Defect detection methods, apparatus, equipment, and storage media for metal photomasks

Abstract

This disclosure relates to a method, apparatus, device, and storage medium for defect detection of metal photomasks. The method includes: acquiring a target image, wherein the target image represents a metal photomask processed by a wire mesh fabrication process; performing defect detection on the target image using an automated optical inspection device to obtain an image region containing defects in the target image; determining the target defect type of the image region and the corresponding confidence level of the target defect type using a pre-trained artificial intelligence model, wherein the target defect type represents a real defect or a false defect; determining whether to perform defect verification on the image region based on the target defect type and confidence level; if it is determined that defect verification should be performed on the image region, determining the verification result of the image region to obtain the target detection result of the image region. This improves the detection efficiency and accuracy of metal photomasks.

Description

Technical Field

[0001] This disclosure relates to the field of industrial inspection technology, and more specifically, to a method, apparatus, equipment, and storage medium for defect detection of metal masks. Background Technology

[0002] Metal masks (including open masks, CVD masks, etc.) are core components in vapor deposition and film formation processes. Their surface morphology and defect state directly determine the film formation accuracy, display yield, and lifespan of the device. In the complete production process of metal masks, high-precision scanning and inspection of the entire surface of the metal mask is usually required to identify various defects, so as to ensure the quality of the metal mask leaving the factory and control the yield of downstream production.

[0003] In related technologies, defect detection of metal masks is generally based on traditional algorithms such as grayscale threshold comparison, contour template matching, and edge feature extraction. The core detection logic is that "non-compliant graphics are defects." However, the detection accuracy of these methods is relatively low. To improve accuracy, the detection results need to be reviewed, requiring the processing of a large amount of data, which leads to low detection efficiency.

[0004] It is evident that improving the detection efficiency and accuracy of metal photomasks is a technical issue worthy of attention. Summary of the Invention

[0005] The main purpose of this disclosure is to provide a method, apparatus, device, and storage medium for defect detection of metal masks, in order to solve the technical problem of how to improve the detection efficiency and accuracy of metal masks in the prior art, and to achieve the technical effect of improving the detection efficiency and accuracy of metal masks.

[0006] To achieve the above objectives, a first aspect of this disclosure provides a method for defect detection in metal photomasks, comprising: Acquire the target image, where the target image represents the metal mask after the wire mesh stretching process; The defect detection of the target image is performed by an automated optical inspection device to obtain the image region in the target image where defects exist; By using a pre-trained artificial intelligence model, the target defect type of the image region and the confidence level corresponding to the target defect type are determined, where the target defect type represents a real defect or a false defect. Based on the target defect type and confidence level, determine whether to perform defect verification on the image region; If it is determined that a defect verification will be performed on the image region, then the verification result of the image region will be determined, and the target detection result of the image region will be obtained.

[0007] In some possible implementations, the artificial intelligence model is trained in the following manner: Obtain a training sample set; wherein the training samples in the training sample set include sample images and sample labels; the sample images include image regions with defects as determined by an automated optical inspection device; the sample labels represent real defects or false defects; real defects include at least one of the following: hole defects, edge burrs, scratches, foreign matter residue, and mesh deformation; false defects include at least one of the following: reflection artifacts, dust adhesion, inherent texture of the substrate, imaging noise, and non-functional traces; The artificial intelligence model is trained by using machine learning algorithms, taking the sample images included in the training sample set as input data and the sample labels included in the training sample set as the expected output data.

[0008] In some possible implementations, the sample images included in the training sample set are generated in the following manner: The defective image is obtained by acquiring the image area with defects as determined by an automated optical inspection device. The defect image is preprocessed to obtain the processed image, wherein the preprocessing includes at least one of the following: size normalization, grayscale calibration, and noise reduction. The processed image is augmented to obtain a sample image, wherein the augmentation process includes at least one of the following: flipping, rotating, brightness adjustment, contrast adjustment, cropping, and stitching.

