OLED defect detection equipment based on machine vision

Abstract

The invention discloses OLED defect detection equipment based on machine vision, and relates to the technical field of visual detection, and the detection equipment comprises the following working processes: driving an OLED panel through a sinusoidal modulation current sequence, synchronously collecting a plurality of light-emitting images, and carrying out registration; constructing a three-dimensional texture tensor and generating a scale weighted texture field through wavelet decomposition; the local entropy density and the multi-scale gradient are fused, and a defect sensitive entropy change field is calculated in combination with the direction constraint matrix; non-linearly enhancing the defect area, and executing anisotropic diffusion filtering to optimize the signal-to-noise ratio; extracting a candidate region, and iteratively solving an energy functional based on an improved Snake model to realize sub-pixel boundary fitting; calculating morphological and optical characteristic parameters such as a defect area, a contrast ratio and a spectrum abnormal index; and performing decision tree classification on the feature vectors, identifying defect types, and generating a report. According to the method, multi-scale texture response of the OLED panel is excited through active current modulation, and electromagnetic feature expression of microdefects is enhanced in combination with an entropy change field algorithm.

Description

Technical Field

[0001] This invention belongs to the field of visual inspection technology, and in particular relates to OLED defect detection equipment based on machine vision. Background Technology

[0002] OLED (Organic Light-Emitting Diode) is a self-emissive display technology made of organic materials. It emits light through carrier injection and recombination in an organic layer driven by current, eliminating the need for a backlight module. Compared to LCD, it boasts advantages such as ultra-thinness, wide viewing angles (no color distortion), microsecond-level response (no ghosting), low power consumption, and flexibility. It also delivers vibrant colors with high contrast and a wide color gamut. It is widely used in small-sized mobile phones and smart wearables, as well as large-screen TVs, automotive displays, and VR headsets. Technically, WOLED is highly mature, offering superior brightness and lifespan, while QD-OLED utilizes quantum dots to enhance color purity. Currently, the bottlenecks of "short lifespan and difficulty in achieving large sizes" have been overcome, and larger, higher-performance products are continuously being developed, making it one of the mainstream directions in display technology.

[0003] Electroluminescence imaging has certain limitations in existing OLED defect detection technologies. When the OLED panel is in a constant driving current state, the difference between the light emission response of microcracks and low-contrast Mura defects and the optical characteristics of normal areas is extremely weak, which makes the defect signal easily drowned out by substrate texture noise.

[0004] Traditional machine vision methods mainly rely on grayscale or gradient analysis of single-frame images, which makes it difficult to effectively separate the electromagnetic response features of such defects. In addition, since the physical scale of micro-defects is close to the pixel resolution limit, their electroluminescence behavior is difficult to form distinguishable texture patterns under static conditions, resulting in insufficient detection system capability to capture sub-pixel-level abnormal responses, which has become a bottleneck restricting detection sensitivity.

[0005] In view of the above problems, the following solutions are proposed. Summary of the Invention

[0006] The purpose of this invention is to provide an OLED defect detection device based on machine vision. By using active current modulation to excite multi-scale electroluminescence response and texture entropy change field algorithm, it can separate the electromagnetic anomaly signal of defects in dynamic feature space, thus solving the problem that the microscale defect response is submerged by substrate texture noise in the prior art.

[0007] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution:

[0008] This invention relates to an OLED defect detection device based on machine vision. The detection device includes a current modulation control module, an image acquisition module, an image registration module, a multi-scale texture analysis module, an entropy change field calculation module, a defect enhancement module, a sub-pixel localization module, a defect quantization module, a classification decision module, and a result output module.

[0009] The current modulation control module, image acquisition module, image registration module, multi-scale texture analysis module, entropy change field calculation module, defect enhancement module, sub-pixel positioning module, defect quantization module, classification decision module, and result output module are connected in sequence. The output end of the multi-scale texture analysis module is unidirectionally connected to the defect quantization module, and the output end of the entropy change field calculation module is unidirectionally connected to the sub-pixel positioning module.

[0010] The workflow of the testing equipment is as follows:

[0011] Step S1, Current Modulation Image Acquisition: Drive the OLED panel with a sinusoidal modulated current sequence, simultaneously acquire multiple luminescent images and perform subpixel-level registration;

[0012] Step S2, Multi-scale texture field construction: Extract image gradient features, construct a three-dimensional texture tensor, and generate a scale-weighted texture field through wavelet decomposition;

[0013] Step S3, Entropy Variation Field Calculation: Combine local entropy density and multi-scale gradient, and calculate the defect-sensitive entropy variation field by combining the directional constraint matrix;

[0014] Step S4, Defect Feature Enhancement: The defect region is nonlinearly enhanced using a Sigmoid-power-law composite function, and anisotropic diffusion filtering is performed to optimize the signal-to-noise ratio;

[0015] Step S5, Subpixel Boundary Localization: Extract candidate regions on the enhanced image, and achieve subpixel boundary fitting by iteratively solving the energy functional based on the improved Snake model;

[0016] Step S6, Defect Quantitative Analysis: Calculate morphological and optical characteristic parameters such as defect area, contrast, and spectral anomaly index;

[0017] Step S7, Classification Decision and Output: Classify the feature vectors using a decision tree based on a predefined set of rules, identify the defect type, and generate a report.

