OLED screen low contrast mura defect detection method

Abstract

This invention relates to a method for detecting low-contrast Mura defects in OLED screens, belonging to the field of OLED screen defect detection technology. It solves the problem of poor Mura defect detection accuracy caused by low contrast in existing technologies. The method includes the following steps: acquiring an actual image of the OLED screen and converting it to grayscale to obtain an initial actual grayscale image; performing image enhancement on the initial actual grayscale image to obtain a final grayscale image; inputting the final grayscale image into a trained background reconstruction network to obtain a reconstructed background image; and obtaining a defect image of the OLED screen based on the reconstructed background image and the final grayscale image.

Description

Technical Field

[0001] This invention relates to the field of OLED screen defect detection technology, and in particular to a method for detecting low-contrast Mura defects in OLED screens. Background Technology

[0002] Organic light-emitting display (OLED) is a current-driven organic light-emitting device. Compared to the widely used thin-film transistor liquid crystal display (TFT-LCD), OLED has superior characteristics such as low power consumption, thin and light panels, high brightness, and good flexibility. Its core advantages of thinness and flexibility can well meet the future consumer demand for display technology and has been regarded as one of the most promising display and lighting products of the 21st century.

[0003] OLED manufacturing processes are complex, primarily employing vapor deposition and inkjet printing. Vapor deposition can lead to various display defects due to factors such as unknown substrate materials, unstable deposition rates, and misalignment errors. Inkjet printing, on the other hand, can cause variations in pixel thickness and flatness due to changes in droplet volume, droplet offset, and film spreading, resulting in defects during display. The most common OLED defect classification is based on geometry: point defects, line defects, and Mura defects. Different defects require different detection methods. Mura defects, characterized by weak contrast and irregular shapes and areas of light and dark, are the most difficult to detect and have the lowest accuracy because they are less noticeable than the background.

[0004] Therefore, there is an urgent need for a method to detect low-contrast Mura defects in OLED screens, which can solve the problem of poor detection accuracy caused by the low contrast of Mura defects in OLED screens. Summary of the Invention

[0005] Based on the above analysis, the present invention aims to provide a method for detecting low-contrast Mura defects in OLED screens, in order to solve the problem of poor detection accuracy of Mura defects caused by low contrast in existing OLED screens.

[0006] This invention provides a method for detecting low-contrast Mura defects in OLED screens, comprising the following steps:

[0007] Acquire the actual image of the OLED screen and convert it to grayscale to obtain the initial actual grayscale image;

[0008] The initial actual grayscale image is enhanced to obtain the actual grayscale image;

[0009] The actual grayscale image is input into the trained background reconstruction network to obtain the reconstructed background image;

[0010] Based on the reconstructed background image and the actual grayscale image, defect images of the OLED screen are obtained.

[0011] Furthermore, the initial actual grayscale image is enhanced in the following way:

[0012] S21. Let the image r0 = I in the initial iteration. d ′, and set the maximum window size and the initial window size; where, I d ′ represents the initial actual grayscale image.

[0013] S22. Determine the image r in the current iteration. g-1 We can find all the local extrema in the set of local maxima and local minima; where g represents the number of iterations.

[0014] S23, Transfer image r g-1 The following steps are performed sequentially on the pixels to obtain image h. g-1 :

[0015] S231. Set the current window as the initial window;

[0016] S232. If the current window is less than or equal to the largest window, then

[0017] Centered on the current pixel, if the number of local minima in the current window is equal to the number of local maxima, then the average pixel value of all pixels in the current window is taken as the pixel value of the current pixel; otherwise, after expanding the window, return to step S231 and execute again.

[0018] If the current window is larger than the maximum window, then the average pixel value of all pixels within the current window will be used as the pixel value of the current pixel.

[0019] S24, Based on image r g-1 and image h g-1 Obtain the image IMF g-1 And let image r g =h g-1 Then return to step S22 for the next iteration, until the number of iterations reaches the set iteration threshold, and then convert the image at this point to IMF. g-1 As the actual grayscale image after image enhancement.

[0020] Furthermore, the image IMF g-1 Represented as:

[0021] IMF g-1 =r g-1 -h g-1 .

[0022] Furthermore, the size of the maximum window is U×U, where U represents:

[0023]

[0024] In the formula, N max N min Let represent the number of all local maxima and all local minima in the actual grayscale image, respectively. Represents a local maximum point i max Distance to the nearest local maximum point Represents a local minimum point i min The distance between the nearest local minimum point and the nearest local minimum point.