[0009] In some possible implementations, after determining the verification result of the image region and obtaining the target detection result of the image region if it is determined to perform defect verification on the image region, the method further includes: Obtain quality feedback images for the metal mask, along with the labels on the quality feedback images; The first training sample is generated based on the quality feedback image and its label. If the verification result does not match the target defect type, then determine the label of the image region; A second training sample is generated based on the image region and its label. Using the first and second training samples, the pre-trained artificial intelligence model is incrementally trained to obtain the incrementally trained artificial intelligence model.

[0010] In some possible implementations, the learning rate during incremental training is smaller than the learning rate during pre-training.

[0011] In some possible implementations, the artificial intelligence model has a convolutional neural network architecture; the loss function of the artificial intelligence model includes the cross-entropy loss function.

[0012] In some possible implementations, determining whether to perform defect verification on the image region based on the target defect type and confidence level includes: If the target defect type represents a real defect, or if the confidence level is less than or equal to a preset threshold, then the image region is determined to undergo defect verification. If the target defect type indicates a false defect and the confidence level is greater than a preset threshold, then it is determined that no defect verification will be performed on the image region.

[0013] Secondly, embodiments of this disclosure provide a defect detection device for a metal mask, comprising: The first acquisition unit is configured to acquire a target image, wherein the target image represents a metal mask after being processed by the wire mesh stretching process; The detection unit is configured to perform defect detection on the target image using an automated optical inspection device to obtain the image region in the target image where a defect exists. The first determining unit is configured to: determine the target defect type of the image region and the confidence level corresponding to the target defect type through a pre-trained artificial intelligence model, wherein the target defect type represents a real defect or a false defect; The second determining unit is configured to: determine whether to perform defect verification on the image region based on the target defect type and confidence level; The third determining unit is configured to: if it is determined that a defect verification of the image region is to be performed, determine the verification result of the image region and obtain the target detection result of the image region.

[0014] In some possible implementations, the artificial intelligence model is trained in the following manner: Obtain a training sample set; wherein the training samples in the training sample set include sample images and sample labels; the sample images include image regions with defects as determined by an automated optical inspection device; the sample labels represent real defects or false defects; real defects include at least one of the following: hole defects, edge burrs, scratches, foreign matter residue, and mesh deformation; false defects include at least one of the following: reflection artifacts, dust adhesion, inherent texture of the substrate, imaging noise, and non-functional traces; The artificial intelligence model is trained by using machine learning algorithms, taking the sample images included in the training sample set as input data and the sample labels included in the training sample set as the expected output data.

[0015] In some possible implementations, the sample images included in the training sample set are generated in the following manner: The second acquisition unit is configured to: acquire the image region with defects as determined by an automatic optical inspection device, and obtain a defect image; The first processing unit is configured to: preprocess the defect image to obtain a processed image, wherein the preprocessing includes at least one of the following: size normalization processing, grayscale value calibration, and noise reduction processing; The second processing unit is configured to perform augmentation processing on the processed image to obtain a sample image, wherein the augmentation processing includes at least one of the following: flipping, rotating, brightness adjustment, contrast adjustment, cropping, and stitching.

[0016] In some possible implementations, after determining the verification result of the image region and obtaining the target detection result of the image region if it is determined to perform defect verification on the image region, the apparatus further includes: The third acquisition unit is configured to acquire a quality feedback image of the metal mask and a label for the quality feedback image. The first generation unit is configured to generate the first training sample based on the quality feedback image and the label of the quality feedback image; The fourth determining unit is configured to: determine the label of the image region if the verification result does not match the target defect type; The second generation unit is configured to generate a second training sample based on the image region and the label of the image region. The training unit is configured to incrementally train the pre-trained artificial intelligence model using the first training sample and the second training sample to obtain the incrementally trained artificial intelligence model.

[0017] In some possible implementations, the learning rate during incremental training is smaller than the learning rate during pre-training.

[0018] In some possible implementations, the artificial intelligence model has a convolutional neural network architecture; the loss function of the artificial intelligence model includes the cross-entropy loss function.

