center. However, since this medium gray level is not the source, the calculated central divergence value may be close to 0 or even negative, such as -0.03.

[0122] In the third case, if there is an abnormal leakage current in the central unit, causing a sharp peak in local brightness, the divergence value may be abnormally concentrated, such as 0.92. The system maps these divergence results into a two-dimensional image, that is, forms a light field divergence distribution map, in which the values ​​at different locations can be further encoded into different gray levels or pseudo-colors for subsequent intrinsic stripping processing.

[0123] As an anomaly handling mechanism, if there are isolated noise points in the potential energy map, direct differentiation may amplify high-frequency noise. The system can perform a light smoothing of the local area before differentiation, but the smoothing kernel should be limited to the range that does not cross multiple adjacent lamp positions to avoid weakening the true divergence peak.

[0124] If the edge pixels cannot obtain the complete left and right or top and bottom neighborhoods, mirror filling, boundary copying or single-sided difference can be used for calculation; if the center value and the surrounding value of a certain area are close to the upper limit due to saturation, the divergence result may be compressed. In this case, the system can first perform local dynamic recovery based on the exposure metadata at the time of acquisition, and then enter the divergence calculation.

[0125] For example, in the online detection of the aforementioned conference display terminal module, the system calculates the horizontal and vertical changes of the full-screen floating-point potential energy map pixel by pixel to generate the gradient vector field of the entire board;

[0126] For a light position in the middle that appears normal in the original image, its gray value is close to that of the surrounding area, but the scatter plot shows that this position is a low value band, indicating that it is not a local light source, but a false bright area formed by the diffusion effect of the surrounding area; conversely, another area in the original image that is only slightly brighter than the background forms a clear sharp peak in the scatter plot and is retained as a key candidate location for subsequent leakage current analysis.

[0127] The purpose of this step is to convert the brightness representation into a divergence distribution that can characterize the causal relationship of local luminescence, so that the system can identify the real source and the passively illuminated area from the structure of image intensity changes, thereby improving the separability of defects in strong crosstalk scenarios.

[0128] In a preferred embodiment of the present invention, the process of applying the Laplace operator to calculate the second spatial derivative of the gradient vector field to obtain the divergence matrix includes the following steps: performing a fast Fourier transform on the gradient vector field to convert the spatial domain data to the frequency domain; performing divergence matrix calculation in the frequency domain to obtain the frequency domain divergence matrix; and performing an inverse fast Fourier transform on the frequency domain divergence matrix to restore it to the spatial domain to obtain the divergence matrix.

[0129] This embodiment provides a frequency domain divergence matrix calculation mechanism; specifically, although the previous solution can obtain the divergence matrix by spatial domain differentiation, on large-size, high-resolution MiniLED direct display devices, if the second-order spatial derivative operation is performed directly pixel by pixel, the calculation delay will increase rapidly as the image size increases, making it difficult to adapt to the online detection cycle.

[0130] Especially in 8K-level or multi-module splicing scenarios, the cumulative computational complexity of spatial domain convolution and difference operations is high. To address this, this embodiment introduces Fast Fourier Transform to transfer divergence calculation to the frequency domain, thereby reducing overall computational complexity and improving the overall board processing efficiency.

[0131] Specifically, the system first performs a fast Fourier transform on the gradient vector field to convert the change information in the spatial domain to a frequency domain representation;

[0132] This can be understood as the gradient changes that were originally distributed in a two-dimensional plane being decomposed into components at different spatial frequencies. This transformation is particularly advantageous for the luminous patterns of MiniLEDs with regular dot matrix structures, because a large number of regular backgrounds and local anomalous peaks can be presented in a more separable form in the frequency domain.