[0018] Furthermore, the current modulation control module is used to generate a sinusoidal modulated current sequence to drive the OLED panel and excite defect response characteristics under different brightness states;

[0019] The image acquisition module is used to synchronously acquire OLED light-emitting images under each current state to obtain the original defect data sequence;

[0020] The image registration module is used to perform sub-pixel level displacement correction on multiple frames of images to eliminate pixel-level offset errors caused by mechanical vibration.

[0021] The multi-scale texture analysis module is used to extract gradient features and construct a three-dimensional texture tensor, and obtain multi-scale texture features through wavelet decomposition.

[0022] The entropy change field calculation module is used to calculate the local entropy density change field with directional constraints, thereby enhancing the defect region;

[0023] The defect enhancement module is used to apply nonlinear transformation and anisotropic diffusion filtering to improve the contrast between defects and the background.

[0024] The subpixel positioning module accurately extracts defect boundaries based on the improved Snake model, achieving superpixel precision contour positioning.

[0025] The defect quantization module is used to calculate the morphological parameters and spectral anomaly index of defects and generate quantized feature vectors.

[0026] The classification decision module is used to automatically identify defect types based on feature vectors and pre-trained decision tree models;

[0027] The result output module is used to generate a spatial coordinate mapping map and a structured detection report, outputting visualized and machine-readable results.

[0028] Further, step S1, current modulation image acquisition, specifically includes the following steps:

[0029] Step S11: Control the OLED driving circuit to generate a stepped current sequence, specifically:

[0030] Based on a preset base drive current value, K different current values ​​are generated using a sinusoidal modulation mode:

[0031]

[0032] In the formula, I k I0 is the driving current value during the k-th modulation, α is the current modulation amplitude coefficient, k is the modulation sequence index, and K is the total number of modulations.

[0033] Step S12: In each current I k Below, a high-resolution CMOS sensor is used to acquire the panel's emission image F. k (x,y);

[0034] Step S13: Perform temporal registration on the image sequence, and output a precisely aligned registered image sequence for subsequent processing.

[0035]

[0036] In the formula, For the registered k-th frame image, w is the subpixel level displacement correction operator, F k (x,y) represents the k-th frame of the original acquisition, F ref (x,y) represents the reference image, and (x,y) represents the pixel coordinates of the image.

[0037] This step drives the OLED panel by actively generating a sinusoidally modulated stepped current sequence, acquires panel emission images under different current intensities, and uses subpixel-level displacement correction to achieve high-precision registration of the image sequence, providing stable and comparable multi-brightness level image data for subsequent analysis.

[0038] Furthermore, step S2, the construction of the multi-scale texture field, specifically includes the following steps:

[0039] Step S21: Extract gradient-invariant features for each registered image:

[0040]

[0041] In the formula, G k (x,y) represents the gradient magnitude of the k-th frame image. For gradient operators, K is the convolution operator. sobel This is the Sobel edge detection convolution kernel, where ||·||2 is the L2 norm of the vector;

[0042] Step S22: Construct the 3D texture tensor:

[0043]

[0044] In the formula, T(x,y,k) is the value of the 3D texture tensor at position (x,y,k), and G... k (x,y) represents the gradient magnitude;

[0045] Step S23: Perform wavelet decomposition along the current dimension:

[0046]

[0047] In the formula, T sc (x,y) represents the scalar values ​​of the multi-scale texture field, s is the wavelet decomposition scale index, S is the maximum decomposition scale number, and ψ s These are scale-weighted coefficients. For a one-dimensional Haar wavelet transform operator, |·| s To obtain the absolute value of the wavelet coefficients at scale s, T(x,y,:) is a one-dimensional signal along the current dimension at a fixed position (x,y);

[0048] This step extracts gradient-invariant features from the registered image sequence, constructs a spatiotemporal three-dimensional texture tensor, and separates texture feature components of different scales through wavelet decomposition to form a multi-scale texture field that can characterize the response characteristics of defects, thereby enhancing the recognizability of micro-defects.

[0049] Furthermore, step S3, the entropy change field calculation, specifically includes the following steps:

[0050] Step S31: Define the local entropy density function:

[0051]

[0052] In the formula, ε(x,y) is the local entropy density value, i is the gray level, and p i Let be the probability of gray level i within a 5×5 window, μ be the mean gray level of the pixels within the window, σ be the standard deviation of the gray level of the pixels within the window, and exp(·) be the fourth-order Lorentz kernel function.