[0019] In some possible implementations, determining whether to perform defect verification on the image region based on the target defect type and confidence level includes: If the target defect type represents a real defect, or if the confidence level is less than or equal to a preset threshold, then the image region is determined to undergo defect verification. If the target defect type indicates a false defect and the confidence level is greater than a preset threshold, then it is determined that no defect verification will be performed on the image region.

[0020] Thirdly, embodiments of this disclosure provide an electronic device, including: Memory, used to store computer programs; A processor is configured to execute a computer program stored in the memory, wherein, when the computer program is executed, it implements the method of any embodiment of the method for defect detection of a metal mask of the first aspect of this disclosure.

[0021] Fourthly, embodiments of this disclosure provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method of any embodiment of the defect detection method for a metal mask as described in the first aspect above.

[0022] Fifthly, embodiments of this disclosure provide a computer program product including computer-readable code, which, when executed by a processor, implements the method of any embodiment of the defect detection method for metal mask described in the first aspect.

[0023] The technical solutions provided by the embodiments of this disclosure may include the following beneficial effects: In this disclosure, a target image can be acquired, representing a metal mask processed by a wire mesh fabrication process. Then, an automated optical inspection device is used to perform defect detection on the target image, obtaining image regions containing defects. Next, a pre-trained artificial intelligence model is used to determine the target defect type and its corresponding confidence level within the image region. The target defect type represents either a real defect or a false defect. Subsequently, based on the target defect type and confidence level, it is determined whether to perform defect verification on the image region. If it is determined that defect verification should be performed, the verification result is determined, resulting in the target detection result for the image region. Therefore, by using an artificial intelligence model to determine the target defect type and confidence level of an image region, and then determining whether to perform defect verification based on the target defect type and confidence level, image regions requiring verification can be selected, thereby reducing the amount of data requiring defect verification and improving the detection efficiency of the metal mask. Furthermore, since the final detection result (i.e., the target detection result) is obtained by combining automated optical inspection equipment, an artificial intelligence model, and defect verification, the detection accuracy of the metal mask can be improved. Attached Figure Description

[0024] The accompanying drawings, which form part of this disclosure, are used to provide a further understanding of the disclosure and to make other features, objects, and advantages of the disclosure more apparent. The illustrative embodiments of the disclosure, along with their descriptions, are used to explain the disclosure and do not constitute an undue limitation thereof. In the drawings: Figure 1 A flowchart illustrating a defect detection method for a metal mask provided in this embodiment of the disclosure; Figure 2 A flowchart of another method for defect detection of a metal mask provided in this disclosure embodiment; Figure 3 A flowchart illustrating yet another method for defect detection of a metal mask provided in this disclosure; Figure 4A schematic diagram of a defect detection device for a metal mask provided in an embodiment of this disclosure; Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this disclosure. Detailed Implementation

[0025] To enable those skilled in the art to better understand the present disclosure, the technical solutions of the present disclosure will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present disclosure, and not all embodiments. Based on the embodiments of the present disclosure, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present disclosure.

[0026] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this disclosure described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0027] In this disclosure, the terms "upper," "lower," "left," "right," "front," "rear," "top," "bottom," "inner," "outer," "middle," "vertical," "horizontal," "lateral," and "longitudinal" indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings. These terms are primarily for the purpose of better describing this disclosure and its embodiments, and are not intended to limit the indicated devices, elements, or components to having a specific orientation, or to be constructed and operated in a specific orientation.

[0028] Furthermore, in addition to indicating location or positional relationship, some of the aforementioned terms may also have other meanings. For example, the term "above" may also be used in certain circumstances to indicate a dependency or connection. Those skilled in the art can understand the specific meaning of these terms in this disclosure according to the specific circumstances.

[0029] Furthermore, the terms "installation," "setup," "equipped with," "connection," "linked," and "socketing" should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral structure; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, or an internal connection between two devices, components, or parts. Those skilled in the art can understand the specific meaning of the above terms in this disclosure according to the specific circumstances.

[0030] like Figure 1 As shown, the method specifically includes: Step 101: Obtain the target image, where the target image represents the metal mask after the wire mesh stretching process.