[0133] For ease of explanation, a simplified one-dimensional illustration can be constructed: if a certain row of gradient change sequence is ([1,3,5,3,1]), where the center is high and the sides are low, it represents a local diffusion source; after conversion to the frequency domain, the low-frequency components mainly reflect the overall slowly changing background, while the high-frequency components mainly reflect local peaks and edge changes; the system directly performs the second-order spatial derivative related calculation in the frequency domain based on the frequency response relationship corresponding to the divergence operation, and obtains the frequency domain divergence matrix; specifically, when performing divergence matrix calculation in the frequency domain, the system pre-constructs a two-dimensional frequency coordinate system with the same dimension as the frequency domain matrix, and performs zero-frequency centering shift processing to ensure that the low-frequency components are clustered at the center of the matrix; the system is based on the formula:

[0134]

[0135] The frequency domain matrix after the potential energy diagram is transformed to the frequency domain is denoted as follows: The centered spatial frequency variables are denoted as follows: and The corresponding frequency domain divergence matrix is ​​denoted as ;

[0136] By extracting the frequency variables in the horizontal and vertical directions respectively and adding them together using complex multiplication, the system can perform equivalent processing to divergence differentiation and Laplace composite calculation in the frequency domain, ultimately obtaining a unified scalar frequency domain divergence matrix; then, by using inverse fast Fourier transform, it is restored to the spatial domain, thus obtaining a divergence matrix that corresponds one-to-one with the original image position.

[0137] In engineering implementation, the advantage of frequency domain computation lies not only in speed, but also in its ease of integration with subsequent point spread functions for unified processing. This is because the convolution, deconvolution, and second derivative operations of the ideal point spread function can all be expressed as multiplication and division relationships on frequency components in the frequency domain, thereby reducing the overhead caused by multiple round trip scans in the spatial domain.

[0138] As an exception handling mechanism, if the input image size is not an integer power of 2, the system can perform a fast Fourier transform after padding the boundaries with zeros, and then crop it back to the effective area after the inverse transform. If padding with zeros causes edge ringing, a window function can be added before the transform to alleviate the boundary abruptness. If some frequency components are close to zero, they may cause numerical instability in subsequent frequency domain division. In this case, a minimum spectral limit can be set to avoid excessive amplification of noise.

[0139] If the panel under test is small and the image resolution is limited, the improvement in frequency domain processing efficiency is limited, and the system can also revert to the spatial domain divergence calculation mode.

[0140] For example, in the mass production testing of the aforementioned conference display terminal module, the resolution of the entire board image reaches tens of millions of pixels; the system performs fast Fourier transform on the full-screen gradient vector field in blocks, completes the divergence matrix calculation in the frequency domain, and then synthesizes the spatial domain result of the entire board; compared with the block-by-block spatial difference and convolution scheme, the processing latency of the entire board is significantly reduced, enabling the system to meet the real-time requirements of continuous online testing, while the output divergence map still maintains the spatial positioning relationship consistent with the theoretical grid;

[0141] The purpose of this step is to improve the processing efficiency of large-size images in the frequency domain without changing the divergence discrimination principle, and to provide a shareable frequency domain calculation basis for subsequent inverse filtering and stripping, thereby achieving low-latency image processing in high-resolution online detection.

[0142] In a preferred embodiment of the present invention, a pre-calibrated ideal point spread function convolution kernel of a defect-free emitting unit is loaded and nonlinear inverse filtering calculation is performed with the light field divergence distribution map; based on the nonlinear inverse filtering calculation result, neighborhood optical flow crosstalk is removed and intrinsic divergence peak matrix is ​​extracted; the method of synchronously generating a stripped residual distribution map containing sub-pixel level distortion data includes: retrieving the pre-calibrated ideal point spread function convolution kernel of the defect-free emitting unit;

[0143] The light field divergence distribution map is deconvolved with the ideal point spread function kernel to obtain the nonlinear inverse filtering calculation result; based on the nonlinear inverse filtering calculation result, the overflow optical flow data between the emitting units is subtracted by matrix subtraction, and the local intrinsic divergence peak value of each emitting unit is extracted to construct the intrinsic divergence peak matrix; the overflow optical flow data subtracted in the matrix subtraction operation is recorded to generate the stripped residual distribution map containing subpixel-level distortion data.

[0144] This embodiment provides an intrinsic divergence stripping and residual extraction mechanism; specifically, although the divergence distribution map can better distinguish between the real light source and the passive light-receiving area than the original brightness map, in a high-density MiniLED array, the point diffusion tails of adjacent lamp positions may still overlap with each other.