[0053] Step S32: Calculate the multiscale entropy-variable field:

[0054]

[0055] in,

[0056] In the formula, ΔE(x,y) is the output value of the entropy-variable field. For the scale-adaptive gradient operator, R s (x,y) is the texture direction constraint matrix, |·| is the modulus of the complex number, and θ s The textural orientation angle is arg(·), where arg(·) is the argument of the complex gradient vector.

[0057] This step proposes a local entropy density function combined with a scale-adaptive gradient operator, and introduces a texture direction constraint matrix to calculate a multi-scale entropy change field. By quantifying the spatial change rate of texture disorder, it reveals the boundary anomaly features of defects such as microcracks at the sub-pixel level.

[0058] Furthermore, step S4, defect feature enhancement, specifically includes the following steps:

[0059] Step S41: Construct a nonlinear enhancement function:

[0060]

[0061] In the formula, M(x,y) is the enhanced defect feature map, β is the Sigmoid steepness coefficient, τ is the activation threshold, and ‖·‖ is the absolute value of the scalar;

[0062] Step S42: Perform anisotropic diffusion filtering:

[0063]

[0064] In the formula, Let be the rate of change of the feature map over time, div be the divergence operator, c(z) be the diffusion coefficient function, z be the independent variable of the diffusion function, and κ be the diffusion threshold parameter.

[0065] This step employs a nonlinear Sigmoid function to selectively enhance the anisotropic entropy-varying field, combined with partial differential equation filtering using an exponential diffusion coefficient, to sharpen defect edges while suppressing background noise and improving the signal-to-noise ratio of the defect region.

[0066] Furthermore, step S5, sub-pixel boundary localization, specifically includes the following steps:

[0067] Step S51: Extract candidate defect regions on the enhanced image M(x,y);

[0068] Step S52: Calculate the boundary energy functional for each connected component Ω:

[0069]

[0070] In the formula, B is the boundary energy functional value. For the arc length parameter of the boundary curve, Here, α and β represent the defect boundary curve, and α and β are the energy weighting coefficients. For feature maps at boundary points The value, The tangent vector of the boundary curve;

[0071] Step S53: Solve iteratively using the improved Snake model:

[0072]

[0073] In the formula, v is the boundary point coordinate vector, t is the iteration time step, and γ is the elasticity coefficient. Let λ be the arc length parameter of the boundary curve, and λ be the gradient weight of the feature map. Let ε be the gradient of the feature map at point v, ε be the weight of the normal vector field, and n(v) be the unit normal vector derived from the entropy-change field.

[0074] This step is based on an improved Snake model with energy functional constraints. It iteratively optimizes the defect contour line on the enhanced image and achieves defect boundary localization with an accuracy of 0.1 pixels by driving the fusion of entropy-varying field normal vector field.

[0075] Furthermore, step S6, the defect quantification analysis, specifically includes the following steps:

[0076] Step S61: Calculate defect morphological parameters:

[0077]

[0078] In the formula, A is the area of ​​the defect region, Ω is the set of defective pixels, δx and δy are the physical dimensions of the pixels, C is the optical contrast of the defect region, and max Ω M represents the maximum pixel value of the enhanced feature map M(x,y) within the defect region Ω, min Ω M is the minimum pixel value of the enhanced feature map M(x,y) within the defect region Ω, and max Panel M is the global maximum pixel value of the enhanced feature map M(x,y) across the entire panel image range;

[0079] Step S62: Extract the spectral anomaly index:

[0080]

[0081] In the formula, SAI is the spectral anomaly index, |Ω| is the total number of pixels in the defect region, and T sc (x,y) represents the multi-scale texture field values. The average value of the texture field in the background region, ||·|| F It is the Frobenius norm;

[0082] This step calculates morphological area parameters, dynamic contrast index, and spectral anomaly index within the located defect area, extracting quantitative physical feature parameters to provide measurable criteria for defect classification.

[0083] Furthermore, step S7, classification decision and output, specifically includes the following steps:

[0084] Step S71: Input feature vector: f = [A, C, SAI, roundness, aspect ratio];

[0085] Step S72: Classify using a pre-trained decision tree model:

[0086]

[0087] In the formula, Class is the defect type label, A is the defect area, C is the contrast, and SAI is the spectral anomaly index;

[0088] Step S73: Generate defect coordinate mapping map D(x,y) and output a JSON format inspection report;

[0089] This step uses a decision tree model to make logical judgments on the quantification parameters based on preset multidimensional feature threshold rules, thereby achieving automated classification and identification of three typical defects: Mura, microcracks, and dead pixels. Finally, it generates vectorized contour data containing defect type, geometric features, and location coordinates, and outputs a structured inspection report and panel quality score.

