Organic electroluminescent substrate detection method and system based on image processing

Abstract

The invention provides an organic electroluminescent substrate detection method and system based on image processing, and relates to the technical field of organic electroluminescent substrate detection. Comprising the following steps: carrying out image feature extraction on preprocessed image information by using key point detection; establishing a defect detection model, and training the defect detection model by adopting historical image information; identifying feature defects in the image features through the trained defect detection model; feature defects are classified and quantified; and in combination with the defect classification data, establishing association between the defect degree data and the product quality, and evaluating the quality grade of the product. According to the method, the defect detection model is combined with the chaos evolution optimization algorithm, and the accuracy of defect classification is remarkably improved by dynamically adjusting the learning rate and enhancing the generalization ability of the model; and the crossing from qualitative identification to quantitative evaluation is realized through comprehensive defect indexes, fatal defects and repairable defects can be accurately distinguished, and dynamic adjustment of production parameters is supported.

Description

Technical Field

[0001] The present invention relates to the technical field of organic electroluminescent substrate detection, and particularly to an organic electroluminescent substrate detection method and system based on image processing. Background Art

[0002] An organic electroluminescent substrate (Organic Electroluminescent Substrate, abbreviated as OEL Substrate) refers to a substrate material used in organic electroluminescent (OLED) devices. This substrate plays a crucial role in the operation of OLEDs and is commonly used in display screens and lighting devices. Organic electroluminescent substrates are usually made of glass or plastics (such as PET or PC), which have good light transmittance and appropriate mechanical strength. In some applications, a conductive layer, such as a transparent conductive oxide (ITO), may be coated on the substrate to help the current distribute evenly. The substrate needs to provide appropriate conductivity to support the injection and migration of electrons and holes of the organic light-emitting materials. It needs to have good light transmittance to emit light effectively, especially in display and lighting applications. The substrate must have good chemical and thermal stability to ensure the service life and reliability of the OLED device.

[0003] When detecting an organic electroluminescent substrate, in addition to detecting its optoelectronic properties, chemical properties, and physical properties, it is also necessary to detect the defects on its appearance.

[0004] Although general organic electroluminescent substrate detection can detect the appearance of the substrate to a certain extent, due to the lack of equipment conditions, it may not be sensitive enough to tiny defects, resulting in these problems not being identified in time, thus affecting the overall quality of the product. Detection algorithms based on traditional image processing methods may not be able to identify complex defect patterns. Especially, the organic materials on the substrate may cause stripes or interference, thus affecting the accurate identification of the algorithm. Summary of the Invention

[0005] The present invention provides an organic electroluminescent substrate detection method and system based on image processing to solve the defects existing in the prior art.

[0006] On the one hand, the present invention provides an organic electroluminescent substrate detection method based on image processing, including: Collecting image information of the organic electroluminescent substrate and preprocessing the image information to generate preprocessed image information; Performing image feature extraction on the preprocessed image information using key point detection; Establishing a defect detection model based on ResNet and training the defect detection model using historical image information; Input the image features into the trained defect detection model to output feature defects; Classify and quantify the feature defects to output defect classification data and defect degree data; Combine the defect classification data to establish the correlation between the defect degree data and the product quality, and evaluate the quality grade of the product.

[0007] According to an organic electroluminescent substrate detection method based on image processing provided by the present invention, the steps of preprocessing the image information include: Use a filter to remove the noise in the image and output a low-noise image; Use adaptive histogram equalization to enhance the contrast of the low-noise image and output a high-contrast image; Sharpen the high-contrast image.

[0008] According to an organic electroluminescent substrate detection method based on image processing provided by the present invention, the steps of image feature extraction include: Perform multi-scale processing on the preprocessed image through Gaussian blur to form a scale space; Identify local extreme points in the scale space and find image key points through Difference of Gaussians (DoG); Assign multiple directions to each key point; Calculate the SIFT feature vectors for each direction and generate descriptors for the regions around each key point.