[0145] If this neighborhood optical flow crosstalk is not further removed, some weak defect locations may still retain positive divergence contributions from their neighbors, thus affecting the final classification. Therefore, in this embodiment, a pre-calibrated ideal point spread function convolution kernel is loaded onto the divergence map to perform deconvolution and matrix subtraction on the local optical field, and the intrinsic divergence peak value of each emitting unit is extracted.

[0146] Specifically, the system retrieves the ideal point diffusion function convolution kernel of the pre-calibrated defect-free light-emitting unit; this convolution kernel can be obtained through standard template calibration, for example, by selecting a defect-free reference module, lighting only a single light-emitting unit, collecting its diffusion distribution on the imaging plane and normalizing it to obtain a local convolution kernel; for illustration, a simplified 3×3 convolution kernel can be set: ([0.05,0.10,0.05;0.10,0.40,0.10;0.05,0.10,0.05]); where the center has the highest weight, and the surrounding area represents the typical diffusion leakage ratio;

[0147] The system performs deconvolution calculations on the light field divergence distribution map and the ideal point spread function kernel to obtain the nonlinear inverse filtering calculation results. The nonlinear inverse filtering calculation here is not a simple direct division operation in the frequency domain, but a deconvolution algorithm that introduces non-negativity constraints and iterative compensation mechanisms, such as the Richardson-Lucy algorithm.

[0148] During the calculation process, the system uses the light field divergence distribution map as the initial observation state. In each iteration, the estimated intrinsic divergence distribution is convolved with the ideal point spread function to obtain the estimated observation distribution. Then, the next intrinsic divergence estimate is updated point by point by the difference between the ratio of the real observation divergence map and the estimated observation distribution.

[0149] Simultaneously, a non-negative truncation condition is applied in the iterative link to force unreasonable negative extreme values ​​appearing in the calculation to be returned to zero, so as to satisfy the non-negativity constraint of the physical properties of the image, avoid abnormal amplification of high-frequency noise and suppress the ripple effect caused by excessive stripping, thereby ensuring the stable convergence of the nonlinear inverse filtering calculation; the result can be understood as follows: if the observed scatter plot is formed by the diffusion superposition of several ideal light sources, then the most likely original source distribution after deconvolution is the calculation result.

[0150] For example, a local observation divergence patch is ([0.08, 0.18, 0.09; 0.20, 0.60, 0.21; 0.07, 0.17, 0.08]). After deconvolution, a more concentrated source distribution may be obtained, where the estimated value at the center location increases to 0.56, while the values ​​at the surrounding locations decrease to between 0.05 and 0.09. This indicates that most of the response at the periphery comes from diffusion from the central unit rather than independent emission from itself.

[0151] The system performs a matrix subtraction operation based on the inverse filtering result, deducts the overflow optical flow data between each emitting unit, and extracts the local intrinsic divergence peak value of each emitting unit;

[0152] Specifically, the matrix subtraction operation refers to: taking the optical flow value belonging to the neighboring luminescent unit diffused to the coordinate position of the current unit in the estimated observation distribution matrix calculated by the nonlinear inverse filtering as the overflow optical flow data, and subtracting it from the initial observation divergence value of the coordinate position of the current unit in the light field divergence distribution map, thereby obtaining the intrinsic divergence peak value that only characterizes the luminous intensity of the current unit itself.

[0153] Taking the aforementioned local block as an example, if the observed value of 0.20 in the left cell is estimated by the model to have 0.13 originating from the diffusion of the central cell, then its eigenvalue is only 0.07.

[0154] If the backflow from the neighbor in the central cell observation of 0.60 is only 0.04, then its eigenvalue can be recorded as 0.56; the system performs this processing grid by grid across the entire map, resulting in the final eigenvalue divergence peak matrix;

[0155] During the matrix subtraction process, the subtracted overflow optical flow data is recorded synchronously to form a stripping residual distribution map. This residual map is not a simple error map, but reflects the degree of deviation between certain regions and the ideal diffusion model. For example, if the lens of a lamp position is slightly off-center, although it has not reached the level of dead lamp or leakage, its actual diffusion shape is inconsistent with the standard convolution kernel. After subtraction, this region will leave directional residual stripes.