[0090] The present invention has the following beneficial effects:

[0091] 1. This invention excites the multi-scale texture response of OLED panels through active current modulation and, combined with the entropy-changing field algorithm, enhances the electromagnetic feature expression of micro-defects. It overcomes the limitation of weak signals being submerged by background noise, enabling micron-level cracks and low-contrast Mura defects to generate quantifiable anomalous responses during gradient-invariant features and texture tensor decomposition. The scale-adaptive gradient operator in the entropy-changing field can effectively separate defect features from the substrate texture, and the nonlinear enhancement function further amplifies the signal-to-noise ratio of the defect region, laying the foundation for sub-pixel localization.

[0092] 2. This invention solves the problem of ambiguous positioning of edge detection in OLED defects by fusing entropy-varying field gradient information with geometric constraints; the boundary energy functional integrates entropy-varying field intensity and curve curvature constraints, and the improved Snake model uses the normal vector field derived from the entropy-varying field as the driving force, so that the boundary convergence process can adapt to irregular defect morphology; this design can accurately characterize the sub-pixel details of the defect contour, providing geometric accuracy guarantee for subsequent quantitative analysis.

[0093] 3. This invention establishes a mapping relationship between the essential characteristics of defects and classification decisions based on a multi-dimensional quantitative system of physical features combined with morphological parameters and spectral anomaly indices; multi-scale texture features generated by current modulation provide photoelectric response fingerprints of defects, and spectral anomaly indices effectively distinguish between material degradation and mechanical damage; the decision tree model constructs stable classification boundaries by integrating complementary features such as area, contrast, and shape factor, reducing the risk of misjudgment of transitional defects by traditional thresholding methods.

[0094] 4. This invention can synergistically suppress environmental interference through anisotropic diffusion filtering and texture direction constraint matrix; the current modulation sequence eliminates fixed pattern noise through time-domain differentiation, and the direction constraint matrix adaptively adjusts the detection sensitivity according to the local texture direction; wavelet scale weighting enhances the cross-scale consistency characteristics of defects, and the exponential term design of the nonlinear enhancement function effectively suppresses random noise; this design enables the system to maintain stable performance in complex industrial environments.

[0095] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0096] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0097] Figure 1 This is a schematic diagram of the OLED defect detection device based on machine vision according to the present invention. Detailed Implementation

[0098] 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.

[0099] Please see Figure 1 As shown, the present invention is an OLED defect detection device based on machine vision, including a current modulation control module, an image acquisition module, an image registration module, a multi-scale texture analysis module, an entropy change field calculation module, a defect enhancement module, a sub-pixel localization module, a defect quantization module, a classification decision module, and a result output module.

[0100] The current modulation control module, image acquisition module, image registration module, multi-scale texture analysis module, entropy field calculation module, defect enhancement module, sub-pixel positioning module, defect quantization module, classification decision module, and result output module are connected in sequence. The output end of the multi-scale texture analysis module is unidirectionally connected to the defect quantization module, and the output end of the entropy field calculation module is unidirectionally connected to the sub-pixel positioning module.

[0101] The current modulation control module is used to generate a sinusoidal modulated current sequence to drive the OLED panel and excite the defect response characteristics under different brightness states.

[0102] The image acquisition module is used to synchronously acquire OLED light-emitting images under each current state to obtain the original defect data sequence;

[0103] The image registration module is used to perform sub-pixel level displacement correction on multiple frames of images to eliminate pixel-level offset errors caused by mechanical vibration.

[0104] The multi-scale texture analysis module is used to extract gradient features and construct a three-dimensional texture tensor, and obtain multi-scale texture features through wavelet decomposition.

[0105] The entropy change field calculation module is used to calculate the local entropy density change field with directional constraints, thereby enhancing the defect region;

[0106] The defect enhancement module is used to apply nonlinear transformation and anisotropic diffusion filtering to improve the contrast between defects and the background;

[0107] The subpixel localization module accurately extracts defect boundaries based on the improved Snake model, achieving superpixel-precision contour localization;

[0108] The defect quantization module is used to calculate the morphological parameters and spectral anomaly index of defects, and generate quantized feature vectors;

[0109] The classification decision module is used to automatically identify defect types based on feature vectors and a pre-trained decision tree model;

[0110] The results output module is used to generate spatial coordinate mapping maps and structured inspection reports, outputting visualized and machine-readable results.

[0111] The workflow of the testing equipment is as follows:

[0112] Step S1, Current Modulation Image Acquisition: Drive the OLED panel with a sinusoidal modulated current sequence, simultaneously acquire multiple luminescent images and perform subpixel-level registration;

[0113] Step S1, current modulation image acquisition, specifically includes the following steps:

[0114] Step S11: Control the OLED driving circuit to generate a stepped current sequence, specifically:

[0115] Based on a preset base drive current value, K different current values ​​are generated using a sinusoidal modulation mode:

[0116]

[0117] In the formula, I k I0 is the driving current value during the kth modulation, α is the base driving current value, α is the current modulation amplitude coefficient (the base value of α is determined by the panel structure; for example, rigid OLEDs require a smaller α, while flexible OLEDs can be appropriately increased), k is the modulation sequence index, and K is the total number of modulations.