[0025] Furthermore, the background reconstruction network includes:

[0026] The feature extraction module is used to extract features from the input image;

[0027] The feature detection module is used to detect normal and abnormal features based on the features extracted by the feature extraction module;

[0028] The feature reconstruction module is used to edit the abnormal features detected by the feature detection module to obtain reconstructed features;

[0029] The feature decoding module is used to generate a reconstructed background image based on the normal features detected by the feature detection module and the reconstructed features of the feature reconstruction module.

[0030] Furthermore, the feature detection module performs detection in the following manner:

[0031] S321. Based on the cluster centers in the cluster center set, obtain the residual between each feature and each cluster center, and then based on the smoothing factor of each cluster center, obtain the feature distance between each feature and each cluster center; assign each feature to the cluster center with the smallest feature distance to obtain each current cluster;

[0032] S322. Calculate the Euclidean distance between each feature based on the feature position of each feature to obtain the neighborhood of each feature; determine whether each feature in the neighborhood of each feature belongs to the same cluster. If so, the feature is the center feature; otherwise, the feature is the boundary feature. The neighborhood of a feature includes each feature that is the nearest number of pre-preset neighboring features to the feature in terms of Euclidean distance.

[0033] S323. Update the cluster center set based on each current cluster, and update the clustering parameters with the goal of minimizing the clustering loss function. Repeat steps S321-S323 iteratively until the termination condition is met. Take each current cluster at this point as the final clustering result. The clustering loss function is constructed based on each center feature, each boundary feature, each feature's neighborhood, each feature's feature distance, and the clustering hyperparameters. The clustering parameters include the smoothing factor of each cluster center, the feature position of each feature, and the clustering hyperparameters.

[0034] Based on the final clustering results, the central boundaries of each cluster are obtained, and then normal and abnormal features are obtained.

[0035] Furthermore, the clustering loss function L clu Represented as:

[0036] L clu =ηL kl +L c +L b

[0037] In the formula, η represents the clustering hyperparameter, L c L b Let L represent the center loss function and the boundary loss function, respectively. kl This represents the KL divergence.

[0038] Furthermore, the central loss function L c Represented as:

[0039]

[0040] In the formula, f j F represents the set of central features c Feature j in Represents the neighborhood of feature j Feature m in j , Representing feature j and feature m j The similarity.

[0041] Furthermore, the boundary loss function L b Represented as:

[0042]

[0043] In the formula, f k F represents the set of boundary features b Feature k in The neighborhood E of feature k k Middle feature m k , Representing feature k and feature m kThe similarity, where θ represents the equilibrium parameter.

[0044] Furthermore, the KL divergence L kl Represented as:

[0045]

[0046] In the formula, N represents the total number of features, K represents the total number of cluster centers, and c k Let k represent the cluster center. The distance score between feature i and cluster center k represents the distance score between i and k. Let i represent the target score of feature i and cluster center k.

[0047] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0048] This invention provides a method for detecting low-contrast Mura defects in OLED screens. The method involves acquiring an actual image of the OLED screen and converting it to grayscale to obtain an initial actual grayscale image. Image enhancement is then performed to obtain a final actual grayscale image. This image is then input into a trained background reconstruction network to obtain a reconstructed background image. Based on the reconstructed background image and the actual grayscale image, a defect image of the OLED screen is obtained. Image enhancement during detection improves the contrast of the Mura defect, thereby increasing detection accuracy and solving the problem of poor Mura defect detection accuracy caused by low contrast in OLED screens. Furthermore, the method improves the sufficiency of model training by constructing defect samples. A Gaussian defect mask is constructed to better simulate the characteristics of Mura defects, making the defect image samples more reliable. An illumination mask is introduced to construct a sample image set, enabling the model to distinguish between illumination effects and Mura defects during detection. Additionally, the proposed feature detection module makes feature separation more obvious and abnormal features more prominent, improving the accuracy of defect identification.

[0049] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0050] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0051] Figure 1This is a flowchart illustrating the low-contrast Mura defect detection method for OLED screens provided in an embodiment of the present invention.

[0052] Figure 2 A schematic diagram of image enhancement provided in an embodiment of the present invention;

[0053] Figure 3 A schematic diagram of the background reconstruction adversarial network provided in an embodiment of the present invention;

[0054] Figure 4 This is a schematic diagram illustrating the effect of the feature detection module provided in an embodiment of the present invention;

[0055] Figure 5 This is a schematic diagram illustrating the simulation effect provided for an embodiment of the present invention. Detailed Implementation

[0056] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0057] A specific embodiment of the present invention discloses a method for detecting low-contrast Mura defects in OLED screens, such as... Figure 1 As shown, it includes the following steps:

[0058] S1. Obtain the actual image of the OLED screen and convert it to grayscale to obtain the initial actual grayscale image.