[0156] These residuals can be further used to generate sub-pixel level optical compensation parameters;

[0157] As a fallback, if the ideal point diffusion function obtained from the calibration has a systematic deviation from the current batch of devices, such as changes in the diffusion scale caused by lens replacement or detection distance adjustment, the system can first normalize the scale of the convolution kernel or recalibrate it;

[0158] If some frequency components are too small during the deconvolution process, causing oscillations in the results, the system can add regularization constraints to limit the upper limit of the inverse filter gain. If a certain region has a large difference between the observation mode and the ideal diffusion due to atypical defects such as local contamination or scratches, the system can mark the region as an abnormal optical morphology and retain the original residual block separately for manual verification while maintaining the intrinsic divergence output.

[0159] For example, in the online detection of the aforementioned conference display terminal module, the original scatter plot of a certain light position in the middle is moderately high, which seems normal, but the two light positions on its left and right also show an increase at the same time; after the system calls the ideal point diffusion function convolution kernel of the corresponding batch of the module to perform deconvolution, it finds that the middle light position has a stable intrinsic peak, while the left and right sides are mainly diffusion leakage. Therefore, the intrinsic values ​​of the left and right positions are lowered and are no longer mistakenly recorded as abnormal bright spots.

[0160] Meanwhile, another area in the upper right corner showed a persistently large residual after stripping, indicating that although there were no hard defects, there was sub-pixel level optical distortion, which provided a direct data source for subsequent image uniformity compensation.

[0161] The purpose of this step is to separate the neighborhood coupling component from the unit's self-luminous component in the divergence map, obtain the intrinsic divergence peak value that is closer to the real electroluminescence state, and at the same time retain the residual information generated by the stripping, so as to achieve the coordinated output of defect detection and optical compensation data extraction.

[0162] In a preferred embodiment of the present invention, the method of generating defect category labels by performing conditional classification based on the intrinsic divergence peak matrix includes: extracting the intrinsic divergence peak of each luminescent unit in the intrinsic divergence peak matrix;

[0163] The intrinsic divergence peak value is compared with a preset critical divergence threshold and a preset extreme divergence threshold, wherein the preset extreme divergence threshold is greater than the preset critical divergence threshold; configured as follows: if the intrinsic divergence peak value is greater than the preset critical divergence threshold and less than the preset extreme divergence threshold, then the light-emitting unit is determined to be normal and a normal category label is generated.

[0164] If the intrinsic divergence peak value is less than or equal to the preset critical divergence threshold, the light-emitting unit is determined to have a non-light-emitting defect, and a dead light dark spot category label is generated; the above threshold is obtained by statistical sampling of the intrinsic divergence distribution of defect-free benchmark samples; if the intrinsic divergence peak value is greater than or equal to the preset extreme divergence threshold, the light-emitting unit is determined to have a leakage defect, and a puncture leakage category label is generated.

[0165] This embodiment provides a conditional classification mechanism based on intrinsic divergence peaks. Specifically, after obtaining the intrinsic divergence peak matrix, although the system can accurately remove neighborhood crosstalk, it is still difficult to convert the numerical results into executable production line judgments if there is a lack of clear classification decision boundaries.

[0166] Especially in actual production, the detection system not only needs to point out abnormalities, but also needs to clarify whether it is an acceptable normal fluctuation, a dead light spot that needs to be repaired, or a puncture leakage that needs to be isolated immediately; therefore, this embodiment introduces a dual threshold condition classification method to map continuous intrinsic divergence peaks to discrete defect category labels.

[0167] Specifically, the system extracts the eigenvalues ​​of each luminescent unit from the eigenvalue peak matrix and compares them with the preset critical divergence threshold and extreme divergence threshold.

[0168] The critical divergence threshold is used to distinguish between those with normal light source characteristics and those lacking self-light source characteristics; the extreme divergence threshold is used to identify abnormally concentrated high divergence responses; the relationship between the two is fixed so that the extreme divergence threshold is greater than the critical divergence threshold.

[0169] The determination mechanism of the puncture leakage category label is as follows: When the positive and negative semiconductor junctions break down or micro short circuits occur inside the light-emitting unit, it will cause the local current density to increase abnormally. In the potential energy diagram and the intrinsic divergence peak matrix, it will be manifested as an abnormal energy convergence and an extreme peak that far exceeds the maximum possible light emission intensity under normal driving. Therefore, the abnormal state where the intrinsic divergence peak is not lower than the extreme divergence threshold is clearly defined and mapped as a puncture leakage defect.