[0118] Step S12: In each current I k Below, a high-resolution CMOS sensor is used to acquire the panel's emission image F. k (x,y);

[0119] Step S13: Perform temporal registration on the image sequence, and output a precisely aligned registered image sequence for subsequent processing.

[0120]

[0121] In the formula, For the registered k-th frame image, For subpixel-level displacement correction operators (the correction operator is not uniquely determined; its specific form can be selected or designed according to actual needs), F k (x,y) represents the k-th frame of the original acquisition, F ref(x,y) is the reference image, and (x,y) is the pixel coordinate of the image.

[0122] Step S2, Multi-scale texture field construction: Extract image gradient features, construct a three-dimensional texture tensor, and generate a scale-weighted texture field through wavelet decomposition;

[0123] Step S2, the construction of the multi-scale texture field specifically includes the following steps:

[0124] Step S21: Extract gradient-invariant features for each registered image:

[0125]

[0126] In the formula, G k (x,y) represents the gradient magnitude of the k-th frame image. For gradient operators, K is the convolution operator. sobel This is the Sobel edge detection convolution kernel, where ||·||2 is the L2 norm of the vector;

[0127] Step S22: Construct the 3D texture tensor:

[0128]

[0129] In the formula, T(x,y,k) is the value of the 3D texture tensor at position (x,y,k), and G... k (x,y) represents the gradient magnitude;

[0130] Step S23: Perform wavelet decomposition along the current dimension:

[0131]

[0132] In the formula, T sc (x,y) represents the scalar values ​​of the multi-scale texture field, s is the wavelet decomposition scale index, S is the maximum decomposition scale number, and ψ s These are scale-weighted coefficients. For a one-dimensional Haar wavelet transform operator, |·| s To take the absolute value of the wavelet coefficients at scale s, T(x,y,:) is a one-dimensional signal along the current dimension at a fixed position (x,y).

[0133] Step S3, Entropy Variation Field Calculation: Combine local entropy density and multi-scale gradient, and calculate the defect-sensitive entropy variation field by combining the directional constraint matrix;

[0134] Step S3, the entropy change field calculation specifically includes the following steps:

[0135] Step S31: Define the local entropy density function:

[0136]

[0137] In the formula, ε(x,y) is the local entropy density value, i is the gray level, and p i Let be the probability of gray level i within a 5×5 window, μ be the mean gray level of the pixels within the window, σ be the standard deviation of the gray level of the pixels within the window, and exp(·) be the fourth-order Lorentz kernel function.

[0138] Step S32: Calculate the multiscale entropy-variable field:

[0139]

[0140] in,

[0141] In the formula, ΔE(x,y) is the entropy-varying field output value, s is the wavelet decomposition scale index, and S is the maximum decomposition scale number. Let R be the scale-adaptive gradient operator, |·| be the modulus of the complex number, and R be the gradient operator. s (x,y) is the texture direction constraint matrix, θ s is the texture direction angle, and arg(·) is the argument of the complex gradient vector.

[0142] Step S4, Defect Feature Enhancement: The defect region is nonlinearly enhanced using a Sigmoid-power-law composite function, and anisotropic diffusion filtering is performed to optimize the signal-to-noise ratio;

[0143] Step S4, defect feature enhancement specifically includes the following steps:

[0144] Step S41: Construct a nonlinear enhancement function:

[0145]

[0146] In the formula, M(x,y) is the enhanced defect feature map, β is the Sigmoid steepness coefficient, τ is the activation threshold, and ‖·‖ is the absolute value of the scalar. The Sigmoid steepness coefficient is a general setting for typical OLED defects (such as Mura and microcracks) and can be regarded as a constant. In practical applications, it can be adjusted according to the panel type. The activation threshold is derived from statistics. If the detection environment noise is higher (such as in an industrial field), it can be adjusted and increased accordingly. These two values ​​are generally determined by experience or experiment.

[0147] Step S42: Perform anisotropic diffusion filtering:

[0148]

[0149] In the formula, denoted as the rate of change of the feature map over time, div is the divergence operator, c(z) is the diffusion coefficient function, z is the independent variable of the diffusion function, and κ is the diffusion threshold parameter, which is a fixed constant and is usually an empirical value verified in the laboratory.