[0059] During implementation, the pixel value of each pixel in the initial actual grayscale image is represented as follows:

[0060] Gray(i,j)=0.299·R(i,j)+0.578·G(i,j)+0.114·B(i,j)

[0061] In the formula, Gray(i,j) represents the pixel value in the i-th row and j-th column of the initial actual grayscale image, R(i,j) represents the pixel value in the i-th row and j-th column of the red primary color of the actual image, G(i,j) represents the pixel value in the i-th row and j-th column of the green primary color of the actual image, and B(i,j) represents the pixel value in the i-th row and j-th column of the blue primary color of the actual image.

[0062] Understandably, since the acquired OLED screen image is a color image, the subsequent calculation and network training of the three-channel image data are more complex. Therefore, this embodiment performs grayscale processing on the image before use: each pixel of the color image is composed of three integer primary colors in the range [0, 255], [R, G, B], which respectively represent the proportion of red, green and blue in the color of that pixel. A grayscale image is an image in which each pixel only has a grayscale value, so it has only one channel. The grayscale value is an integer distributed in [0, 255]. Considering that the Mura defect itself has low contrast, this embodiment performs grayscale processing by weighted averaging to obtain a grayscale image, which has a better effect.

[0063] S2. Enhance the initial actual grayscale image to obtain the actual grayscale image.

[0064] During implementation, the initial actual grayscale image is enhanced using the following methods:

[0065] S21. Let the image r0 = I in the initial iteration. d ′, and set the maximum window size and the initial window size; where, I d ′ represents the initial actual grayscale image.

[0066] S22. Determine the image r in the current iteration. g-1 We can find all the local extrema in the set of local maxima and local minima; where g represents the number of iterations.

[0067] S23, Transfer image r g-1 The following steps are performed sequentially on the pixels to obtain image h. g-1 :

[0068] S231. Set the current window as the initial window;

[0069] S232. If the current window is less than or equal to the largest window, then

[0070] Centered on the current pixel, if the number of local minima in the current window is equal to the number of local maxima, then the average pixel value of all pixels in the current window is taken as the pixel value of the current pixel; otherwise, after expanding the window, return to step S231 and execute again.

[0071] If the current window is larger than the maximum window, then the average pixel value of all pixels within the current window will be used as the pixel value of the current pixel.

[0072] S24, Based on image r g-1 and image h g-1 Obtain the image IMF g-1 And let image r g =h g-1Then return to step S22 for the next iteration, until the number of iterations reaches the set iteration threshold, and then convert the image at this point to IMF. g-1 As the actual grayscale image after image enhancement.

[0073] Specifically, the IMF g-1 =r g-1 -h g-1 .

[0074] Specifically, the size of the maximum window is U×U, where U is represented as:

[0075]

[0076] In the formula, N max N min Let represent the number of all local maxima and all local minima in the actual grayscale image, respectively. Represents a local maximum point i max Distance to the nearest local maximum point Represents a local minimum point i min The distance between the nearest local minimum point and the nearest local minimum point.

[0077] Specifically, expanding the window means increasing the current window size by 2.

[0078] Specifically, the iteration threshold is set to 4, and the initial window size is set to 3×3.

[0079] It is understandable that the above enhancements can effectively enhance the contrast of Mura defects, such as... Figure 2 As shown.

[0080] S3. Input the actual grayscale image into the trained background reconstruction network to obtain the reconstructed background image.

[0081] During implementation, the trained background reconstruction network is obtained in the following way:

[0082] S31. Obtain a defect-free image of the OLED screen, convert the defect-free image to grayscale to obtain a grayscale image, generate a defect-free grayscale image and a defect image based on the grayscale image, and then construct a sample image set based on the defect-free grayscale image and the corresponding defect image.

[0083] S32. Construct a background reconstruction adversarial network, which includes a background reconstruction network and a discriminant network; train the background reconstruction adversarial network based on the sample image set to obtain a trained background reconstruction network.

[0084] In specific implementation, step S31, generating the defect-free grayscale image and the defective image based on the grayscale image, includes:

[0085] Construct a lighting mask, and obtain a defect-free grayscale image based on the lighting mask and the grayscale image;

[0086] A Gaussian defect mask is constructed, and a simulated defect image is obtained based on the Gaussian defect mask and a defect-free grayscale image; the simulated defect image and the corresponding Gaussian defect mask constitute a defect image.

[0087] The illumination mask, Gaussian defect mask, grayscale image, defect-free grayscale image, and defect image are all the same size.

[0088] Specifically, multiple defect-free grayscale images and defective images are generated based on a single grayscale image. During the generation of each defect-free grayscale image and defective image, the corresponding illumination mask and Gaussian defect mask are reconstructed.