[0031] In this embodiment, the target image can be an image obtained by capturing images of the metal mask after the wire mesh stretching process using an image acquisition device. In some optional implementations, the image acquisition device can be located in an AOI (Automated Optical Inspection) device. Therefore, an AOI device can be used to perform a full-surface scan of the metal mask after the wire mesh stretching process, thereby outputting the original inspection image, which can then be used as the target image.

[0032] The wire mesh stretching process is a step in the metal photomask production process. Here, the wire mesh stretching process is used to stretch and fix the metal photomask.

[0033] A metal mask is a physical masking template made of high-precision metal sheets (such as Invar, stainless steel, molybdenum, nickel, etc.) with micron / submicron level openings or patterns. Metal masks can be used in processes such as vacuum evaporation, sputtering, and vapor deposition to precisely deposit film materials only in designated areas, achieving patterned processing. In some optional implementations, the aforementioned metal mask can be an OLED (Organic Light-Emitting Diode) metal mask, a CIS (Complementary Metal Oxide Semiconductor Image Sensor) metal mask, etc. Here, the OLED metal mask is a core component in the OLED evaporation and film deposition process; its surface morphology and defect state directly determine the film deposition accuracy, display yield, and lifespan of the OLED device. As examples, the type of OLED metal mask can be an Open Mask, a CVD Mask, etc.

[0034] Automated optical inspection equipment can be used to scan and detect defects across the entire surface of a metal mask after the wire mesh stretching process, thereby outputting target images and defect detection data. The defect detection data may include, for example, the coordinates of the detected defects, a partial screenshot of the defect (i.e., the image area), and preliminary judgment data on the defect type (e.g., grayscale defects, planar defects, point defects, etc.).

[0035] In some alternative implementations, the target image output by the AOI device can be obtained by communicating with the AOI device via Ethernet.

[0036] Step 102: Defect detection is performed on the target image using an automated optical inspection device to obtain the image region in the target image that contains defects.

[0037] Defect detection can be used to identify various defects in target images, such as holes, burrs on edges, scratches, foreign matter residue, and deformation of the mesh.

[0038] The image area can be a local screenshot corresponding to the defect detected by the AOI equipment.

[0039] In some alternative implementations, AOI devices can use template matching (comparing with a standard defect-free image), pixel-level / region-level threshold segmentation, edge detection, morphological analysis, and other methods to detect defects in the target image.

[0040] In some alternative implementations, the AOI device can output defect detection data after completing the scan, thereby enabling the aforementioned execution entity to extract the image region containing defects from the target image based on the defect detection data.

[0041] Step 103: Using a pre-trained artificial intelligence model, determine the target defect type of the image region and the confidence level corresponding to the target defect type, where the target defect type represents a real defect or a false defect.

[0042] In this embodiment, the artificial intelligence model can be used to classify defects in an image region, thereby determining the type of target defect in the image region, i.e., whether the defect in the image region is a real defect or a false defect. Furthermore, the artificial intelligence model can also determine the confidence level of the target defect type.

[0043] In some alternative implementations, the artificial intelligence model may include a convolutional neural network architecture.

[0044] The target defect type can represent a real defect or a false defect. In practice, real and false defects can be classified based on different criteria. For example, real defects can include holes and blemishes, while false defects can include reflection artifacts.

[0045] In some alternative implementations, a real defect can be one that affects the performance of the metal mask. A fake defect can be one that does not affect the performance of the mask.

[0046] Confidence level indicates the degree to which an artificial intelligence model can trust the current classification result (real defect or false defect).

[0047] In some alternative implementations, the image region can be input into a pre-trained artificial intelligence model, which outputs the target defect type of the image region (i.e., the binary classification result of real defects / false defects) and the confidence level corresponding to the target defect type.

[0048] Step 104: Based on the target defect type and confidence level, determine whether to perform defect verification on the image region.

[0049] In this embodiment, defect review can be used to perform secondary confirmation of the judgment results of the artificial intelligence model.