[0150] Step S5, Subpixel Boundary Localization: Extract candidate regions on the enhanced image, and achieve subpixel boundary fitting by iteratively solving the energy functional based on the improved Snake model;

[0151] Step S5, sub-pixel boundary localization specifically includes the following steps:

[0152] Step S51: Extract candidate defect regions on the enhanced image M(x,y);

[0153] Step S52: Calculate the boundary energy functional for each connected component Ω:

[0154]

[0155] In the formula, B is the boundary energy functional value. For the arc length parameter of the boundary curve, The curve represents the defect boundary, where α and β are energy weighting coefficients, both of which are preset fixed constants. For feature maps at boundary points The value, The tangent vector of the boundary curve;

[0156] Step S53: Solve iteratively using the improved Snake model:

[0157]

[0158] In the formula, v is the boundary point coordinate vector, t is the iteration time step, and γ is the elasticity coefficient. Let λ be the arc length parameter of the boundary curve, and λ be the gradient weight of the feature map. Let ε be the gradient of the feature map at point v, ε be the weight of the normal vector field, and n(v) be the unit normal vector derived from the entropy-varying field.

[0159] Step S6, Defect Quantitative Analysis: Calculate morphological and optical characteristic parameters such as defect area, contrast, and spectral anomaly index;

[0160] Step S6, the defect quantification analysis specifically includes the following steps:

[0161] Step S61: Calculate defect morphological parameters:

[0162]

[0163] In the formula, A is the area of ​​the defect region, Ω is the set of defective pixels, δx and δy are the physical dimensions of the pixels, C is the optical contrast of the defect region, and maxΩ M represents the maximum pixel value of the enhanced feature map M(x,y) within the defect region Ω, min Ω M is the minimum pixel value of the enhanced feature map M(x,y) within the defect region Ω, and max Panel M is the global maximum pixel value of the enhanced feature map M(x,y) across the entire panel image range;

[0164] Step S62: Extract the spectral anomaly index:

[0165]

[0166] In the formula, SAI is the spectral anomaly index, |Ω| is the total number of pixels in the defect region, and T sc (x,y) represents the multi-scale texture field values. The average value of the texture field in the background region, ||·|| F It is the Frobenius norm.

[0167] Step S7, Classification Decision and Output: Classify the feature vectors using a decision tree based on a predefined set of rules, identify the defect type, and generate a report;

[0168] Step S7, classification decision and output, specifically includes the following steps:

[0169] Step S71: Input feature vector: f = [A, C, SAI, roundness, aspect ratio];

[0170] Step S72: Classify using a pre-trained decision tree model:

[0171]

[0172] In the formula, Class is the defect type label, A is the defect area, C is the contrast, and SAI is the spectral anomaly index;

[0173] Step S73: Generate the defect coordinate mapping map D(x,y) and output a JSON format inspection report.

[0174] One specific application of this embodiment is:

[0175] Test subject: 6.8-inch OLED panel, resolution 3168×1440;

[0176] Step S1: Current Modulation Image Acquisition

[0177] Set the base drive current I0 = 15mA, modulation parameter α = 0.05, and cycle number K = 8; generate the current sequence:

[0178]

[0179] The obtained current values ​​(mA) are: 15.00, 15.74, 16.32, 16.67, 16.74, 16.52, 16.04, 15.37;

[0180] Images were captured in a darkroom environment using an 8K CMOS camera:

[0181] The object distance was fixed at 300mm; three images were acquired at each current (exposure time 50ms); the median filter result was taken as the F. k (x,y);

[0182] Image registration (based on k=0):

[0183]

[0184] in,

[0185] Typical registration parameters:

[0186] k value Δx (pixels) Δy (pixels) θ(°) RMS error 1 +0.35 -0.18 +0.03 0.025 2 +0.28 -0.22 -0.01 0.027 3 -0.31 +0.15 +0.04 0.026 4 -0.42 +0.27 -0.05 0.028 5 -0.38 +0.33 +0.02 0.030 6 +0.25 -0.29 -0.03 0.029 7 +0.19 -0.31 +0.02 0.031

[0187] Registration quality verification:

[0188] Calculate the normalized cross-correlation values ​​for all image pairs:

[0189]

[0190] Result: NCC k >0.998 (k=1,...,7);

[0191] Generate a spatial transformation field:

[0192]

[0193] In the formula, C k Let D be the global affine transformation matrix. k (x,y) represents a local non-rigid deformation field (maximum deformation < 0.1 pixels);

[0194] Final output registered image sequence:

[0195] Spatial alignment accuracy < 0.3 pixels.