[0089] It should be noted that there is no publicly available image library for OLED screens. Therefore, in Mura defect detection, images are obtained from high-resolution industrial cameras capturing high-definition photos of screens with Mura defects. Industrial cameras are highly susceptible to lighting effects during photography, and uneven light distribution can easily be mistaken for defects. Therefore, when constructing the sample image set, the influence of lighting direction is considered, and a lighting mask is constructed to avoid misidentifying light spots as Mura defects in subsequent detection, thus further enhancing the reliability and accuracy of the detection. Furthermore, in actual manufacturing processes, very few OLED screens exhibit Mura defects, insufficient to support model learning; therefore, it is necessary to construct defect images.

[0090] Specifically, the lighting mask is constructed in the following way:

[0091] Set the darkest pixel value, the brightest pixel value, and the lighting direction; wherein the lighting direction is set according to each side of the lighting mask;

[0092] The pixel values ​​of each row of pixels in the illumination mask are obtained sequentially based on the darkest and brightest pixel values ​​along the illumination direction, thus obtaining the illumination mask.

[0093] More specifically, the pixel values ​​of each row of pixels are represented as follows:

[0094]

[0095] In the formula, L n L represents the pixel value of the nth row of pixels along the lighting direction of the light mask. min L min These represent the darkest and brightest pixel values ​​under illumination, respectively, and N represents the total number of rows of pixels along the illumination direction of the illumination mask.

[0096] More specifically, the darkest and brightest pixel values ​​are randomly selected and set within the preset darkest and brightest ranges each time a lighting mask is built; the lighting direction is randomly set within each side of the lighting mask each time a lighting mask is built.

[0097] Preferably, the range of the darkest area is set to (0.9, 1], and the range of the brightest area is set to [1, 1.1].

[0098] Specifically, the Gaussian defect mask is constructed in the following way:

[0099] An initial Gaussian defect mask of the same size is generated based on the grayscale image, wherein the pixel value of each pixel of the initial Gaussian defect mask is 0;

[0100] The defect region is determined in the initial Gaussian defect mask based on randomly set defect center point and location parameters; wherein, the location parameters include row variance and column variance;

[0101] Based on the defect center point and position parameters, the pixel values ​​of each pixel point are generated in the defect region of the initial Gaussian defect mask to obtain the Gaussian defect mask.

[0102] More specifically, the pixel values ​​of the pixels in the defect region of the Gaussian defect mask are represented as follows:

[0103]

[0104] In the formula, I m,wh Let A represent the initial pixel value of the pixel in the w-th row and h-th column of the Gaussian defect mask, where A represents the amplitude, w0 and h0 represent the row and column of the defect center point in the Gaussian defect mask, respectively, and σ represents the initial pixel value of the pixel in the w-th row and h-th column of the Gaussian defect mask. W σ H These represent the row variance and column variance, respectively.

[0105] More specifically, the amplitude, row variance, and column variance are randomly selected and set within the preset amplitude range, row variance, and column variance range each time a Gaussian defect mask is constructed; the defect center point position is randomly selected in the initial Gaussian defect mask each time a Gaussian defect mask is constructed.

[0106] Preferably, the amplitude range is (0, 255), and the row variance and column variance range are both (1, 10).

[0107] Preferably, there may be multiple defect center points, that is, multiple defects are constructed in the Gaussian defect mask.

[0108] It is understandable that the pixel values ​​of Mura defects often exhibit a characteristic of decaying from the center of the defect to the surrounding area. In this embodiment, a two-dimensional Gaussian distribution is used for simulation, which can more accurately simulate Mura defects.

[0109] Specifically, the defect-free grayscale image I o Represented as:

[0110] I o =L light ⊙I′ o

[0111] In the formula, L light Indicates a light mask, I o ′ represents a grayscale image, and ⊙ represents matrix dot product; where matrix dot product is the multiplication of elements at corresponding positions in a matrix.

[0112] The simulated defect image I d Represented as:

[0113] I d =I o +vI m

[0114] In the formula, v represents the adjustment coefficient, and I m This represents a Gaussian defect mask.

[0115] More specifically, the adjustment coefficient v ranges from (-1, 1).

[0116] In specific implementation, in step S32, such as Figure 3 As shown, the background reconstruction network includes:

[0117] The feature extraction module is used to extract features from the input image;

[0118] The feature detection module is used to detect normal and abnormal features based on the features extracted by the feature extraction module;

[0119] The feature reconstruction module is used to edit the abnormal features detected by the feature detection module to obtain reconstructed features;

[0120] The feature decoding module is used to generate a reconstructed background image based on the normal features detected by the feature detection module and the reconstructed features of the feature reconstruction module.