[0009] According to an organic electroluminescent substrate detection method based on image processing provided by the present invention, the method of forming a scale space includes: Set a Gaussian function, perform Gaussian blur on the preprocessed image at different scales to generate Gaussian images; Generate multiple blurred images by convolving the preprocessed image information with Gaussian functions of different standard deviations; Select multiple groups of different standard deviations, apply Gaussian blur to the original image to obtain multiple layers of blurred images; Combine the multiple layers of blurred images into a pyramid structure, with the bottom layer being the preprocessed image and the subsequent layers being blurred images with gradually increasing standard deviations; Calculate the differences between adjacent scale blurred images to generate Difference of Gaussian (DoG) images.

[0010] According to an organic electroluminescent substrate detection method based on image processing provided by the present invention, the steps of finding image key points through Difference of Gaussians (DoG) include: Generate a DoG image for key point detection based on the Difference of Gaussian (DoG) image; In the DoG image, traverse each pixel and compare it with the surrounding neighborhood to find local maxima and minima, and generate initial key points; Calculate the Hessian matrix for each initial key point; Generate image key points based on the intensity of local extrema and the calculation results of the Hessian matrix.

[0011] According to a method for detecting an organic electroluminescent substrate based on image processing provided by the present invention, the steps of training a defect detection model include: Divide the defect data in the historical image data into a training set and a validation set; Set the network depth, loss function, learning rate, and accuracy of the defect detection model; Use the training set to train the defect detection model; Use the chaotic evolution optimization algorithm to optimize the learning rate; Use the validation set to verify the loss of the defect detection model. When the loss no longer significantly decreases during the training cycle, stop the training process.

[0012] According to a method for detecting an organic electroluminescent substrate based on image processing provided by the present invention, the steps of optimizing the learning rate include: Set an initial learning rate; Utilize the hyperchaotic characteristics of the two-dimensional memory chaotic ultra-efficient mapping to determine the evolution direction of the initial learning rate; Use the binomial crossover operation to mutate the initial learning rate; Select the learning rate after the mutation operation to obtain the optimal learning rate.

[0013] According to a method for detecting an organic electroluminescent substrate based on image processing provided by the present invention, the steps of classifying and quantifying feature defects include: Classify the feature defects using a preset classification standard to generate multiple defect sets; Add defect labels to the multiple defect sets and output the defect classification data; Combine the defect classification data and the preset defect standards to evaluate the defects of all subsets within the multiple defect sets and output the defect degree data.

[0014] According to a method for detecting an organic electroluminescent substrate based on image processing provided by the present invention, the steps of evaluating the quality grade of a product include: Define the core quality parameters of the product according to different defect labels; Assign weights to each defect type and define a scoring standard according to the magnitude of the weights; Calculate the comprehensive defect index according to the scoring standard; Combine the comprehensive defect index and the preset defect disposal standards to determine the disposal measures of the product.

[0015] On the other hand, the present invention also provides an organic electroluminescent substrate detection system based on image processing, including: An image acquisition device for acquiring image information of an organic electroluminescent substrate; An image processing module for preprocessing the image information; A feature extraction module for extracting image features from the preprocessed image information; An intelligent processing module for detecting image features through a trained defect detection model; the intelligent processing module includes a model training unit and a defect detection unit. The model training unit is used to train the defect detection model in combination with historical image information, and the defect detection unit is used to identify feature defects in the image features; A defect quantification module for classifying and quantifying the feature defects, and outputting defect classification data and defect degree data; the defect quantification module includes a defect classification unit and a defect quantification unit. The defect classification unit is used to classify the feature defects, and the defect quantification unit is used to quantify the classified feature defects; A quality assessment module for calculating a comprehensive defect index of the organic electroluminescent substrate according to a preset scoring standard and determining a disposal measure for the product.