[0170] For ease of explanation, the critical divergence threshold can be set to 0.10, and the extreme divergence threshold to 0.85. The intrinsic divergence peak values ​​of five emitting units in a certain local row are 0.61, 0.08, 0.57, 0.91, and 0.13, respectively. The system classifies them as follows: 0.61 is greater than 0.10 and less than 0.85, and is marked as normal; 0.08 is not greater than 0.10, and is marked as a dead light dark spot.

[0171] 0.57 is also considered normal; 0.91 is not less than 0.85 and is marked as puncture leakage; 0.13 is close to the critical threshold but still higher than 0.10, so it is temporarily classified as normal; in this way, the system can integrate continuous numerical results into a category label sequence.

[0172] Furthermore, in engineering implementation, the threshold can be derived from historical sample data statistics, reference module calibration results, or preset according to product specifications; for different brightness levels, different package sizes, or different detection exposure configurations, the threshold can be configured as different templates.

[0173] The specific configuration logic is as follows: The system preloads a batch of illuminated image data of defect-free reference samples, extracts the intrinsic divergence peak values ​​of all normal luminous units, and calculates their statistical distribution mean and standard deviation.

[0174] The mean minus three standard deviations is set as the critical divergence threshold, and the mean plus three standard deviations is set as the extreme divergence threshold. This achieves adaptive classification boundary setting based on data distribution, avoiding the blindness of manual experience setting. If the overall intrinsic divergence level of the same batch of products shifts slightly under changes in ambient temperature, the system can also add a small-range dynamic bias on the basis of the fixed template to maintain classification stability.

[0175] As a fallback, if the intrinsic divergence peak value of a certain light-emitting unit is exactly equal to the critical divergence threshold, it is treated as a non-light-emitting defect to avoid the boundary lamp position being allowed; if it is exactly equal to the extreme divergence threshold, it is treated as a puncture leakage to improve the interception rate of high-risk defects.

[0176] If the intrinsic values ​​of some units are negative, they can be directly classified into the dead light dark spot category, indicating that the location not only lacks source features after stripping, but also shows a convergence or collapse trend; if the confidence of the intrinsic peak value is insufficient due to the large residual in a certain area, the system can add a low confidence label in addition to the main label for manual review or secondary image taking.

[0177] For example, in the aforementioned mass production sampling inspection of the conference display terminal module, after the system completed the traversal of the full-screen intrinsic divergence peak matrix, it found that the intrinsic value of the 32nd row and 48th column was 0.04, which was judged as a dead light dark spot; the intrinsic value of the 18th row and 73rd column was 0.89, which was judged as a puncture leakage.

[0178] The intrinsic values ​​of most lamp positions are concentrated in the range of 0.40 to 0.70, and are classified as normal. Thus, the detection of this system no longer relies on subjective observation of brightness, but directly generates objective defect label results based on a unified divergence threshold rule.

[0179] The purpose of this step is to stably transform the intrinsic divergence values ​​after crosstalk stripping into structured defect categories, thereby enabling automatic classification output suitable for online rejection, rework, and quality statistics.

[0180] In a preferred embodiment of the present invention, the defect category label is mapped to the light-emitting array topology mesh model to generate a structured defect error correction mask map;

[0181] The method for generating an optical compensation reference matrix based on the stripping residual distribution map and outputting the structured defect correction mask map and the optical compensation reference matrix includes: extracting the two-dimensional coordinates of the light-emitting units corresponding to the defect category labels; spatially binding the two-dimensional coordinates of the light-emitting units with the topological mesh model of the light-emitting array to generate the structured defect correction mask map;

[0182] The subpixel-level distortion data is the residual value reflecting the offset of the light emission center or the asymmetric diffusion characteristics in the stripped residual distribution map; the subpixel-level distortion data is extracted from the stripped residual distribution map; the reverse compensation parameters are calculated based on the subpixel-level distortion data to generate the optical compensation reference matrix; the structured defect correction mask map and the optical compensation reference matrix are output synchronously and sent to the display control terminal for image compensation.