[0121] Specifically, the feature extraction module includes five sequentially connected convolutional layers. The convolutional kernels of the first to fourth layers are all 3×3. The stride and number of channels of the first convolutional layer are 1 and 16, respectively. The stride and number of channels of the second to fourth layers increase in multiples of 2. The convolutional kernel of the fifth layer is 1×1, and the stride is 1.

[0122] Specifically, the feature decoding module includes five sequentially connected deconvolutional layers corresponding to the feature extraction module, and the layers of the feature extraction module and the feature decoding module are connected in a jump-like manner.

[0123] Specifically, such as Figure 4 As shown, the feature detection module performs detection in the following manner:

[0124] S321. Based on the cluster centers in the cluster center set, obtain the residual between each feature and each cluster center, and then based on the smoothing factor of each cluster center, obtain the feature distance between each feature and each cluster center; assign each feature to the cluster center with the smallest feature distance to obtain each current cluster.

[0125] More specifically, the cluster centers in the cluster center set are selected features, which are randomly selected and set initially, and updated in subsequent steps.

[0126] More specifically, the residual between the features and the cluster centers is expressed as:

[0127]

[0128] In the formula, f represents the residual between feature i and cluster center k. i Representing features i, c k Let k represent the cluster center.

[0129] More specifically, the feature distance between a feature and a cluster center is expressed as:

[0130]

[0131] In the formula, The residuals represent feature i and cluster center k. The smoothing factor represents the cluster center k.

[0132] It should be noted that the smoothing factor of each cluster center is randomly assigned at the beginning, for example, all of them are assigned a value of 1, and then updated in subsequent iterations according to the clustering loss function.

[0133] Preferably, the number of cluster centers is 16.

[0134] S322. Calculate the Euclidean distance between each feature based on the feature position of each feature to obtain the neighborhood of each feature; determine whether each feature in the neighborhood of each feature belongs to the same cluster. If so, the feature is the central feature; otherwise, the feature is the boundary feature. The neighborhood of a feature includes each feature that is the nearest number of pre-preset neighboring features to the feature in terms of Euclidean distance.

[0135] More specifically, the number of neighboring features is set to 10.

[0136] It should be noted that each feature has a corresponding position in the feature space, and clustering of each feature is achieved by changing the feature position of each feature through subsequent iterations.

[0137] It is understood that in this embodiment, the neighborhood of each feature is obtained based on the Euclidean distance, and the features in the same cluster are divided into two parts: central features and boundary features. For the central features, it is desirable that they be as compact as possible, while for the boundary features, it is desirable that they be as far away from dissimilar neighboring features as possible and as close as similar neighboring features. Thus, by changing the position of the features in subsequent iterations, the distribution of normal background features is made very compact, which helps in the detection of abnormal features in the future.

[0138] S323. Update the cluster center set based on each current cluster, and update the clustering parameters with the goal of minimizing the clustering loss function. Repeat steps S321-S323 for iteration until the termination condition is met. Take each current cluster at this time as the final clustering result. The clustering loss function is constructed based on each center feature, each boundary feature, each feature's neighborhood, each feature's feature distance, and the clustering hyperparameters. The clustering parameters include the smoothing factor of each cluster center, the feature position of each feature, and the clustering hyperparameters.

[0139] More specifically, the cluster center set is updated by taking the average of the feature distances of each feature in each current cluster as the new cluster centers.

[0140] More specifically, the termination condition is that the clustering loss function converges and the cluster center set does not change within a set number of consecutive iterations. Preferably, the number of consecutive iterations is set to 20.

[0141] More specifically, the clustering loss function L clu Represented as:

[0142] L clu =ηL kl +L c +L b

[0143] In the formula, η represents the clustering hyperparameter, L c L b Let L represent the center loss function and the boundary loss function, respectively. kl This represents the KL divergence.

[0144] Furthermore,

[0145] Central loss function L c Represented as:

[0146]

[0147] In the formula, f j F represents the set of central features c Feature j in Represents the neighborhood of feature j Feature m inj , Representing feature j and feature m j The similarity.

[0148] Boundary loss function L b Represented as:

[0149]

[0150] In the formula, f k F represents the set of boundary features b Feature k in The neighborhood E of feature k k Middle feature m k , Representing feature k and feature m k The similarity is represented by θ, which is a balance parameter. This balance parameter is used to balance the Euclidean distances between features of the same and different clusters; a larger value indicates that the distance between features of the same cluster is greater and more dominant.

[0151] Preferably, the balance parameter θ is set to 10.