[0016] An organic electroluminescent substrate detection method and system provided by the present invention, by adopting a ResNet deep learning model combined with a chaotic evolutionary optimization algorithm, dynamically adjusting the learning rate and enhancing the model generalization ability, significantly improving the accuracy of defect classification. At the same time, using the multi-scale space construction and key point descriptor generation technology of the SIFT feature extraction algorithm, effectively overcoming the feature drift problem under complex scenarios such as uneven illumination and noise interference. Through the cascaded preprocessing process of adaptive histogram equalization and differential Gaussian pyramid, the image processing speed is improved. The chaotic optimization strategy based on binomial crossover operation speeds up the model convergence speed and shortens the single-batch substrate detection time. Through the comprehensive defect index, it realizes the leap from qualitative identification to quantitative evaluation, can accurately distinguish fatal defects from repairable defects, and supports dynamic adjustment of production parameters. Description of the Drawings

[0017] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained according to these drawings.

[0018] Figure 1 It is a schematic flowchart of an organic electroluminescent substrate detection method based on image processing provided by an embodiment of the present invention; Figure 2 It is a step diagram for establishing a defect detection model based on ResNet; Figure 3 It is a schematic structural diagram of an organic electroluminescent substrate detection system based on image processing provided by an embodiment of the present invention. Detailed implementation manners

[0019] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below with reference to the accompanying drawings in the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without making creative efforts shall fall within the protection scope of the present invention.

[0020] The following Figure 1 - Figure 2 describes an organic electroluminescent substrate detection method and system based on image processing of the present invention.

[0021] Embodiment 1: Figure 1 It is a schematic flowchart of an organic electroluminescent substrate detection method based on image processing provided by an embodiment of the present invention.

[0022] Figure 2 It is a step diagram for establishing a defect detection model based on ResNet.

[0023] As Figure 1 shown, an organic electroluminescent substrate detection system and method based on image processing provided by an embodiment of the present invention, the execution subject may be an organic electroluminescent substrate detection method based on image processing, including: Collect image information of the organic electroluminescent substrate and preprocess the image information. High-resolution industrial cameras are usually used to collect image information to ensure sufficient pixel density to capture micron-level defects. Select the transmission mode or the reflection mode according to the defect types in history. For the detection of dark or bright spots, uniform backlighting is used to highlight the brightness abnormality. For line defects or color differences, multi-spectral light sources or polarized light are used to assist in the analysis. The steps for preprocessing the image information include: Use a filter to remove the noise in the image and output a low-noise image. During the acquisition and transmission of the image, it may be interfered by noise from different sources, such as Gaussian noise, salt-and-pepper noise, etc. Using a filter can effectively remove this noise and improve the image quality. Common filtering methods include mean filtering, median filtering, Gaussian filtering, etc. After filtering, the noise level of the image is reduced, providing a more reliable input for subsequent processing.

[0024] Enhance the contrast of low-noise images using adaptive histogram equalization and output high-contrast images. By adjusting the histogram of the image, the brightness distribution becomes more uniform, thereby enhancing the contrast of the image and making the details in the image clearer.

[0025] Sharpen the high-contrast image. Sharpening can enhance the edges and details of the image, making the image clearer. Sharpening is usually achieved through high-pass filters, such as Laplacian filtering, Sobel filtering, etc. These filters can highlight the high-frequency components in the image, that is, the edge and detail parts, thereby enhancing the clarity of the image.

[0026] Use key-point detection to extract image features from the preprocessed image information. The steps include: Perform multi-scale processing on the preprocessed image through Gaussian blur to form a scale space. This step is to simulate the performance of the image at multiple scales, so as to better capture the key features in the image. The methods for forming the scale space include: Set the Gaussian function and generate a set of Gaussian images by performing Gaussian blur on the preprocessed image at different scales. The formula of the Gaussian function is expressed as:

[0027] In the formula, (x, y) are the pixel coordinates in the image, and σ is the standard deviation of Gaussian blur, which is used to control the degree of blur. The larger the value of σ, the higher the degree of blur of the image. The standard deviation determines the width of the Gaussian function, thus affecting the range of the blur effect. The value of the Gaussian function G(x, y, σ) represents the degree of blur of the image at the given pixel position (x, y).