[0196] Step S2: Multi-scale texture field construction

[0197] Calculate the gradient magnitude (taking the image with k=4 as an example):

[0198]

[0199] (Sobel operator convolution kernel size 3×3);

[0200] Construct the texture tensor (along the current dimension): T(1000,500,4)=(G2-2G4+G6)×G4=(42.1-2×38.6+40.3)×38.6=196.2;

[0201] Three-scale wavelet decomposition (Haar wavelet): T sc (1000, 500.) = 2 -0.5 ×|d1|+2 -1 ×|d2|+2 -1.5 ×|d3|=0.707×12.3+0.5×8.7+0.354×5.2=16.8;

[0202] Where d1, d2, and d3 are detail coefficients;

[0203] Step S3: Original Entropy Variation Field Calculation

[0204] Local entropy density calculation (5×5 window): The mean gray value μ at the center of the window (1000, 500) is 127.3; the standard deviation σ is 15.2.

[0205] Calculate the entropy value:

[0206] A grayscale matrix of 25 pixels is extracted centered on the target pixel. The frequency H(i) of each grayscale level i is calculated. Calculate the probability distribution, and finally calculate the entropy value:

[0207]

[0208] (The histogram is divided into 16 intervals);

[0209] Entropy-variant field calculation (scale s = 2): Scale gradient

[0210] Texture orientation angle

[0211] Directional constraint matrix:

[0212]

[0213] Entropy-variable field components:

[0214] Final amplitude:

[0215] Step S4: Defect Feature Enhancement

[0216] Nonlinear enhancement (parameters β = 0.8, τ = 0.05):

[0217] Anisotropic diffusion (5 iterations):

[0218] initial gradient

[0219] Conductivity:

[0220] Pixel value update amount after diffusion:

[0221] Step S5: Subpixel boundary localization

[0222] Candidate region extraction: threshold segmentation (threshold = 0.001); obtain the microcrack connected region Ω (85 pixels);

[0223] Calculate the boundary energy functional:

[0224]

[0225] Snake model iteration (γ = 0.1):

[0226]

[0227] Where the normal vector field n = [-0.342, 0.940] (derived from step S3);

[0228] Step S6: Defect Quantification Analysis

[0229] Morphological parameter calculation:

[0230] Actual area: A = 85 × (3.4 μm) 2 ≈0.000983mm 2 ;

[0231] Contrast Ratio: (Maximum panel M = 0.02);

[0232] Spectral anomaly index: (background mean)

[0233] Step S7: Classification Decision and Output

[0234] Input feature vector: f = [0.000983, 0.7, 1.92, 0.15, 8.5] (roundness, aspect ratio);

[0235] Decision rule trigger: Micro-crack ← (SAI>1.8) ∧ (Aspect Ratio>5) ∧ (10 2 μm 2 ≤A≤10 4 μm 2),Right now:

[0236] Micro-crack meets the following conditions: SAI > 1.8; and aspect ratio > 5 and 10. 2 μm 2 ≤A≤10 4 μm 2 ;

[0237] Finally, a test report is generated:

[0238] One microcrack was detected, located near panel coordinates (1234.56, 789.01), with a length of 85.3 micrometers and a width of 2.5 micrometers. Its outline is precisely described by 32 sub-pixel points. The system determined this defect to be a microcrack with a 96% confidence level. The overall defect density of the panel is 0.32 defects / mm², and the quality score is 94.2 out of 100.

[0239] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0240] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A machine vision-based OLED defect detection device, characterized in that, The detection equipment includes a current modulation control module, an image acquisition module, an image registration module, a multi-scale texture analysis module, an entropy change field calculation module, a defect enhancement module, a sub-pixel localization module, a defect quantization module, a classification decision module, and a result output module. The current modulation control module, image acquisition module, image registration module, multi-scale texture analysis module, entropy change field calculation module, defect enhancement module, sub-pixel positioning module, defect quantization module, classification decision module, and result output module are connected in sequence. The output end of the multi-scale texture analysis module is unidirectionally connected to the defect quantization module, and the output end of the entropy change field calculation module is unidirectionally connected to the sub-pixel positioning module. The workflow of the testing equipment is as follows: Step S1, Current Modulation Image Acquisition: Drive the OLED panel with a sinusoidal modulated current sequence, simultaneously acquire multiple luminescent images and perform subpixel-level registration; Step S2, Multi-scale texture field construction: Extract image gradient features, construct a three-dimensional texture tensor, and generate a scale-weighted texture field through wavelet decomposition; Step S3, Entropy Variation Field Calculation: Combine local entropy density and multi-scale gradient, and calculate the defect-sensitive entropy variation field by combining the directional constraint matrix; Step S4, Defect Feature Enhancement: The defect region is nonlinearly enhanced using a Sigmoid-power-law composite function, and anisotropic diffusion filtering is performed to optimize the signal-to-noise ratio; Step S5, Subpixel Boundary Localization: Extract candidate regions on the enhanced image, and achieve subpixel boundary fitting by iteratively solving the energy functional based on the improved Snake model; Step S6, Defect Quantitative Analysis: Calculate defect area, contrast, and spectral anomaly index morphological and optical characteristic parameters; Step S7, Classification Decision and Output: Classify the feature vectors using a decision tree based on a predefined set of rules, identify the defect type, and generate a report.

