An automatic detection and classification method and system for micro-defects in OLED displays

Abstract

This invention discloses an automatic detection and classification method and system for micro-defects in OLED displays, relating to the field of display panel quality inspection technology. The method includes acquiring OLED display images and preprocessing them; modeling pixel responses based on the preprocessed images and generating an ideal background image; performing differential processing and multi-scale residual feature enhancement to obtain an enhanced residual image; segmenting the enhanced residual image to extract a set of candidate defect regions and extracting global features; based on the set of candidate defect regions, performing frequency domain enhancement on the global features using a local frequency domain prior-guided complex filter to obtain enhanced global context features; based on the enhanced global context features, using a meta-learning mechanism to obtain new screen type adaptation parameters for defect category probability prediction; and performing uncertainty estimation based on the defect category probability prediction results. This invention enables more accurate, flexible, and reliable OLED display defect detection and classification.

Description

Technical Field

[0001] This invention relates to the field of display panel quality inspection technology, and in particular to an automatic detection and classification method and system for micro-defects in OLED displays. Background Technology

[0002] As a core component of high-end mobile phones and televisions, OLED screens are highly susceptible to defects during production, even minor ones (such as bright spots, dark spots, and uneven color patches), which directly impact product quality. Currently, the industry primarily relies on machine vision and artificial intelligence technologies to automatically detect these defects, replacing traditional manual inspection. Mainstream methods can be divided into two categories: one uses image processing algorithms (such as filtering and segmentation) to identify defect areas; the other uses deep learning models (such as CNNs) to directly determine the defect type. To improve adaptability to different screen models, some research has also begun to utilize transfer learning and data augmentation methods.

[0003] However, these existing technologies still have many problems in practical applications, especially in the comprehensiveness of defect feature extraction, adaptability to different screens, and reliability assessment of detection results. First, most methods only focus on the spatial features of the image, without fully considering the frequency distribution of defects, leading to inaccurate identification of defects with specific textures (such as fine-line unevenness). Second, ordinary deep learning models have a limited "seeing" range, making it difficult to notice the overall periodic defect patterns of the screen. Expanding the detection range increases computational burden and loses details. More importantly, when the production line changes to a new screen model, it usually requires re-labeling a large amount of data to train the model, which is very costly. Moreover, the model often cannot provide a "sure" assessment of the likelihood of identifying some difficult-to-judge defects, making it difficult to effectively focus the workload of manual re-inspection. Therefore, existing OLED defect detection technologies have significant shortcomings in terms of insufficient feature extraction, weak adaptability to different screens, and difficulty in assessing the reliability of detection results. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an automatic detection and classification method and system for micro-defects in OLED displays, which solves the problems of insufficient feature extraction, weak adaptability to different screens, and difficulty in evaluating the reliability of detection results.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides an automatic detection and classification method for micro-defects in OLED displays, comprising,

[0008] Acquire OLED display images and perform preprocessing. Based on the preprocessed images, perform pixel response modeling and generate an ideal background image. Perform differential processing and multi-scale residual feature enhancement to obtain an enhanced residual image.

[0009] The enhanced residual image is segmented to extract a set of candidate defect regions, and global features are extracted. Based on the set of candidate defect regions, a complex filter guided by local frequency domain prior is used to enhance the global features in the frequency domain to obtain enhanced global context features.

[0010] Based on enhanced global context features, a meta-learning mechanism is used to obtain new screen type adaptation parameters for defect category probability prediction. Uncertainty estimation is performed based on the defect category probability prediction results, and active labeling and updating are triggered by predicting uncertainty.

[0011] As a preferred embodiment of the automatic detection and classification method for micro-defects in OLED displays described in this invention, the step of acquiring OLED display images and performing preprocessing refers to controlling the OLED screen under test to sequentially display a set of standard gray levels covering the working dynamic range in a light-shielding environment. After the display state of each standard gray level stabilizes, original display images are continuously acquired, and the original display images under the same gray level are preprocessed to obtain an image sequence.

[0012] As a preferred embodiment of the automatic detection and classification method for micro-defects in OLED displays described in this invention, the step of performing pixel response modeling based on preprocessed images and generating an ideal background image for differential processing and multi-scale residual feature enhancement to obtain an enhanced residual image refers to receiving an image sequence as input, extracting the brightness response value under different grayscale conditions for each pixel in the display area, establishing a pixel-level grayscale-brightness response model, calculating the theoretical brightness of each pixel in the entire screen, generating an ideal background image, subtracting the ideal background image from the linearized image under the current standard grayscale pixel by pixel to obtain an initial residual image, constructing a Laplacian pyramid on the initial residual image for multi-scale decomposition, and performing weighted fusion to form the final enhanced residual feature map, and constructing a screen type background prior vector based on the mean and standard deviation of the pixel-level grayscale-brightness response model parameters combined with the fixed properties of the OLED screen.

[0013] As a preferred embodiment of the automatic detection and classification method for micro-defects in OLED displays described in this invention, the step of segmenting the enhanced residual image to extract a set of candidate defect regions refers to performing binarization processing on the final enhanced residual feature map to obtain a binary image, performing opening operations, performing connected component analysis based on the processed binary image, identifying all independent candidate defect regions, and extracting a set of geometric features for each candidate defect region, including bounding box coordinates, area, perimeter, major axis and minor axis, and shape factor, to form a set of candidate defect regions.