[0183] This embodiment provides a defect mapping and compensation closed-loop output mechanism; specifically, after the defect classification is completed, if only a set of scattered abnormal coordinates and category labels are output, although it can support quality inspection rejection, it cannot directly serve the subsequent online correction or factory rework process of the display control terminal.

[0184] Especially for areas that are not completely ineffective but have slight uneven light emission, a single defect label cannot reflect how they should be compensated. Therefore, based on the classification results, this embodiment further constructs a structured defect correction mask map and uses the stripping residual distribution map to generate an optical compensation reference matrix, thereby realizing an integrated closed loop of detection and compensation.

[0185] Specifically, the system first extracts the two-dimensional coordinates of the light-emitting unit corresponding to each defect category label, and then spatially binds them with the aforementioned light-emitting array topology mesh model to form a structured defect correction mask. This mask not only records the abnormal location, but also the abnormal type. For example, a multi-value encoding method can be used: 0 represents normal, 1 represents dead light dark spot, 2 represents puncture leakage, and 3 represents pending re-inspection.

[0186] For ease of explanation, assume that the classification result of a certain 2×4 local grid is: first row: [normal, dead light dark spot, normal, normal]; second row: [normal, normal, puncture leakage, normal]; then its mask coding matrix can be represented as ([0,1,0,0;0,0,2,0]); this matrix can be directly used to highlight the target light position in the repair terminal, and can also be used by the display controller to disable or reduce the weight of the corresponding pixel area during operation.

[0187] On the other hand, the system extracts subpixel-level distortion data from the stripped residual distribution map; here, subpixel-level means that the residual does not strictly fall on the center of a single lamp position, but may be manifested as leftward, rightward, oblique tailing or local asymmetric diffusion; the system calculates the reverse compensation parameters based on these residual shapes.

[0188] The rule for calculating the reverse compensation parameter is as follows: using the ideal target brightness or target divergence preset by the host computer or display control terminal as the reference value, the sub-pixel level distortion data is compared with the reference value to generate a proportional adjustment coefficient or bias voltage adjustment amount to offset the residual, which is used as the reverse compensation parameter; wherein, for positive residuals, an attenuation proportional coefficient less than 1 is generated, and for negative residuals, a gain proportional coefficient greater than 1 is generated.

[0189] For example, if the residual on the right side of a certain lamp position is consistently high, it indicates that its actual light diffusion is slightly offset to the right. In this case, the compensation matrix can be used to assign a reverse attenuation coefficient to that lamp position and its adjacent right side region. If the overall residual of a certain region is low, it indicates that there is a slight light deficiency at that location. In this case, an appropriate gain parameter can be assigned. Finally, the compensation parameters of each region are combined to form an optical compensation reference matrix.

[0190] For ease of explanation, the residual statistics corresponding to four lamp positions in a certain area can be set as ([+0.03,-0.01,+0.05,0.00]). The system can convert this into reverse compensation coefficients ([0.97,1.01,0.95,1.00]), where positive residuals correspond to lowering the gain and negative residuals correspond to raising the gain. If combined with directional residuals, the single gain coefficient can be extended into a local two-dimensional compensation block. In this way, the output is not a simple suggestion to make the lamp brighter or darker, but a structured compensation reference matrix that can be directly loaded by the display control terminal.

[0191] As a fallback, if a lamp position is determined to be a dead lamp or dark spot, the system can mark it as an uncompensable hard defect in the compensation matrix to avoid the control end blindly increasing the drive and causing greater risks; if a lamp position is determined to be a puncture leakage, a high-priority disable mark can be added to the mask image and the corresponding area can be prohibited from participating in the regular Demura gain calculation.

[0192] If the residual distribution map has excessive noise in a certain area, causing the compensation parameter to fluctuate beyond the limit, the system can use neighborhood smoothing compensation for that area or directly retain it for manual review; if the display control terminal only supports low-dimensional compensation tables, the system can also perform block summarization of the high-resolution compensation matrix and output a version that matches the capabilities of the control terminal.