[0028] Generate a series of blurred images by convolving the preprocessed image with Gaussian functions of different standard deviations to form a scale space. The formula is expressed as:

[0029] In the formula, L(x, y, σ) represents the blurred image, I(x, y) is the original image, and * represents the convolution operation. The convolution operation is an integral process, that is, multiplying the template by the pixel values at this position in the image and then adding up all these products to obtain the pixel value of the blurred image at this position. The convolution operation can apply the blur effect of the Gaussian function to the original image, thereby generating a blurred version of the image, and this image is an element in the scale space. Repeating this process at different standard deviations σ can construct the entire scale space.

[0030] Select multiple different standard deviations σ, apply Gaussian blur to the original image, and obtain multiple blurred images. These standard deviations σ are usually selected according to a certain proportional relationship. For example, the standard deviation is doubled each time. In this way, each standard deviation value corresponds to a different degree of image blur. These images are arranged in ascending order of blur degree, forming a multi-level image set.

[0031] Combine the multiple blurred images into a pyramid structure, with the bottom layer being the preprocessed image and the subsequent layers being the blurred images with gradually increasing standard deviations. In this structure, the bottom layer is the preprocessed original image or a slightly blurred image. As the number of layers increases, the degree of image blur also gradually increases. Such a pyramid structure allows us to analyze the image at different scales, thereby capturing information from fine details to large structures.

[0032] Calculate the difference between adjacent scale blurred images to form a Difference of Gaussian (DoG) image. The formula is expressed as:

[0033] In the formula, σ1 and σ2 represent two adjacent scale parameters (i.e., standard deviations).

[0034] Identify local extreme points in the scale space and find image key points through the Difference of Gaussian. The steps for identifying local extreme points include: Define a neighborhood range (usually a 3x3x3 cube, including the near plane and adjacent scales in space), so that the current pixel D(x,y,σ) can be compared with its 8 adjacent pixels and the pixels at the same position in the adjacent upper and lower scales, and a total of 26 pixel values need to be considered.

[0035] Traverse each pixel in D(x,y,σ) and perform local neighborhood analysis to ensure that all pixels are evaluated, thereby detecting possible key points and ensuring complete coverage of the image information.

[0036] Compare the value of the current pixel D(x,y,σ) with the values of the other 26 pixels in its neighborhood to determine whether this point is a local extreme.

[0037] Determine whether the current pixel D(x,y,σ) is greater (local maximum) or less (local minimum) than all pixel values in its neighborhood. If so, mark the current pixel as a potential key point.

[0038] After local extreme value judgment, set an intensity threshold to eliminate key points that are not significant enough. The screening criteria are usually directly based on the comparison results of the previous steps to ensure that the retained key points have sufficient significance and robustness in terms of feature performance.

[0039] The steps to find image key points through Difference of Gaussians include: Calculate the Difference of Gaussians (DoG) image between two adjacent scales to form the DoG image for key point detection. This process can simulate the local extreme points in the scale space, and these extreme points correspond to the key points in the image, such as corners, edges, etc.

[0040] In the DoG image, traverse each pixel and compare it with the surrounding neighborhood to find local maxima and minima, generating key points.

[0041] To determine these key points, we need to calculate the Hessian matrix for each key point. The Hessian matrix is a second - order derivative matrix, which can help us judge the nature of the extreme point (maximum or minimum). The definition of the Hessian matrix is as follows:

[0042] Where \(L_{xx}\), \(L_{xy}\), and \(L_{yy}\) are the second - order derivatives of the image along the x - axis and y - axis at this point respectively. These derivatives can be calculated through the derivatives of the Gaussian function. For each extreme point, use the trace and determinant of the Hessian matrix to eliminate unstable edge response points.

[0043] The trace - determinant of the Hessian matrix is expressed as:

[0044] Where \(Tr(H)\) is the trace of the Hessian matrix, that is, the sum of the diagonal elements of the matrix. The trace can be used to judge the nature of the extreme point. If the trace is positive, the extreme point tends to be an edge; if the trace is negative, the extreme point tends to be a corner.