[0014] As a preferred embodiment of the automatic detection and classification method for micro-defects in OLED displays described in this invention, the following steps are described: global features are extracted, and based on a set of candidate defect regions, a local frequency domain prior-guided complex filter is used to enhance the global features in the frequency domain to obtain enhanced global context features. The coordinates of the center point of the bounding box are calculated based on the rectangular bounding box corresponding to each candidate defect region in the set of candidate defect regions. With the coordinates of the center point of the bounding box of each candidate defect region as the center, local residual blocks are cropped from the enhanced residual feature map. At the same time, the enhanced residual feature map is used as the global input image and input into a lightweight convolutional backbone network constructed based on FFC.

[0015] The backbone network consists of convolutional blocks and includes batch normalization, ReLU activation function and downsampling operation. In the shallow stage of the backbone network, only local convolutional paths are retained. In the deep stage of the network, the complete FFC module is enabled to fuse local and global information and output a global feature map.

[0016] The bounding boxes of candidate defect regions are mapped to the coordinate system of the global feature map. Global contextual feature regions are extracted using RoIAlign. Simultaneously, local residual blocks are input into a shared lightweight encoder to obtain local feature maps, and response values ​​are calculated to form local frequency domain description vectors. The conditional input supernetwork generates the final complex frequency domain filter. For global context feature regions Each channel undergoes DFT and FFTshift, resulting in the final complex frequency domain filter. Element-wise multiplication, followed by inverse FFT shift and IDFT to obtain the real part, yields the enhanced channel features. These are then concatenated to obtain the final enhanced global context features. .

[0017] As a preferred embodiment of the automatic detection and classification method for micro-defects in OLED displays described in this invention, the step of using a meta-learning mechanism to obtain new screen type adaptation parameters and predict defect category probabilities based on enhanced global context features refers to, for each candidate defect region, using the corresponding final enhanced global context features... Perform global average pooling to obtain the feature vector. Simultaneously, the structured screen background prior vectors are mapped to embedding vectors through a small fully connected network. and will , The combined feature vector is obtained by concatenating the features, which is then input into a multilayer perceptron and followed by a classification layer to form a learnable classification function.

[0018] The classification function is trained using the MAML framework. On the historical OLED screen dataset, each screen type is treated as an independent task, and shared initial parameters are optimized. When the production line switches to a new screen type, labeled samples are collected to form a support set. Starting with the optimal initial parameters obtained after optimization, gradient descent is performed to obtain the adaptation parameters. The adaptation parameters and the comprehensive feature vector of each candidate region are used as input, and forward propagation is performed through the classification function to obtain the class probability vector.

[0019] As a preferred embodiment of the automatic detection and classification method for micro-defects in OLED displays described in this invention, wherein: the uncertainty estimation based on the probability prediction results of defect categories, and the active labeling and updating triggered by the prediction uncertainty, refers to using only the intra-model uncertainty estimation based on MCDropout or the inter-model uncertainty estimation based on multi-model consistency enhancement within any detection cycle based on the probability prediction results of the defect categories of each candidate defect region;

[0020] Based on the mean probability vector obtained by any method, Shannon entropy is calculated as the prediction uncertainty. When the prediction uncertainty is less than or equal to the uncertainty threshold... When the prediction uncertainty exceeds the uncertainty threshold, the automatic classification result is accepted. When candidate defect regions are included in a high-risk sample pool, a redundancy removal screening based on feature similarity is performed on the high-risk sample pool. When the number of candidate defect regions after screening exceeds the preset review budget, the previous ones are selected in descending order of uncertainty. Each candidate defect region enters the manual review queue. For the selected candidate defect regions, a manual review data package is constructed and stored in the screen-specific cache set. After the end of each production day, incremental fine-tuning is performed on the newly added samples in the cache set to obtain new adaptation parameters.

[0021] Secondly, this invention provides an automatic detection and classification system for micro-defects in OLED displays, comprising,

[0022] The image acquisition and preprocessing module is used to acquire OLED images under multi-grayscale conditions, and after registration, averaging and brightness linearization, output a low-noise linearized image sequence.

[0023] The pixel response modeling and enhancement module is used to generate an ideal background image based on a pixel-level grayscale-brightness model, and obtain an enhanced residual feature map through differential and multi-scale residual fusion.

[0024] The candidate defect region extraction module is used to perform adaptive threshold segmentation and morphological processing on the enhanced residual map, extract candidate defect regions, and calculate geometric and statistical features.

[0025] The frequency domain feature fusion and classification module enhances the global context through a frequency domain attention mechanism, enabling the fusion and classification of local and global features.

[0026] The Meta-Learning and Uncertainty Learning module is used to perform cross-screen adaptive classification using a meta-learning mechanism, quantify and predict uncertainty, trigger active labeling, and achieve incremental model updates.

[0027] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the automatic detection and classification method for micro-defects in an OLED display screen as described in the first aspect of the present invention.

[0028] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the automatic detection and classification method for micro-defects in an OLED display screen as described in the first aspect of the present invention.

[0029] The beneficial effects of this invention are as follows: By fusing local frequency domain features with global contextual attention, this invention significantly improves the feature representation ability and classification accuracy of micro-defects (especially texture-type and periodic defects). At the same time, by leveraging the meta-learning cross-screen adaptation framework and the dual-path uncertainty quantification strategy, it achieves rapid deployment and high-confidence decision-making on new screen types, greatly reducing annotation costs and manual review burden, and ultimately forming a high-precision, highly generalizable, and self-evolving OLED micro-defect detection system. Attached Figure Description

[0030] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Fig. 1 This is a flowchart of an automatic detection and classification method for micro-defects in an OLED display screen according to Example 1.