2. The OLED defect detection device based on machine vision according to claim 1, characterized in that, The current modulation control module is used to generate a sinusoidal modulated current sequence to drive the OLED panel and excite the defect response characteristics under different brightness states. The image acquisition module is used to synchronously acquire OLED light-emitting images under each current state to obtain the original defect data sequence; The image registration module is used to perform sub-pixel level displacement correction on multiple frames of images to eliminate pixel-level offset errors caused by mechanical vibration. The multi-scale texture analysis module is used to extract gradient features and construct a three-dimensional texture tensor, and obtain multi-scale texture features through wavelet decomposition. The entropy change field calculation module is used to calculate the local entropy density change field with directional constraints, thereby enhancing the defect region; The defect enhancement module is used to apply nonlinear transformation and anisotropic diffusion filtering to improve the contrast between defects and the background. The subpixel positioning module accurately extracts defect boundaries based on the improved Snake model, achieving superpixel precision contour positioning. The defect quantization module is used to calculate the morphological parameters and spectral anomaly index of defects and generate quantized feature vectors. The classification decision module is used to automatically identify defect types based on feature vectors and pre-trained decision tree models; The result output module is used to generate a spatial coordinate mapping map and a structured detection report, outputting visualized and machine-readable results.

3. The OLED defect detection equipment based on machine vision according to claim 1, characterized in that, Step S1, current modulation image acquisition, specifically includes the following steps: Step S11: Control the OLED driving circuit to generate a stepped current sequence, specifically based on the preset base driving current value, and use a sine wave modulation mode to generate K different current values; Step S12: In each current I k Below, images of the panel's light emission are acquired using a high-resolution CMOS sensor; Step S13: Perform temporal registration on the image sequence and output a precisely aligned registered image sequence for subsequent processing.

4. The OLED defect detection device based on machine vision according to claim 1, characterized in that, Step S2, the construction of the multi-scale texture field, specifically includes the following steps: Step S21: Extract gradient-invariant features for each registered image. Specifically, calculate the spatial derivative of the image using the Sobel operator and obtain the Euclidean norm of the gradient magnitude to form a basic gradient feature map. Step S22: Construct a three-dimensional texture tensor to form a space-current three-dimensional tensor; Step S23: Perform wavelet decomposition along the current dimension, extract the absolute values ​​of each scale component, and apply exponentially decaying scale weighting coefficients to generate scale-invariant texture field features.

5. The OLED defect detection device based on machine vision according to claim 1, characterized in that, Step S3, the entropy change field calculation, specifically includes the following steps: Step S31: Define the local entropy density function and calculate the local entropy density value for each pixel; Step S32: Calculate the multi-scale entropy change field, average the magnitude of the entropy density gradient under direction constraints at each scale, and generate an entropy change field map.

6. The OLED defect detection device based on machine vision according to claim 1, characterized in that, Step S4, defect feature enhancement, specifically includes the following steps: Step S41: Construct a nonlinear enhancement function to suppress background noise while enhancing the response of the defect area; Step S42: Perform anisotropic diffusion filtering, construct the diffusion coefficient based on the gradient-sensitive fourth-power exponential decay function, and describe the filtering process through partial differential equations.

7. The OLED defect detection device based on machine vision according to claim 1, characterized in that, Step S5, sub-pixel boundary localization, specifically includes the following steps: Step S51: Extract the connected regions of candidate defects on the defect feature map after nonlinear enhancement; Step S52: Construct the boundary energy functional. Calculate the boundary energy functional for each connected domain Ω to provide an energy minimization convergence target for subsequent active contour iterations and guide sub-pixel level boundary localization. Step S53: Use the improved Snake model to solve iteratively. The iteration process continues until the rate of change of the total energy at the boundary is less than the threshold, and the closed contour with sub-pixel accuracy is output.

8. The OLED defect detection device based on machine vision according to claim 1, characterized in that, Step S6, the defect quantification analysis, specifically includes the following steps: Step S61: Calculate the morphological parameters of the defects to provide a basis for quantifying the differences in physical size and visual significance for subsequent defect classification; Step S62: Extract the spectral anomaly index and quantify the dynamic response differences between the defective region and the normal background in multi-scale texture features.

9. The OLED defect detection device based on machine vision according to claim 1, characterized in that, Step S7, the classification decision and output, specifically includes the following steps: Step S71: Input the feature vector obtained from the quantitative analysis. The feature vector includes five parameters: defect area, contrast, spectral anomaly index, roundness, and aspect ratio. Step S72: Classify using a pre-trained decision tree model. The classification process is implemented using a decision tree model, and the output is a structured result containing defect type labels. Step S73: Generate a defect coordinate mapping map and output a JSON-formatted inspection report.

Images