[0045] The determinant of the Hessian matrix is expressed as:

[0046] Where \(Det(H)\) is the determinant of the Hessian matrix, that is, the product of all eigenvalues of the matrix. The determinant can be used to judge the stability of the extreme point. If the determinant is positive, the extreme point is stable in the scale space. If the determinant is negative, the extreme point may be caused by noise or edges and is not stable enough.

[0047] Generate the final set of key points according to the intensity of the local extreme and the calculation results of the Hessian matrix.

[0048] Assign multiple directions to each key point. This step is to ensure the rotation invariance of SIFT features. By calculating the gradient direction histogram within the neighborhood of the key point, the main direction of the key point can be determined. In addition, to increase the number of feature points, the key point can also be copied to nearby scale space positions, and a direction is assigned to each copied point.

[0049] By calculating the SIFT feature vector, a descriptor for the area around each key point is generated. The descriptor is a unique feature representation of the key point, which enables the key point to be matched and recognized between different images. The SIFT feature vector is a 128-dimensional vector, which is constructed by statistically analyzing the gradient direction and magnitude in the area around the key point. Specifically, the area around the key point is divided into 16 small sub-regions, and the gradient direction and magnitude within each sub-region are statistically analyzed into an 8-dimensional feature vector. In this way, the entire area around the key point constitutes a 128-dimensional SIFT feature vector. This feature vector not only contains the gradient information in the area around the key point, but also ensures the locality of the feature vector through Gaussian weighting. This means that the feature vector is robust to noise and small deformations. In addition, since the SIFT features are calculated in the scale space, they also have scale invariance.

[0050] Normalize the generated 128-dimensional feature vector to reduce the influence of illumination changes. The specific representation of the normalization process is:

[0051] In the formula, is the original 128-dimensional feature vector, is the Euclidean norm of the feature vector, is the normalized feature vector. The length of the normalized feature vector is scaled to the unit length, further enhancing the stability of the feature vector.

[0052] Finally, each key point is described by its position, scale, direction, and 128-dimensional feature vector.

[0053] As Figure 2 shown, a defect detection model based on ResNet is established and trained by combining historical image information. The steps to establish the defect detection model include: Divide the defect data in the historical image data into a training set, a validation set, and a test set.

[0054] Set the network depth, loss function, learning rate, and accuracy of the defect detection model. The loss function is expressed as follows:

[0055] Where N is the total number of the training set, C is the number of defect categories, aij is the true label of the i-th defect sample in category j, and is the predicted probability of the i-th defect sample in category j.

[0056] Use the training set to train the defect detection model, and monitor the training and validation losses as well as the accuracy.

[0057] Use the Chaotic evolution optimization (CEO) algorithm to optimize the learning rate, and improve the efficiency and final performance of model training. The main inspiration of CEO comes from the chaotic evolution process of two-dimensional discrete memristive mapping. By utilizing the hyperchaotic characteristics of the memristive mapping, a mathematical model of the CEO algorithm is established, and a random search direction for the evolution process is introduced. The overall framework of CEO includes mutation, crossover, and selection operations. CEO uses a two-dimensional discrete memristive hyperchaotic mapping to provide a mutation direction for each individual.

[0058] The steps of using the CEO algorithm to optimize the model learning rate include: Set the initial population of the learning rate and set the mutation operator of the population. The formula is expressed as follows:

[0059] Where p t is the initial learning rate, p t+1 is the mutated individual of the learning rate, b is the search step size, and d t is the evolution direction.

[0060] Utilize the hyperchaotic characteristics of the two-dimensional memory chaotic hyper-efficient mapping to determine the evolution direction of the learning rate. The formula is expressed as:

[0061]

[0062] Where p' and r' represent the chaotic positions of the two learning rate individuals p and r after mapping, and the values are in the intervals [-0.5, 0.5] and [-0.25, 0.25] respectively. In order to enable the proposed CEO algorithm to effectively utilize the hyperchaotic characteristics, the two individuals p t and r t selected in CEO must be mapped to the intervals [-0.5, 0.5] and [-0.25, 0.25]. lb and ub are the lower and upper bounds of the current population variables respectively.