[0032] Fig. 2 This is a structural diagram of an automatic detection and classification system for micro-defects in an OLED display screen, as shown in Example 1.

[0033] Fig. 3 This is a flowchart of the global feature enhancement method based on local frequency domain priors in Example 1.

[0034] Fig. 4 This is a flowchart of uncertainty estimation and active learning in Example 1. Detailed Implementation

[0035] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0036] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0037] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0038] Example 1, referring to Figs. 1-4 This is the first embodiment of the present invention, which provides an automatic detection and classification method for micro-defects in an OLED display screen, including the following steps:

[0039] S1. Acquire OLED display images and perform preprocessing. Based on the preprocessed images, perform pixel response modeling and generate an ideal background image. Perform differential processing and multi-scale residual feature enhancement to obtain an enhanced residual image.

[0040] Specifically, acquiring and preprocessing OLED display images involves using a light shield or conducting OLED screen display tests in a dark room to ensure stable lighting conditions in the acquisition environment, thereby reducing the impact of ambient light on the results, and controlling the OLED screen under test to sequentially display a standard grayscale set. The standard grayscale in, where, The z-th standard grayscale (from darkest to brightest) is the standard grayscale set G, which covers the working dynamic range of the OLED display (e.g., 0 (complete black), 64, 128, 192, 255 (complete white)). After each standard grayscale is displayed, the brightness output of the OLED screen is monitored until the change is lower than the preset stability threshold (set according to the camera signal-to-noise ratio or panel specifications, typically [0.5%, 2%] maximum brightness fluctuation) and maintained for a stable time window (greater than the panel response time and power ripple period, typically [50, 300] ms). After the display state is considered stable, N (typically 10 frames) frames of original display images are continuously acquired from the OLED display area using an industrial camera.

[0041] For the original display images acquired at the same standard grayscale, a phase-correlation-based subpixel registration method is used to compensate for translational displacement, and bilinear interpolation is used to obtain aligned images. To reduce the impact of noise, all aligned images are averaged. The averaging formula is as follows:

[0042]

[0043] in, It is the standard grayscale after averaging. The image below, It is the pixel value at position (x,y) in the original displayed image after averaging, and N is the number of images acquired;

[0044] Considering the non-linear brightness response characteristics of OLED displays, the averaged image... To perform inverse gamma linearization, first calculate the maximum value of the grayscale. Normalize, then press the panel for gamma index. (Based on screen type calibration, a typical value of 2.2) Perform a power transform to obtain the linearized image:

[0045]

[0046] in, It is the standard grayscale after averaging. The linearized image below;

[0047] The final result is a multi-grayscale, low-noise, spatially aligned, and brightness-linearized image sequence. This provides a reliable data foundation for subsequent response modeling and defect visualization.

[0048] By employing a triple collaborative processing approach of steady-state grayscale control, sub-pixel multi-frame alignment averaging, and inverse gamma linearization, the effects of ambient light interference, inter-frame shift, random noise, and OLED nonlinear response are effectively eliminated. This results in a multi-grayscale image sequence with high signal-to-noise ratio, precise spatial alignment, and linearized brightness, significantly improving defect visibility capabilities. In particular, it enhances the robustness of detection and the reliability of quantitative analysis for low-contrast, grayscale-sensitive defects, providing a key prerequisite for high-precision automated optical inspection.

[0049] Furthermore, pixel response modeling is performed based on the preprocessed image, and an ideal background image is generated for differential processing and multi-scale residual feature enhancement to obtain an enhanced residual image. This image sequence is then used for further processing. As input, for each pixel in the OLED display area, the brightness response value of the pixel under different grayscale conditions is extracted to establish a pixel-level grayscale-brightness response model.

[0050] The pixel-level grayscale-luminance response model is used to describe the luminous characteristics of a pixel under normal operating conditions, and is expressed in power-law form as follows:

[0051]

[0052] Where L(x,y,g) is the actual luminance of pixel (x,y) at standard gray level g. It is the maximum luminous gain of the pixel (x,y) (related to the efficiency of organic materials and driving current). It is the gamma exponent of the pixel (x,y), reflecting the degree of non-linearity between brightness and grayscale. It is the dark state offset of the pixel (x,y) (mainly caused by leakage current or dark current). , , Each pixel is independently fitted using a nonlinear least squares method with physical constraints, with typical initial values ​​of 1.0, 2.2, and 0.0.

[0053] The fitted pixel-level grayscale-luminance response model is applied to the current standard grayscale to calculate the theoretical luminance of each pixel in the entire screen under the condition of "no defects and only normal spatial non-uniformity", thus generating an ideal background image. It accurately reflects the uniform light emission state that an OLED panel should have under current display conditions, eliminating abnormal factors such as bright spots, dark spots, and mura, and retaining only the inherent non-uniformity caused by the manufacturing process, thus achieving an ideal background image. Linearized image under current standard grayscale Perform pixel-by-pixel subtraction to obtain the initial residual image:

[0054]

[0055] in, It is the initial residual image. It is a linearized image The pixel value at position (x, y);

[0056] For the initial residual image Construct a Laplacian pyramid for multi-scale decomposition, making the initial Gaussian pyramid equal to the initial residual image. For the current layer Gaussian blur and downsampling are performed to obtain a higher layer. And upsampled to using bilinear interpolation The resolution, denoted as Calculate the Laplace residual plot:

[0057]

[0058] in, is the l-th layer Laplacian residual plot, used to characterize the defect details at different scales, and J is the total number of Laplacian pyramid layers;

[0059] To emphasize defect features at different scales, a weighted fusion method is used based on the importance of each Laplacian pyramid layer to form the final enhanced residual feature map:

[0060]

[0061] in, It is an enhanced residual feature map. It is the weight of the l-th layer of weights, which is set according to the sensitivity of each scale to micro-defects, and the typical value range is [0.4, 1.0].