[0152] Furthermore, feature j and feature m j similarity Represented as:

[0153]

[0154] in,

[0155]

[0156] In the formula, Representing feature j and feature m j Normalized weights; Representing feature j and feature m j The Euclidean distance is given by softmax(), where softmax() represents the Softmax function and distance() represents the Euclidean distance function.

[0157] It is understandable that in this embodiment, considering that the neighboring regions of features may have area differences due to sparsity, in order to fairly handle the features of dense regions within a cluster and sparse regions at the cluster boundary, the Euclidean distance is adjusted by normalization using the Softmax function so that it is distributed in (0,1), making it more sensitive to local characteristics. The role of the softmax function is to amplify the weights corresponding to smaller distances and compress the weights corresponding to larger distances, thereby enhancing local information.

[0158] It should be noted that in this embodiment, the calculation methods for each parameter are the same, with only the features being transformed.

[0159] Furthermore, the KL divergence L kl Represented as:

[0160]

[0161] In the formula, N represents the total number of features, K represents the total number of cluster centers, and c k Let k represent the cluster center. The distance score between feature i and cluster center k represents the distance score between i and k. Let represent the target score between feature i and cluster center k. The distance score measures how normal each feature is to a given cluster center; normal features are closer to the center.

[0162] Furthermore, the target score of feature i and cluster center k Represented as:

[0163]

[0164] in,

[0165]

[0166] In the formula, The smoothing factor representing the cluster center k. Let |i| represent the residual between feature i and cluster center k, and |||| denotes the modulo operation.

[0167] S324. Based on the final clustering results, the central boundaries of each cluster are obtained, and then the normal and abnormal features are obtained.

[0168] Specifically, the central boundaries of each cluster center are represented as follows:

[0169]

[0170] In the formula, T k N represents the central boundary of cluster k. k This represents the number of features in cluster k. In cluster k, feature i represents the feature i. k The characteristic distance σ to the cluster center k of cluster k k It represents the standard deviation of the feature distances between all features in cluster k and the cluster center k of cluster k.

[0171] More specifically, normal and abnormal features are obtained in the following ways:

[0172] If the feature is more than the cluster center of its cluster than the feature distance of its cluster, then the feature is an anomalous feature; otherwise, it is a normal feature.

[0173] Specifically, the reconstructed features of the abnormal features in the feature reconstruction module are represented as follows:

[0174]

[0175] In the formula, Indicates abnormal feature i E The reconstruction features, Representing normal feature j E B represents the set of anomalous features. Indicates abnormal feature i E and normal features j E The normalized fraction.

[0176] More specifically, anomalous feature i E and normal features j E normalized score Represented as:

[0177]

[0178] In the formula, Indicates abnormal feature i E , <> indicates inner product calculation.

[0179] In specific implementation, in step S32, the discrimination network is used to make discrimination based on the input image to obtain a discrimination score, which is the probability that each pixel in the image comes from a defect-free grayscale image; wherein, the input image includes defect-free grayscale images in the sample image set, and reconstructed background images generated by the background reconstruction network from the defect-free grayscale images and the corresponding simulated defect images.

[0180] Specifically, the discriminant network consists of four convolutional layers and four transposed convolutional layers connected in sequence. The convolutional kernels of both the convolutional and transposed convolutional layers are 3×3, and the stride is 2. The number of filter kernels in the first convolutional layer is 32, and the number of filter kernels in the second and third layers increases by a factor of 2. The number of filter kernels in the first transposed convolutional layer is 256, and the number of filter kernels in the second and third layers decreases by a factor of 2.

[0181] In specific implementation, step S32 involves training the background reconstruction adversarial network in the following manner:

[0182] Sa1. Based on the discriminant network model and defect-free grayscale images in the sample image dataset, the feature extraction module and feature decoding module in the background reconstruction network model are trained. The goal is to minimize the first loss function and the second loss function, and the trained feature extraction module, feature decoding module and discriminant network model are obtained.

[0183] Sa2. Update the background reconstruction adversarial network model based on the trained feature extraction module, feature decoding module and discriminant network model; train the updated background reconstruction adversarial network model based on defect-free grayscale images in the sample image dataset, with the goal of minimizing the first loss function, the second loss function and the clustering loss function, to obtain the trained background reconstruction adversarial network model.

[0184] Sa3. The background reconstruction adversarial network model is trained based on the defect-free grayscale images and defective images in the sample image dataset. The goal is to minimize the third loss function to obtain the trained background reconstruction adversarial network model.