[0063] Generate M chaotic candidate individuals p chaos and r chaos , and the formula is expressed as follows:

[0064]

[0065] Wherein, M represents the number of chaotic samples, m is the label value of any sample, and k is the chaotic mapping control parameter, which is used to control the exploration range and fineness of the algorithm in the search space, thereby affecting the performance and convergence speed of the algorithm. The generated chaotic candidate individuals p chaos ={p chaos1 ,...,p chaosM} and r chaos ={r chaos1 ,...,r chaosM} are inverse mapped to the actual positions p’ chaos ={p’ chaos1 ,...,p’ chaosM} and r’ chaos ={r’ chaos1 ,...,r’ chaosM}, and the mapping formula is:

[0066] Wherein, m is the label value of any sample.

[0067] The learning rate is mutated using the binomial crossover operator. The specific crossover process is expressed as:

[0068]

[0069] Wherein, and are the mutation values of the m-th sample, o is the random number generated each time, and E o is the crossover control parameter. In CEO, E o is a random number within the interval [0, 1] for each iteration.

[0070] According to the initial learning rate in the initial population, new trial vectors are generated through the selection operator as the optimal solution of the learning rate. The selection operator is as follows:

[0071]

[0072] Wherein, p t+1 and r t+1 are the states of two samples at time point t + 1, and p tr and r tr represent the states of the trial vectors generated at time point t, and these trial vectors are potential next states for evaluating whether they are better than the current state. f(pt ) and f(r t ) represent the value of the objective function f at the current state p t and r t . The objective function is used to evaluate the quality of the state. f(p t ) and f(r t ) represent the value of the objective function f at the trial vectors p tr and r tr .

[0073] Verify the loss of the defect detection model using the validation set. If the loss no longer decreases significantly during the training cycle, stop the training process.

[0074] Identify the feature defects in the image features through the trained defect detection model.

[0075] Classify and quantify the feature defects, and output the defect classification data and defect degree data. The steps for classifying and quantifying the feature defects include: Classify the feature defects using the preset classification criteria to generate multiple defect sets.

[0076] Add different defect labels to the multiple defect sets and output the defect classification data.

[0077] Combine the defect classification data and the preset defect criteria to evaluate all subsets within the multiple defect sets and output the defect degree data.

[0078] Combine the defect classification data to establish the association between the defect degree data and the product quality, and evaluate the quality grade of the product. The specific steps include: Define the core quality parameters of the product according to different defect labels.

[0079] Assign weights to each defect type and define the defect scoring criteria according to the magnitude of the weights.

[0080] According to the scoring criteria, calculate the comprehensive defect index. Expressed as:

[0081] In the formula, CDI is the comprehensive defect index, S is the number of defects, θ is the defect score, and ε is the total weight.

[0082] Determine the disposal measures for the product according to the comprehensive defect index.

[0083] In summary, this embodiment provides a method for detecting an organic electroluminescent substrate based on image processing. By adopting a ResNet deep learning model combined with a chaotic evolutionary optimization algorithm, the accuracy of defect classification is significantly improved by dynamically adjusting the learning rate and enhancing the generalization ability of the model. At the same time, the multi-scale space construction and key point descriptor generation techniques of the SIFT feature extraction algorithm are used to effectively overcome the feature drift problem in complex scenarios such as uneven illumination and noise interference. The image processing speed is improved through the cascaded preprocessing process of adaptive histogram equalization and differential Gaussian pyramid. The chaotic optimization strategy based on binomial crossover operation speeds up the model convergence speed and shortens the single-batch substrate detection time. By means of the comprehensive defect index, the leap from qualitative recognition to quantitative evaluation is achieved, and fatal defects and repairable defects can be accurately distinguished, supporting the dynamic adjustment of production parameters.

[0084] Example 1: Highlight defect: luminance value > threshold (200 cd / ㎡), area < 50 μ㎡.

[0085] Line defect: length > 200 μm, gray level gradient difference > 15%.