[0062] To support rapid cross-screen adaptation in the future, a structured screen-type background prior vector is constructed. Specifically, it is constructed as the maximum luminous gain in the pixel-level grayscale-luminance response model. The full-screen mean and standard deviation, gamma index Dark state offset The full-screen mean is obtained, and the fixed attributes of the current OLED screen (from the product specification or MES system) are obtained, including resolution, pixel density and sub-pixel arrangement type (such as RGB Stripe, Pentile, Diamond Pixel, encoded as integer ID). The values ​​are concatenated into a fixed-length vector, which is defined as the screen type background prior vector.

[0063] By accurately separating process non-uniformity from real defects through pixel-level physical response modeling, and combining multi-scale Laplacian residual weighted fusion, the explicit expression of various types of defects such as point, line and area Mura is effectively enhanced. At the same time, a structured screen type prior vector is constructed to explicitly encode the inherent properties of the panel, significantly improving the defect contrast and cross-screen type generalization ability of the residual image, laying the core feature foundation for high-sensitivity and low-false-alarm automated optical inspection.

[0064] S2. Segment the enhanced residual image to extract a set of candidate defect regions and extract global features. Based on the set of candidate defect regions, perform frequency domain enhancement on the global features through a local frequency domain prior-guided complex filter to obtain enhanced global context features.

[0065] Specifically, segmenting the enhanced residual image to extract a set of candidate defect regions refers to using an adaptive thresholding method on the enhanced residual feature map. Binarization is performed, and the local threshold is calculated using the Gaussian weighted mean method.

[0066]

[0067] Where Y(x,y) is the local threshold of pixel (x,y). and These are the average and standard deviation of the pixel values ​​in the neighborhood, respectively, and k is an empirical constant used to adjust the threshold sensitivity;

[0068] Will The pixels marked as foreground and the rest as background are used to obtain a binary image. For binary images Perform an opening operation (erosion followed by dilation) to remove small noise while maintaining the integrity of the target defect area:

[0069]

[0070] in, This is the processed binarized image. It is a structural element, a 3×3 square core. It is a corrosion operation. It is an expansion operation;

[0071] Based on the processed binarized image A breadth-first search algorithm was used to perform connected component analysis, identifying all independent candidate defect regions and extracting a set of detailed geometric features for each candidate defect region, including bounding box coordinates. The area (total number of pixels within the region), perimeter (sum of the lengths of the boundary pixels), major and minor axes (calculated based on the minimum bounding rectangle), and shape factor (used to describe the roundness of the region) form a set of candidate defect regions. ,in, It is the nth candidate defect region.

[0072] By effectively addressing the spatial inhomogeneity of the residual map through local adaptive thresholding, the detection capability of weak defects is improved. Combined with small kernel opening operation, noise is accurately removed without damaging the true geometric structure of defects. Then, multi-dimensional geometric features are extracted through connected component analysis, realizing reliable localization and structured characterization of candidate defects, significantly reducing the false alarm rate, ensuring the input quality of subsequent frequency domain enhancement and classification, and providing a robust regional foundation for high-precision defect detection.

[0073] Furthermore, global features are extracted, and based on the candidate defect region set, a local frequency domain prior-guided complex filter is used to enhance the global features in the frequency domain, resulting in enhanced global contextual features based on the candidate defect region set. The rectangular bounding box corresponding to each candidate defect region. Calculate the coordinates of the center point of the bounding box:

[0074]

[0075] in, These are the coordinates of the center point;

[0076] Coordinates of the center point of the bounding box for each candidate defect region Centered on the final enhanced residual feature map A fixed-size local residual block is clipped, containing the candidate defect region body and neighborhood context information, used to characterize the spatial structural features of point defects, linear defects, and local brightness anomalies:

[0077]

[0078] in, It is a local residual block. It is the half-length of the local clipping block, and is set based on experience. A typical value is 32.

[0079] When the cropped region exceeds the image boundary, zero-padding is used to fill in the pixel values, and the final enhanced residual feature map is then used. The global input image is fed into a lightweight convolutional backbone network built on Fast Fourier Convolution (FFC);

[0080] The backbone network consists of six cascaded convolutional blocks, each containing batch normalization, a ReLU activation function, and a downsampling operation, with a total spatial downsampling factor of [missing value]. In the backbone network, differentiated local / global information processing strategies are adopted at different stages. In the shallow stages of the network (e.g., in the first few convolutional blocks), the global branch ratio parameter is set to... This means disabling the frequency domain processing path and retaining only the local convolutional path to maintain high-resolution spatial details. In deeper layers of the network (e.g., in subsequent convolutional blocks), the global branch ratio parameter is set to... Enable the full FFC module to enhance global context modeling capabilities;

[0081] In the FFC module, the input feature tensor (The output of the previous module) is split into a local stream and a global stream according to the channel dimension. The local stream is processed by a regular spatial convolution (standard 3×3 convolution + BN + ReLU) to obtain the local stream output. The global stream is mapped to the frequency domain using a Fast Fourier Transform (FFT), then subjected to a 1×1 convolution operation in the frequency domain, and finally returned to the spatial domain using an Inverse Fast Fourier Transform (IFFT) to obtain the global stream output. After processing and The data is concatenated along the channel dimension and fused using a 1×1 convolution, finally combined with the input feature tensor. The residuals are added together to form a residual connection, which outputs the feature map of the FFC module. The global feature map is obtained by processing all FFC modules and used for subsequent frequency domain enhancement modeling.