[0193] For example, after the terminal module in the aforementioned conference completes the whole board inspection, the system maps the coordinates of the dead light dark spot in the 32nd row and 48th column and the coordinates of the puncture leakage current in the 18th row and 73rd column to a 64×96 topological mesh to generate the corresponding structured defect correction mask map.

[0194] Meanwhile, statistical and directional analysis of the full-screen stripping residuals is performed to obtain an optical compensation reference matrix. This mask image is sent to the rework station to guide maintenance personnel to prioritize the handling of hard defects. The compensation reference matrix is ​​then sent to the display control terminal to pre-compensate for local brightness unevenness during the trial display stage, thereby reducing the need for secondary image acquisition and repeated correction.

[0195] The purpose of this step is to transform the results of the aforementioned detection chain into executable, rewritable, and compensable structured data output, thereby achieving closed-loop linkage between defect identification, rework location, and image uniformity compensation.

[0196] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. An intelligent detection system for MiniLED direct-display lighting defects based on image recognition, characterized in that, include: Image acquisition unit: used to acquire surface-emitting images of the MiniLED direct-view display device as initial image data; Light field potential energy initialization unit: receives the initial image data, establishes a topological mesh model of the luminescent array; performs geometric distortion correction on the initial image data based on the topological mesh model of the luminescent array, and extracts the pixel brightness matrix; converts the pixel brightness matrix into a floating-point light field potential energy map; Optical field divergence decoupling unit: Analyzes the floating-point optical field potential energy map and calculates the gradient vector field; calculates the divergence matrix based on the gradient vector field and generates an optical field divergence distribution map; Intrinsic divergence stripping unit: Load the ideal point spread function convolution kernel of the pre-calibrated defect-free luminescent unit, and perform nonlinear inverse filtering calculation with the light field divergence distribution map; based on the nonlinear inverse filtering calculation result, remove the neighborhood optical flow crosstalk and extract the intrinsic divergence peak matrix; Simultaneously generate a stripping residual distribution map containing sub-pixel level distortion data; Defect classification and compensation closed-loop unit: Performs conditional classification based on the intrinsic divergence peak matrix and generates defect category labels; The defect category labels are mapped to the light-emitting array topology mesh model to generate a structured defect error correction mask map; An optical compensation reference matrix is ​​generated based on the stripping residual distribution map, and the structured defect correction mask map and the optical compensation reference matrix are output to the display control terminal so that the terminal can shield the defective unit based on the structured defect correction mask map and perform brightness compensation on the normal light-emitting unit based on the optical compensation reference matrix.

2. The image recognition-based intelligent detection system for MiniLED direct-display lighting defects according to claim 1, characterized in that, The method of receiving the initial image data and establishing the topological mesh model of the luminescent array includes: Obtain the initial image data; Extract the theoretical arrangement parameters of the MiniLED direct-view device; A two-dimensional coordinate system is constructed based on the theoretical arrangement parameters to generate the topological mesh model of the light-emitting array.

3. The image recognition-based intelligent detection system for MiniLED direct-display lighting defects according to claim 2, characterized in that, The method of performing geometric distortion correction on the initial image data based on the light-emitting array topological mesh model and extracting the pixel brightness matrix; and converting the pixel brightness matrix into a floating-point light field potential energy map includes: Extract edge localization features from the initial image data; The edge localization features are spatially aligned with the topological mesh model of the light-emitting array, and the affine transformation parameters are calculated. Based on the affine transformation parameters, the initial image data is spatially resampled to generate corrected image data; Extract the grayscale value of each pixel in the corrected image data to construct the pixel brightness matrix; The discrete grayscale values ​​in the pixel brightness matrix are mapped to continuous potential energy height values ​​to generate the floating-point optical field potential energy map.

4. The image recognition-based intelligent detection system for MiniLED direct-display lighting defects according to claim 3, characterized in that, The floating-point optical potential energy diagram is analyzed, and the gradient vector field is calculated. The methods for generating a light field divergence distribution map based on the gradient vector field to calculate the divergence matrix include: Spatial differentiation is performed on the floating-point optical potential energy map to calculate the partial derivatives of each pixel in the horizontal and vertical directions; The gradient vector field is generated by combining the partial derivatives in the horizontal and vertical directions. The second-order spatial derivative of the gradient vector field is calculated by applying the Laplacian operator to obtain the divergence matrix; wherein the Laplacian operator adopts a four-neighborhood difference template to quantify the energy source and sink intensity of the current pixel relative to its neighborhood. The divergence matrix is ​​mapped to a two-dimensional image format to generate the light field divergence distribution map.