[0050] In some alternative implementations, the decision to perform defect verification on an image region can be based on filtering conditions determined by the target defect type and confidence level. For example, if the target defect type represents a real defect, then the image region is determined to undergo defect verification. As another example, if the confidence level is less than or equal to a preset threshold, then the image region is determined to undergo defect verification.

[0051] Step 105: If it is determined that the image region will be subject to defect verification, then the verification result of the image region is determined, and the target detection result of the image region is obtained.

[0052] In this embodiment, the review result can be the defect type determined after review (i.e., real defect / false defect).

[0053] In some alternative implementations, defects can be reviewed by human or other means, or another AOI device can be used for defect review.

[0054] The target detection result can be the final detection result of an image region. The target detection result can indicate whether a real defect exists in the image region or not.

[0055] In some optional implementations of this embodiment, the artificial intelligence model is trained in the following manner: First, obtain the training sample set.

[0056] The training sample set includes sample images and sample labels. Sample images comprise image regions with defects identified by automated optical inspection equipment. Sample labels indicate whether the defects are real or false. Real defects include at least one of the following: hole defects, edge burrs, scratches, foreign matter residue, and mesh deformation. False defects include at least one of the following: reflection artifacts, dust adhesion, inherent substrate texture, imaging noise, and non-functional traces.

[0057] The training sample set can include multiple training samples. Sample labels can be determined by manually or automatically annotating the sample images.

[0058] In some alternative implementations, the sample image can be an image obtained by standardizing (e.g., preprocessing, augmentation, etc.) the image region (defect local screenshot) detected by AOI. Alternatively, the sample image can also be the original image region detected by AOI.

[0059] Then, a machine learning algorithm is used to train an artificial intelligence model by taking the sample images included in the training sample set as input data and the sample labels included in the training sample set as expected output data.

[0060] In some alternative implementations, the sample images included in the training samples in the training sample set can be used as the input data of the initial model (e.g., a convolutional neural network), and the sample labels included in the training samples in the training sample set can be used as the expected output data of the initial model. The sample images and corresponding sample labels are input into the initial model, and the initial model outputs predicted labels. The error between the predicted labels and the sample labels is calculated through a loss function. The model parameters of the initial model are updated using the backpropagation algorithm. The model is iterated and optimized repeatedly until the accuracy and recall of the model on the validation set both reach the preset indicators, thereby obtaining the trained artificial intelligence model.

[0061] It is understandable that by acquiring a training sample set containing sample images and sample labels, where the sample images originate from defect areas detected by AOI and the sample labels accurately distinguish between real and false defects, higher-quality and more relevant supervisory information can be provided for the training of the artificial intelligence model. Using machine learning algorithms, with sample images as input data and sample labels as the expected output data for training, the artificial intelligence model can learn the mapping relationship between the original defect image and the target defect type (real defect / false defect). Therefore, the trained artificial intelligence model can more accurately distinguish between real and false defects, thereby further improving the detection accuracy of metal masks.

[0062] In some application scenarios of the above optional implementation methods, the sample images included in the training sample set are generated in the following way: The first step is to acquire the image area with defects as determined by an automated optical inspection device, thus obtaining the defect image.

[0063] The defect image can be a partial screenshot of the defect as described above.

[0064] The second step is to preprocess the defect image to obtain the processed image. The preprocessing includes at least one of the following: size normalization, grayscale calibration, and noise reduction.

[0065] Size normalization can uniformly scale defective images to the fixed pixel size required by the input of artificial intelligence models.

[0066] Grayscale calibration can normalize the pixel values ​​of defective images to the 0-1 range to eliminate interference caused by differences in imaging brightness between different AOI devices.

[0067] Denoising can be achieved using Gaussian filtering and median filtering algorithms to remove salt-and-pepper noise and random noise during AOI imaging.

[0068] The third step is to perform augmentation processing on the processed image to obtain the sample image. The augmentation processing includes at least one of the following: flipping, rotating, brightness adjustment, contrast adjustment, cropping, and stitching.

[0069] Augmentation processing can be used to enrich the training sample set and avoid overfitting of artificial intelligence models.