[0082]

[0083] in, This is the global feature map, where H and W are the spatial dimensions of the global feature map, and C is the number of channels;

[0084] The bounding box of the candidate defect region Mapping to global feature map The coordinate system is mapped to floating-point coordinates, preserving sub-pixel precision, for use by RoIAlign.

[0085]

[0086] in, It is the mapped bounding box. It is the total spatial downsampling rate of the backbone network;

[0087] In global feature map Above, based on the mapped bounding box The RoIAlign operation (using bilinear interpolation for feature sampling) is used to extract a fixed-size global context feature region to represent the periodic or quasi-periodic structural information of the candidate defect region in the global context.

[0088]

[0089] in, It is a global context feature region. It is a fixed size extracted;

[0090] Local residual blocks A lightweight convolutional encoder with shared parameters (4 layers of Conv-BN-ReLU, stride=2 per layer, kernel size 3×3) is used to obtain local feature maps. For local feature maps For each channel, the response value is calculated at a preset K (selected via "Zigzag scan order", typically 16) low-frequency two-dimensional discrete cosine transform (DCT-II) frequency points:

[0091]

[0092]

[0093] in, It is the response value. It is a preset low-frequency point. It is the sequence number of the preset frequency point. It is pi. It is the height of the local feature map. It is the width of the local feature map. It is a local feature map The pixel value at position (x, y) and channel c. It is the DCT-II normalization factor. It is a component of the DCT frequency domain coordinates, i.e., the frequency index;

[0094] Finally, obtain the response values ​​for all channels and frequencies. The vector is then concatenated and flattened along the channel and frequency dimensions to form a vector of length C·K. This vector is then input into a multilayer perceptron (MLP, consisting of several fully connected layers) for compression mapping to obtain a local frequency domain description vector. Using the local frequency domain to describe vectors As a conditional input, candidate defect regions are generated through a hypernetwork. The corresponding complex frequency domain filter;

[0095] The supernetwork consists of two fully connected layers, which output the real and imaginary parts of the filter, respectively. , Small variance orthogonal initialization (typical standard deviation of 0.01) is used to make the initial filter response close to unity gain, ensuring that the original spectral structure is not destroyed in the early stage of training, and the real part is... and the virtual part The combination is a complex low-frequency core, anchored at the spectral center and embedded into a full-spectrum grid of size H×W. The remaining frequency positions are fixed at 1. To strictly ensure that the inverse Fourier transform result is a real signal, the complex frequency domain filter... By constructing a conjugate symmetry model, the upper half-spectrum (containing the DC component) is directly output from the supernetwork, and the lower half-spectrum is obtained by conjugating the upper half-spectrum, i.e.:

[0096]

[0097] in, It is a complex conjugate. , It is a frequency index;

[0098] Construct the final complex frequency domain filter:

[0099]

[0100] in, It is the final complex frequency domain filter, where j is the imaginary unit;

[0101] Using the final complex frequency domain filter Global context feature regions Perform a two-dimensional discrete Fourier transform (DFT) on each channel to obtain a complex spectral representation. For complex spectrum Perform the fftshift operation to center the zero-frequency component in the spectrum, and then apply the shifted complex spectrum. With the final complex frequency domain filter Perform element-wise complex multiplication:

[0102]

[0103] in, It is the enhanced spectrum after spectral shifting and the application of a complex domain filter. It is element-wise complex multiplication;

[0104] Enhanced spectrum after spectral shifting and application of complex domain filters Perform inverse FFT shift, then calculate the two-dimensional inverse discrete Fourier transform (IDFT). Take the real part of the IDFT result (the imaginary part is theoretically zero due to conjugate symmetry) to obtain the enhanced channel features. Concatenate all the enhanced channels to obtain the final enhanced global context features. .

[0105] By precisely locating candidate defect regions and cropping local residual blocks containing rich contextual information, the recognition accuracy of various defects in complex backgrounds is greatly improved. Next, a lightweight convolutional backbone network based on FFC is adopted, flexibly adjusting local and global information processing strategies. This preserves key details while enhancing the understanding of global patterns, making the model more adaptable and robust. Furthermore, by combining RoIAlign operations and frequency domain analysis techniques, the periodic or quasi-periodic structural information of candidate defect regions is accurately captured, significantly enhancing the model's ability to distinguish subtle structural changes. Finally, by generating complex frequency domain filters through a supernetwork and performing spectral enhancement on global contextual features, not only is the original spectral structure preserved, but the stability and reliability of the model in practical applications are also improved, significantly increasing the accuracy and efficiency of defect detection.

[0106] S3. Based on enhanced global context features, a meta-learning mechanism is used to obtain new screen type adaptation parameters for defect category probability prediction. Uncertainty is estimated based on the defect category probability prediction results, and active labeling and updating are triggered by predicting uncertainty.