[0185] More specifically, the first loss function L adv Represented as:

[0186]

[0187] In the formula, W and H represent the total number of rows and columns of the image, respectively, E represents the expected value, and D represents the desired value. wh (I o ) represents the output value in row w and column h when the input to the discrimination network is a defect-free grayscale image Io; D wh (I ob ) indicates that the input to the discrimination network is a defect-free grayscale image I. o Reconstructed background image I generated by the background reconstruction network ob The output value in row w and column h; D wh (I db ) indicates that the input to the discrimination network is a defect-free grayscale image I. o Corresponding simulated defect image I d Reconstructed background image I generated by the background reconstruction network db The output value in row w and column h.

[0188] More specifically, the second loss function L rec_o Represented as:

[0189] L rec_o =Ε[||I o -I ob ||2]

[0190] In the formula, |||2 represents the L2 norm.

[0191] More specifically, the third loss function L is expressed as:

[0192] L=λ1(L rec_o +L rec_d +L rec_m )+λ2L clu +λ3L adv

[0193] in,

[0194] L rec_d =Ε[||I o -I db ||2]

[0195] L rec_m =Ε[β||I m ⊙(I o -I db )||2]

[0196] In the formula, λ1, λ2, and λ3 are the first, second, and third loss weights, respectively, and I m β represents the magnification factor, which is used to simulate the Gaussian defect mask corresponding to the defect image.

[0197] Furthermore, the amplification factor β is expressed as:

[0198]

[0199] In the formula, sum(I m ) represents Gaussian defect mask I m The sum of the pixel values ​​of all pixels in the image.

[0200] Specifically, each training process terminates when the corresponding loss function converges.

[0201] Preferably, λ1 = 1000, λ2 = 0.1, and λ3 = 1.

[0202] S4. Based on the reconstructed background image and the actual grayscale image, obtain the defect image of the OLED screen.

[0203] In implementation, in step S4, the defect image is represented as follows:

[0204]

[0205] In the formula, I bin (w,h) represents the pixel value of the pixel in the w-th row and h-th column of the defect image of the OLED screen, and T1 and T2 represent the first and second residual thresholds, respectively. res (w,h) represents the pixel value of the pixel in the w-th row and h-th column of the residual image between the reconstructed background image and the actual grayscale image.

[0206] Specifically, the residual image I of the reconstructed background image and the actual grayscale image. res Represented as:

[0207] I res =|I′ d -I′ db |

[0208] In the formula, I d ′、I d ′ b These represent the actual grayscale image and the reconstructed background image, respectively.

[0209] Specifically, the first and second residual thresholds T1 and T2 are expressed as follows:

[0210]

[0211] In the formula, μ and σ represent the mean and standard deviation of the residual image, respectively, and ε represents the control coefficient for segmentation sensitivity.

[0212] Preferably, the control coefficient for ε segmentation sensitivity is set to 3.5.

[0213] To verify the effectiveness of this method, simulation verification was performed, such as... Figure 5 The image shown is a simulation diagram of this embodiment, consisting of the actual grayscale image, the image after image enhancement, the reconstructed background image, and the defect image. It can be seen that this method can perform better in detecting low-contrast Mura defects.

[0214] Compared with existing technologies, this embodiment provides a method for detecting low-contrast Mura defects in OLED screens. It acquires an actual image of the OLED screen and converts it to grayscale to obtain an initial actual grayscale image. Image enhancement is then performed to obtain a final actual grayscale image. This actual grayscale image is then input into a trained background reconstruction network to obtain a reconstructed background image. Based on the reconstructed background image and the actual grayscale image, a defect image of the OLED screen is obtained. Image enhancement during image detection improves the contrast of the Mura defect, thereby improving detection accuracy and solving the problem of poor Mura defect detection accuracy caused by low contrast in OLED screens. Furthermore, the method improves the sufficiency of model training by constructing defect samples; it also better simulates the characteristics of Mura defects by constructing a Gaussian defect mask, making the defect image samples more reliable; and it introduces an illumination mask to construct a sample image set, enabling the model to distinguish between illumination effects and Mura defects during detection. In addition, the proposed feature detection module makes feature separation more obvious and abnormal features more prominent, improving the accuracy of defect identification.

[0215] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0216] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting low-contrast Mura defects in OLED screens, characterized in that, Includes the following steps: Acquire the actual image of the OLED screen and convert it to grayscale to obtain the initial actual grayscale image; The initial actual grayscale image is enhanced to obtain the actual grayscale image; The actual grayscale image is input into the trained background reconstruction network to obtain the reconstructed background image; Based on the reconstructed background image and the actual grayscale image, defect images of the OLED screen are obtained.