5. The image recognition-based intelligent detection system for MiniLED direct-display lighting defects according to claim 4, characterized in that, The process of applying the Laplacian operator to the gradient vector field to calculate the second spatial derivative and obtain the divergence matrix includes the following steps: Perform a Fast Fourier Transform on the gradient vector field to convert the spatial domain data to the frequency domain; Perform divergence matrix calculation in the frequency domain to obtain the frequency domain divergence matrix; Perform an inverse fast Fourier transform on the frequency domain divergence matrix to restore it to the spatial domain and obtain the divergence matrix.

6. The image recognition-based intelligent detection system for MiniLED direct-display lighting defects according to claim 5, characterized in that, The ideal point spread function convolution kernel of the pre-calibrated defect-free light-emitting unit is loaded and nonlinearly inversely filtered with the light field divergence distribution map; based on the nonlinear inverse filtering calculation result, neighborhood optical flow crosstalk is removed and the intrinsic divergence peak matrix is ​​extracted. Methods for synchronously generating stripped residual distribution maps containing sub-pixel level distortion data include: Retrieve the ideal point spread function convolution kernel of the pre-calibrated defect-free light-emitting unit; The light field divergence distribution map is deconvolved with the ideal point spread function kernel to obtain the nonlinear inverse filtering calculation result. Based on the nonlinear inverse filtering calculation results, the overflow optical flow data between the light-emitting units is deducted by matrix subtraction operation, the local intrinsic divergence peak value of each light-emitting unit is extracted, and the intrinsic divergence peak value matrix is ​​constructed. Record the overflow optical flow data that is subtracted during the matrix subtraction operation to generate the stripping residual distribution map containing subpixel-level distortion data.

7. The image recognition-based intelligent detection system for MiniLED direct-display lighting defects according to claim 6, characterized in that, The method for performing conditional classification based on the intrinsic divergence peak matrix and generating defect category labels includes: Extract the eigendivergence peak value of each luminescent unit in the eigendivergence peak matrix; The intrinsic divergence peak value is compared with a preset critical divergence threshold and a preset extreme divergence threshold, wherein the preset extreme divergence threshold is greater than the preset critical divergence threshold. The configuration is as follows: if the intrinsic divergence peak value is greater than the preset critical divergence threshold and less than the preset extreme divergence threshold, the light-emitting unit is determined to be normal, and a normal category label is generated; if the intrinsic divergence peak value is less than or equal to the preset critical divergence threshold, the light-emitting unit is determined to have a non-luminous defect, and a dead light dark spot category label is generated; the above thresholds are obtained based on statistical sampling of the intrinsic divergence distribution of defect-free benchmark samples; if the intrinsic divergence peak value is greater than or equal to the preset extreme divergence threshold, the light-emitting unit is determined to have a leakage defect, and a puncture leakage category label is generated. The normal category label, the dead light / dark spot category label, and the puncture / leakage category label are integrated to generate the defect category label.

8. The image recognition-based intelligent detection system for MiniLED direct-display lighting defects according to claim 7, characterized in that, The defect category labels are mapped to the light-emitting array topology mesh model to generate a structured defect error correction mask map; The methods for generating an optical compensation reference matrix based on the stripping residual distribution map and outputting the structured defect correction mask map and the optical compensation reference matrix include: Extract the two-dimensional coordinates of the light-emitting unit corresponding to the defect category label; Spatially bind the two-dimensional coordinates of the light-emitting unit to the topological mesh model of the light-emitting array to generate the structured defect correction mask map; The subpixel-level distortion data is the residual value in the stripped residual distribution map that reflects the offset of the luminescence center or the asymmetric diffusion characteristics; the subpixel-level distortion data in the stripped residual distribution map is extracted. The inverse compensation parameters are calculated based on the subpixel-level distortion data to generate the optical compensation reference matrix; The structured defect correction mask and the optical compensation reference matrix are output synchronously and sent to the display control terminal for image compensation.

Images