[0086] The specific classification results are shown in Table 1: Table 1:

[0087] According to the comprehensive defect index calculation formula, the comprehensive defect index of this substrate can be obtained as 3.3. Then, according to the preset grading standard (shown in Table 2), the disposal measures can be obtained.

[0088] Table 2:

[0089] The comprehensive defect index of the current substrate is 3.3. According to Table 2, the quality grade belongs to Class B, so random inspection and retest are required.

[0090] Embodiment 2: Based on the same general inventive concept, the present invention also protects a system for detecting an organic electroluminescent substrate based on image processing. The following describes a system for detecting an organic electroluminescent substrate based on image processing provided by the present invention. The system for detecting an organic electroluminescent substrate based on image processing described below can be mutually corresponded and referred to the method for detecting an organic electroluminescent substrate based on image processing described above.

[0091] Figure 3 It is a schematic structural diagram of a system for detecting an organic electroluminescent substrate based on image processing provided by an embodiment of the present invention.

[0092] As Figure 3As shown in the figure, an organic electroluminescent substrate detection system based on image processing provided by an embodiment of the present invention includes: An image acquisition device for acquiring image information of an organic electroluminescent substrate. To ensure the clarity and accuracy of the image, the image acquisition device usually needs to have the characteristics of high resolution and high sensitivity.

[0093] An image processing module for preprocessing the image information to improve the quality of the image, making subsequent feature extraction and defect detection more accurate and reliable.

[0094] A feature extraction module for extracting image features from the preprocessed image information. The features may include information such as the texture, shape, color, and edges of the image. Effective feature extraction can significantly improve the accuracy and efficiency of defect detection.

[0095] An intelligent processing module for detecting image features through a trained defect detection model. The intelligent processing module includes a model training unit and a defect detection unit. The model training unit uses historical image information to train the defect detection model to improve the detection ability and accuracy of the model. The defect detection unit is responsible for identifying feature defects in the image features, which is a key step in defect detection.

[0096] A defect quantification module for classifying and quantifying the feature defects, and outputting defect classification data and defect degree data. The defect quantification module includes a defect classification unit and a defect quantification unit. The defect classification unit is used to classify the feature defects, usually classified according to the type, size, shape, etc. of the defects. The defect quantification unit is responsible for quantifying the classified feature defects, that is, evaluating the severity of the defects, which is usually achieved by quantifying parameters such as the area, length, and density of the defects.

[0097] A quality evaluation module for calculating the comprehensive defect index of the organic electroluminescent substrate according to a preset scoring standard and determining the disposal measures for the product. The output of the module is the final result of the entire system, which is directly related to the quality control and quality improvement of the product. The quality evaluation module needs to consider various types and severities of defects, as well as their impacts on product performance, so as to give a comprehensive defect index. Based on this index, the producer can decide the next disposal of the product, such as whether reprocessing, repair, or scrapping is required.

[0098] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. A person of ordinary skill in the art can understand and implement it without creative efforts.

[0099] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the above technical solution, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0100] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An organic electroluminescent substrate detection method based on image processing, characterized in that Including: Collecting image information of an organic electroluminescent substrate and preprocessing the image information to generate preprocessed image information; Performing image feature extraction on the preprocessed image information using keypoint detection; Establishing a defect detection model based on ResNet and training the defect detection model using historical image information; Inputting the image features into the trained defect detection model to output feature defects; Classifying and quantifying the feature defects to output defect classification data and defect degree data; Combining the defect classification data to establish an association between the defect degree data and product quality and evaluating the quality grade of the product.

2. The organic electroluminescent substrate detection method based on image processing according to claim 1, wherein, The steps of preprocessing the image information include: Using a filter to remove noise in the image and output a low-noise image; Using adaptive histogram equalization to enhance the contrast of the low-noise image and output a high-contrast image; Performing sharpening processing on the high-contrast image.

3. The organic electroluminescent substrate detection method based on image processing according to claim 1, wherein, The steps of performing the image feature extraction include: Performing multi-scale processing on the preprocessed image through Gaussian blur to form a scale space; Identifying local extreme points within the scale space and finding image keypoints through Difference of Gaussians; Assigning multiple directions to each of the keypoints; Calculating the SIFT feature vector for each direction to generate a descriptor for the area around each of the keypoints.