2. The method for detecting low-contrast Mura defects in OLED screens according to claim 1, characterized in that, The initial actual grayscale image is enhanced using the following method: S21. Let the image r0 = I in the initial iteration. d ′, and set the maximum window size and the initial window size; where, I d ′ represents the initial actual grayscale image. S22. Determine the image r in the current iteration. g-1 We can find all the local extrema in the set of local maxima and local minima; where g represents the number of iterations. S23, Transfer image r g-1 The following steps are performed sequentially on the pixels to obtain image h. g-1 : S231. Set the current window as the initial window; S232. If the current window is less than or equal to the largest window, then Centered on the current pixel, if the number of local minima in the current window is equal to the number of local maxima, then the average pixel value of all pixels in the current window is taken as the pixel value of the current pixel; otherwise, after expanding the window, return to step S231 and execute again. If the current window is larger than the maximum window, then the average pixel value of all pixels within the current window will be used as the pixel value of the current pixel. S24, Based on image r g-1 and image h g-1 Obtain the image IMF g-1 And let image r g =h g-1 Then return to step S22 for the next iteration, until the number of iterations reaches the set iteration threshold, and then convert the image at this point to IMF. g-1 As the actual grayscale image after image enhancement.

3. The method for detecting low-contrast Mura defects in OLED screens according to claim 2, characterized in that, The image IMF g-1 Represented as: IMF g-1 =r g-1 -h g-1 。 4. The method for detecting low-contrast Mura defects in OLED screens according to claim 3, characterized in that, The maximum window size is U×U, where U represents: In the formula, N max N min Let represent the number of all local maxima and all local minima in the actual grayscale image, respectively. Represents a local maximum point i max Distance to the nearest local maximum point Represents the local minimum point i min The distance between the nearest local minimum point and the nearest local minimum point.

5. The method for detecting low-contrast Mura defects in OLED screens according to claim 1, characterized in that, The background reconstruction network includes: The feature extraction module is used to extract features from the input image; The feature detection module is used to detect normal and abnormal features based on the features extracted by the feature extraction module; The feature reconstruction module is used to edit the abnormal features detected by the feature detection module to obtain reconstructed features; The feature decoding module is used to generate a reconstructed background image based on the normal features detected by the feature detection module and the reconstructed features of the feature reconstruction module.

6. The method for detecting low-contrast Mura defects in OLED screens according to claim 5, characterized in that, The feature detection module performs detection in the following ways: S321. Based on the cluster centers in the cluster center set, obtain the residual between each feature and each cluster center, and then based on the smoothing factor of each cluster center, obtain the feature distance between each feature and each cluster center; assign each feature to the cluster center with the smallest feature distance to obtain each current cluster; S322. Calculate the Euclidean distance between each feature based on the feature position of each feature to obtain the neighborhood of each feature; determine whether each feature in the neighborhood of each feature belongs to the same cluster. If so, the feature is the center feature; otherwise, the feature is the boundary feature. The neighborhood of a feature includes each feature that is the nearest number of pre-preset neighboring features to the feature in terms of Euclidean distance. S323. Update the cluster center set based on each current cluster, and update the clustering parameters with the goal of minimizing the clustering loss function. Repeat steps S321-S323 iteratively until the termination condition is met. Take each current cluster at this point as the final clustering result. The clustering loss function is constructed based on each center feature, each boundary feature, each feature's neighborhood, each feature's feature distance, and the clustering hyperparameters. The clustering parameters include the smoothing factor of each cluster center, the feature position of each feature, and the clustering hyperparameters. Based on the final clustering results, the central boundaries of each cluster are obtained, and then normal and abnormal features are obtained.

7. The method for detecting low-contrast Mura defects in OLED screens according to claim 7, characterized in that, The clustering loss function L clu Represented as: L clu =ηL kl +L c +L b In the formula, η represents the clustering hyperparameter, L c L b Let L represent the center loss function and the boundary loss function, respectively. kl This represents the KL divergence.

8. The method for detecting low-contrast Mura defects in OLED screens according to claim 7, characterized in that, The central loss function L c Represented as: In the formula, f j F represents the set of central features c Feature j in Represents the neighborhood of feature j Feature m in j , Representing feature j and feature m j The similarity.

9. The method for detecting low-contrast Mura defects in OLED screens according to claim 8, characterized in that, The boundary loss function L b Represented as: In the formula, f k F represents the set of boundary features b Feature k in The neighborhood E of feature k k Middle feature m k , Representing feature k and feature m k The similarity, where θ represents the equilibrium parameter.

10. The method for detecting low-contrast Mura defects in OLED screens according to claim 9, characterized in that, KL divergence L kl Represented as: In the formula, N represents the total number of features, K represents the total number of cluster centers, and c k Let k represent the cluster center. The distance score between feature i and cluster center k represents the distance score between i and cluster center k. Let i represent the target score of feature i and cluster center k.

Images