[0107] Specifically, based on enhanced global context features, a meta-learning mechanism is used to obtain new screen type adaptation parameters for defect category probability prediction. This means that for each candidate defect region, the corresponding final enhanced global context features are used. Perform global average pooling (or max pooling) to obtain the feature vector. Simultaneously, the structured screen background prior vector Mapped to embedding vectors via a small fully connected network. and will , Concatenation yields the comprehensive feature vector Input a learnable classification function ;

[0108] The classification function It consists of a multilayer perceptron (MLP) followed by a classification layer, outputting the probability of defect categories. It is trained using the Model-Agnostic Meta-Learning (MAML) framework, on multiple historical OLED screen type datasets, with each screen type treated as an independent task. By optimizing shared initial parameters, the classification function is optimized. High accuracy can be achieved on any new task with only a small number of samples and a few gradient updates. The meta-training objective is:

[0109]

[0110] in, These are the optimal initial parameters obtained after optimization. This is the 0th screen type task. It is the inner loop learning rate. It is cross-entropy loss. yes In terms of tasks, starting from the initial parameters Starting from this point, the adaptation parameters are obtained after one (or several) gradient updates. These are the "initial parameters to be optimized". From task distribution Random sampling task Seeking expectations;

[0111] When the production line switches to the new screen type, collected labeled samples form a support set. The optimal initial parameters obtained after optimization Starting from the first parameter, perform gradient descent to obtain the adaptation parameters:

[0112]

[0113] in, These are the adaptation parameters when deployed to a new screen production line;

[0114] To adapt parameters and the combined feature vector of each candidate region As input, through the classification function Perform one deterministic forward propagation (with Dropout disabled) to obtain the defect category prediction probability:

[0115]

[0116] in, It is the predicted probability that the nth candidate defect region belongs to the mth type of defect. It is an equivalent model generated due to different Dropout random masks, where M is the preset number of defect categories (including bright spots, dark spots, linear Mura, planar Mura, normal, etc.).

[0117] By simplifying feature representation through global pooling and enhancing model understanding by combining background prior information, classification accuracy is significantly improved. Secondly, the adoption of the MAML framework enables the model to quickly adapt to new tasks, greatly improving the flexibility and efficiency during production line switching. Finally, the personalized adaptation process for new screen types ensures high performance of the model in various application scenarios. This method improves the accuracy and efficiency of defect detection.

[0118] Furthermore, uncertainty estimation is performed based on the probability prediction results of defect categories. Active labeling and updating are triggered by the prediction uncertainty, which refers to the probability prediction results of defect categories based on each candidate defect region. In order to avoid calibration distortion caused by the mixing of different probability spaces, only intra-model uncertainty estimation based on MC Dropout and inter-model uncertainty estimation based on multi-model consistency enhancement (MCA) are enabled in any detection cycle, which is determined by the production line configuration strategy.

[0119] The model internal uncertainty estimation based on MC Dropout is performed with "Single Model Uncertainty Assessment" enabled in the system configuration. The process includes temporarily enabling the Dropout layer during the inference phase (with a fixed inactivation rate of 0.5), performing T independent forward propagations on the comprehensive feature representation of the same candidate region, and obtaining the probability set of each class. :

[0120]

[0121] in, It represents the predicted probability of the nth candidate defect region belonging to each category during the tth independent forward propagation. It is the equivalent model formed by the different Dropout random mask in the t-th iteration;

[0122] Calculate the MC mean probability vector based on the probability output of each category:

[0123]

[0124] in, It is the MC mean probability vector. It is the probability that the nth candidate defect region is predicted to be a defect of type m in the tth forward propagation.

[0125] The inter-model uncertainty estimation based on Multi-Model Consistency Enhancement (MCA) is enabled when the system configuration selects "Multi-Model Consistency Assessment" and a valid historical snapshot exists. The process includes constructing a set of classification head parameters. The parameter set is derived from model versions that have undergone effective incremental updates under the same screen type (e.g., "effective" means that the number of newly added samples in the cache set on that day is ≥5). The backbone network parameters are frozen, and only the classification head parameters differ. The prediction probability is calculated for each member model:

[0126]

[0127] in, It is a prediction probability. These are the adaptation parameters for the r-th member model;

[0128] Based on predicted probability Calculate the probability of the MCA mean:

[0129]

[0130] in, It is the MCA mean probability. It is the probability that the r-th member model predicts the n-th candidate defect region as a defect of the m-th type;

[0131] The mean probability vector obtained by any method Calculate Shannon entropy as uncertainty:

[0132]

[0133] in, This represents the fundamental forecast uncertainty, and log(·) is the natural logarithm. It is the numerical stability constant. It is the mean probability that the nth candidate defect region is predicted as the mth type defect;

[0134] When forecast uncertainty Less than or equal to the uncertainty threshold (When the validation set AUC-PR is maximized, typically ranging from [0.5, 1.0]), the automatic classification result is accepted, and the label is:

[0135]

[0136] in, This is the result of automatic classification;

[0137] When forecast uncertainty Greater than the uncertainty threshold At that time, candidate regions are included in the high-risk sample pool. To avoid duplicate submissions of highly similar defects on the same screen or in adjacent areas, a high-risk sample pool is selected. Perform feature similarity-based redundancy removal screening (ENMS). For any two candidate regions, define a comprehensive similarity to determine whether they are duplicate detections of the same type of defect.