4. The organic electroluminescent substrate detection method based on image processing according to claim 3, characterized in that, The method of forming the scale space includes: Setting a Gaussian function and performing Gaussian blur on the preprocessed image at different scales to generate Gaussian images; Generating multiple blurred images by convolving the preprocessed image information with Gaussian functions of different standard deviations; Selecting multiple groups of different standard deviations and applying Gaussian blur to the original image to obtain multiple layers of blurred images; Combining the multiple layers of blurred images into a pyramid structure with the bottom layer being the preprocessed image and subsequent layers being blurred images with gradually increasing standard deviations; Calculating the difference between adjacent scale blurred images to generate Difference of Gaussians images.

5. A method for detecting an organic electroluminescent substrate based on image processing according to claim 4, characterized in that, The steps of finding the image keypoints through Difference of Gaussians include: Generating a DoG image for keypoint detection based on the Difference of Gaussians image; In the DoG image, traversing each pixel and comparing it with the surrounding neighborhood to find local maxima and minima to generate initial keypoints; Calculating the Hessian matrix for each of the initial keypoints; Generating the image keypoints based on the intensity of the local extreme points and the calculation results of the Hessian matrix.

6. The method for detecting an organic electroluminescent substrate based on image processing according to claim 4, characterized in that, The steps of training the defect detection model include: Dividing the defect data in the historical image data into a training set and a validation set; Setting the network depth, loss function, learning rate, and accuracy of the defect detection model; Training the defect detection model using the training set; Optimizing the learning rate using a chaotic evolutionary optimization algorithm; Validating the loss of the defect detection model using the validation set and stopping the training process when the loss no longer significantly decreases during the training period.

7. The organic electroluminescent substrate detection method based on image processing according to claim 6, characterized in that, The steps of optimizing the learning rate include: Setting an initial learning rate; Using the hyperchaotic characteristics of a two-dimensional memory chaotic ultra-efficient mapping to determine the evolutionary direction of the initial learning rate; Perform mutation operation on the initial learning rate using binomial crossover operation; Select the learning rate after the mutation operation to obtain the optimal learning rate.

8. A method for detecting an organic electroluminescent substrate based on image processing according to claim 1, characterized in that, The steps of classifying and quantifying the feature defects include: Classify the feature defects using a preset classification criterion to generate multiple defect sets; Add defect labels to the multiple defect sets and output defect classification data; Combine the defect classification data and the preset defect criterion to evaluate the defects of all subsets within the multiple defect sets and output defect degree data.

9. The method for detecting an organic electroluminescent substrate based on image processing according to claim 8, wherein The steps of performing quality level evaluation of the product include: Define the core quality parameters of the product according to different defect labels; Assign weights to each defect type and define a scoring criterion according to the magnitudes of the weights; Calculate the comprehensive defect index according to the scoring criterion; Combine the comprehensive defect index and the preset defect handling criterion to determine the handling measures of the product.

10. An organic electroluminescent substrate detection system based on image processing, which adopts an organic electroluminescent substrate detection method based on image processing according to any one of claims 1 to 9, characterized in that, The detection system includes: An image acquisition device for acquiring the image information of the organic electroluminescent substrate; An image processing module for preprocessing the image information; A feature extraction module for extracting image features from the preprocessed image information; An intelligent processing module for detecting the image features through a trained defect detection model; the intelligent processing module includes a model training unit and a defect detection unit, the model training unit is used to train the defect detection model in combination with historical image information, and the defect detection unit is used to identify the feature defects in the image features; A defect quantification module for classifying and quantifying the feature defects and outputting defect classification data and defect degree data; the defect quantification module includes a defect classification unit and a defect quantification unit, the defect classification unit is used to classify the feature defects, and the defect quantification unit is used to quantify the classified feature defects; A quality evaluation module for calculating the comprehensive defect index of the organic electroluminescent substrate according to a preset scoring criterion and determining the handling measures of the product.

Images