[0138]

[0139] in, is the "comprehensive similarity" between the nth candidate defect region and the sth candidate defect region, cos(·) is the cosine similarity of the feature vectors, and IoU(·) is the intersection-union ratio of the candidate region bounding boxes. This is a weighting coefficient, set empirically, with a typical value range of [0.4, 0.6]. It is the comprehensive feature vector of the s-th candidate defect region. It is the bounding box of the nth candidate region. It is the bounding box of the s-th candidate region. It is the comprehensive feature vector of the nth candidate defect region;

[0140] when At that time, among them, This is a comprehensive similarity threshold, set empirically and through offline parameter tuning, with a typical value range of [0.6, 0.8]. Only candidate regions with higher basic prediction uncertainty are retained. When the number of candidate defective regions after screening exceeds the preset review budget... When the specific production line scale, product complexity, and manpower allocation depend on the actual production line size, product complexity, and workforce allocation, select the products from highest to lowest uncertainty. Each candidate defect region enters the manual review queue. For the selected candidate defect regions, a manual review data package is constructed, including the candidate region bounding box in the original image, the original image ROI, the corresponding ROI in the enhanced residual image, the predicted defect category, the classification confidence score, and the uncertainty value. The manual review obtains the true label, and the samples are stored in the screen-specific cache set. After the end of each production day, an incremental fine-tuning is performed on the newly added samples in the cache set, including freezing the backbone network, updating only the classification head parameters, and training and updating using the cross-entropy loss function to obtain new adaptation parameters.

[0141] By constructing a consistent uncertainty estimation mechanism within a single cycle (MC Dropout or MCA), probability calibration distortion is effectively avoided, providing a reliable confidence basis for automatic discrimination. Uncertainty is predicted using Shannon entropy quantification, and offline optimized thresholds are combined to achieve automatic release of high-confidence samples and accurate interception of high-risk samples. Furthermore, based on the feature and location fusion-based ENMS redundancy removal strategy, duplicate submissions are eliminated, ensuring that manual review focuses on truly difficult samples. On this basis, a review queue sorted by uncertainty priority and subject to budget constraints is established, along with a multimodal review data package, to improve annotation efficiency and quality. Finally, new annotated samples are efficiently integrated into the model through daily incremental fine-tuning, forming a "detection-evaluation-review-learning" closed loop, significantly reducing the misjudgment rate and labor costs, while continuously enhancing the model's adaptability to new screen types and new defects.

[0142] This embodiment also provides an automatic detection and classification system for micro-defects in OLED displays, including:

[0143] The image acquisition and preprocessing module is used to acquire OLED images under multi-grayscale conditions, and after registration, averaging and brightness linearization, output a low-noise linearized image sequence.

[0144] The pixel response modeling and enhancement module is used to generate an ideal background image based on a pixel-level grayscale-brightness model, and obtain an enhanced residual feature map through differential and multi-scale residual fusion.

[0145] The candidate defect region extraction module is used to perform adaptive threshold segmentation and morphological processing on the enhanced residual map, extract candidate defect regions, and calculate geometric and statistical features.

[0146] The frequency domain feature fusion and classification module enhances the global context through a frequency domain attention mechanism, enabling the fusion and classification of local and global features.

[0147] The Meta-Learning and Uncertainty Learning module is used to perform cross-screen adaptive classification using a meta-learning mechanism, quantify and predict uncertainty, trigger active labeling, and achieve incremental model updates.

[0148] This embodiment also provides a computer device applicable to an automatic detection and classification method for micro-defects in an OLED display screen, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the automatic detection and classification method for micro-defects in an OLED display screen as proposed in the above embodiment.

[0149] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0150] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the automatic detection and classification method for micro-defects in an OLED display screen as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

Claims

1. An automatic detection and classification method for micro-defects in OLED displays, characterized in that: include, Acquire OLED display images and perform preprocessing. Based on the preprocessed images, perform pixel response modeling and generate an ideal background image. Perform differential processing and multi-scale residual feature enhancement to obtain an enhanced residual image. The enhanced residual image is segmented to extract a set of candidate defect regions, and global features are extracted. Based on the set of candidate defect regions, a complex filter guided by local frequency domain prior is used to enhance the global features in the frequency domain to obtain enhanced global context features. Based on enhanced global context features, a meta-learning mechanism is used to obtain new screen type adaptation parameters for defect category probability prediction. Uncertainty estimation is performed based on the defect category probability prediction results, and active labeling and updating are triggered by predicting uncertainty.

2. The automatic detection and classification method for micro-defects in an OLED display screen as described in claim 1, characterized in that: The process of acquiring and preprocessing OLED display images refers to controlling the OLED screen under test to sequentially display a set of standard gray levels covering the working dynamic range in a light-shielding environment. After the display state of each standard gray level stabilizes, original display images are continuously acquired, and the original display images under the same gray level are preprocessed to obtain an image sequence.

3. The automatic detection and classification method for micro-defects in an OLED display screen as described in claim 2, characterized in that: The process of modeling pixel response based on preprocessed images and generating an ideal background image for differential processing and multi-scale residual feature enhancement to obtain an enhanced residual image involves receiving an image sequence as input, extracting brightness response values, establishing a pixel-level grayscale-brightness response model, calculating the theoretical brightness of each pixel in the entire screen, generating an ideal background image, subtracting the ideal background image from the linearized image under the current standard grayscale pixel by pixel to obtain an initial residual image, constructing a Laplacian pyramid on the initial residual image for multi-scale decomposition, and performing weighted fusion to form the final enhanced residual feature map. Based on the mean and standard deviation of the pixel-level grayscale-brightness response model parameters, and combined with the fixed properties of the OLED screen, a prior vector for the screen background is constructed.

4. The automatic detection and classification method for micro-defects in an OLED display screen as described in claim 3, characterized in that: The step of segmenting the enhanced residual image to extract a set of candidate defect regions refers to binarizing the final enhanced residual feature map to obtain a binary image, performing an opening operation, performing connected component analysis based on the processed binary image, identifying all independent candidate defect regions, and extracting a set of geometric features for each candidate defect region, including bounding box coordinates, area, perimeter, major and minor axes, and shape factor, to form a set of candidate defect regions.

5. The automatic detection and classification method for micro-defects in an OLED display screen as described in claim 4, characterized in that: The process of extracting global features involves using a set of candidate defect regions and a local frequency domain prior-guided complex filter to enhance the global features in the frequency domain, thereby obtaining enhanced global context features. This involves calculating the coordinates of the center point of the bounding box corresponding to each candidate defect region in the set of candidate defect regions, and cropping local residual blocks from the enhanced residual feature map with the coordinates of the center point of the bounding box of each candidate defect region as the center. At the same time, the enhanced residual feature map is used as the global input image and input into a lightweight convolutional backbone network constructed based on FFC. The backbone network consists of convolutional blocks and includes batch normalization, ReLU activation function and downsampling operation. In the shallow stage of the backbone network, only local convolutional paths are retained. In the deep stage of the network, the complete FFC module is enabled to fuse local and global information and output a global feature map. The bounding boxes of candidate defect regions are mapped to the coordinate system of the global feature map. Global contextual feature regions are extracted using RoIAlign. Simultaneously, local residual blocks are input into a shared lightweight encoder to obtain local feature maps, and response values ​​are calculated to form local frequency domain description vectors. The conditional input supernetwork generates the final complex frequency domain filter. For global context feature regions Each channel undergoes DFT and FFTshift, resulting in the final complex frequency domain filter. Element-wise multiplication, followed by inverse FFT shift and IDFT to obtain the real part, yields the enhanced channel features. These are then concatenated to obtain the final enhanced global context features. .

6. The automatic detection and classification method for micro-defects in an OLED display screen as described in claim 5, characterized in that: The method of using a meta-learning mechanism to obtain new screen type adaptation parameters based on enhanced global context features for defect category probability prediction refers to, for each candidate defect region, using the corresponding final enhanced global context features... Perform global average pooling to obtain the feature vector. Simultaneously, the structured screen background prior vectors are mapped to embedding vectors through a small fully connected network. and will , The combined feature vector is obtained by concatenating the features, which is then input into a multilayer perceptron and followed by a classification layer to form a learnable classification function. The classification function is trained using the MAML framework. On the historical OLED screen dataset, each screen type is treated as an independent task, and shared initial parameters are optimized. When the production line switches to a new screen type, labeled samples are collected to form a support set. Starting with the optimal initial parameters obtained after optimization, gradient descent is performed to obtain the adaptation parameters. The adaptation parameters and the comprehensive feature vector of each candidate region are used as input, and forward propagation is performed through the classification function to obtain the class probability vector.

7. The automatic detection and classification method for micro-defects in an OLED display screen as described in claim 6, characterized in that: The uncertainty estimation based on the defect category probability prediction results, and the active labeling and updating triggered by the prediction uncertainty, refer to the use of only intra-model uncertainty estimation based on MCDropout or inter-model uncertainty estimation based on multi-model consistency enhancement in any detection cycle based on the defect category probability prediction results of each candidate defect region. Based on the mean probability vector obtained by any method, Shannon entropy is calculated as the prediction uncertainty. When the prediction uncertainty is less than or equal to the uncertainty threshold... When the prediction uncertainty exceeds the uncertainty threshold, the automatic classification result is accepted. When candidate defect regions are included in a high-risk sample pool, a redundancy removal screening based on feature similarity is performed on the high-risk sample pool. When the number of candidate defect regions after screening exceeds the preset review budget, the previous ones are selected in descending order of uncertainty. Each candidate defect region enters the manual review queue. For the selected candidate defect regions, a manual review data package is constructed and stored in the screen-specific cache set. After the end of each production day, incremental fine-tuning is performed on the newly added samples in the cache set to obtain new adaptation parameters.

8. An automatic detection and classification system for micro-defects in an OLED display screen, based on the automatic detection and classification method for micro-defects in an OLED display screen according to any one of claims 1 to 7, characterized in that: include, The image acquisition and preprocessing module is used to acquire OLED images under multi-grayscale conditions, and after registration, averaging and brightness linearization, output a low-noise linearized image sequence. The pixel response modeling and enhancement module is used to generate an ideal background image based on a pixel-level grayscale-brightness model, and obtain an enhanced residual feature map through differential and multi-scale residual fusion. The candidate defect region extraction module is used to perform adaptive threshold segmentation and morphological processing on the enhanced residual map, extract candidate defect regions, and calculate geometric and statistical features. The frequency domain feature fusion and classification module enhances the global context through a frequency domain attention mechanism, enabling the fusion and classification of local and global features. The Meta-Learning and Uncertainty Learning module is used to perform cross-screen adaptive classification using a meta-learning mechanism, quantify and predict uncertainty, trigger active labeling, and achieve incremental model updates.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the automatic detection and classification method for micro-defects in an OLED display screen as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the automatic detection and classification method for micro-defects in an OLED display screen as described in any one of claims 1 to 7.

Images