[0070] It is understandable that by preprocessing the defect images identified by AOI, we can obtain processed images that conform to the input specifications of artificial intelligence models and eliminate equipment differences and noise interference. Augmentation processing of these processed images can enhance both the quantity and diversity of the sample images. This improves the generalization ability and robustness of the artificial intelligence model, avoids overfitting, and further improves the detection accuracy of metal masks.

[0071] In some optional implementations of this embodiment, the artificial intelligence model has a convolutional neural network architecture. The loss function of the artificial intelligence model includes the cross-entropy loss function.

[0072] The convolutional neural network architecture can use ResNet50 (a 50-layer residual network) as the backbone network, with the layers connected in the following order: Backbone network layer: ResNet50 serves as the core of feature extraction, extracting deep features from the input image (such as image regions) through multiple convolutional blocks (including residual connections), gradually reducing the spatial dimension and increasing the channel dimension.

[0073] Global average pooling layer: Connected after the backbone network, it performs global average pooling on the feature maps output by convolution, compressing high-dimensional features into feature vectors of fixed length.

[0074] Fully connected layer: The pooled feature vector is input into two fully connected layers, and the feature information is further integrated through non-linear transformation.

[0075] Softmax classification layer: As the output layer, it receives the feature vectors from the fully connected layer and outputs the target defect type and corresponding confidence level representing the real / false defects.

[0076] Thus, the input image region is preprocessed and then enters the ResNet50 backbone network. The extracted features pass through the global average pooling layer and the fully connected layer in sequence, and finally the classification result is output by the Softmax layer. The modules are connected sequentially through the data flow to form an end-to-end inference link.

[0077] In some alternative implementations, the pooling layer can be a global average pooling (GAP) layer.

[0078] It is understandable that by using the above-mentioned artificial intelligence model structure and adopting the cross-entropy loss function as the loss function of the artificial intelligence model, the classification accuracy and inference efficiency of the artificial intelligence model can be improved.

[0079] It should be noted that, where there is no conflict, the technical features described in different alternative implementations can be included in the same embodiment. For the sake of brevity, they will not be elaborated here.

[0080] Based on the embodiments of this disclosure, a target image can be acquired, where the target image represents a metal mask processed by a wire mesh forming process. Then, an automated optical inspection device is used to perform defect detection on the target image, obtaining image regions containing defects. Next, a pre-trained artificial intelligence model is used to determine the target defect type and its corresponding confidence level for the image region. The target defect type represents either a real defect or a false defect. Subsequently, based on the target defect type and confidence level, it is determined whether to perform defect verification on the image region. If it is determined that defect verification should be performed, the verification result for the image region is determined, resulting in the target detection result for the image region. Therefore, by using an artificial intelligence model to determine the target defect type and confidence level of the image region, and then determining whether to perform defect verification based on the target defect type and confidence level, image regions requiring verification can be filtered out, thereby reducing the amount of data requiring defect verification and improving the detection efficiency of the metal mask. Furthermore, since the final detection result (i.e., the target detection result) is obtained by combining automated optical inspection equipment, an artificial intelligence model, and defect verification, the detection accuracy of the metal mask can be improved.

[0081] Figure 2 This is a flowchart illustrating another defect detection method for a metal mask provided in an embodiment of this disclosure. Figure 2 As shown, in Figure 1 Following step 105 shown, the method further includes: Step 201: Obtain the quality feedback image for the metal mask and the label of the quality feedback image.

[0082] In this embodiment, the quality feedback image can be: sample data fed back by the production line during the production process regarding situations where the artificial intelligence model misjudges or misses defects.

[0083] Labels on quality feedback images can be manually annotated or automatically labeled using algorithms. These labels can indicate whether the quality feedback image contains a real defect or a false defect.

[0084] Step 202: Generate the first training sample based on the quality feedback image and its label.

[0085] In this embodiment, the first training sample can be a training sample generated based on the quality feedback image and the label of the quality feedback image. As an example, the sample image in the first training sample can be the quality feedback image, and the sample label in the first training sample can be the label of the quality feedback image.

[0086] Step 203: If the verification result does not match the target defect type, then determine the label of the image region.

[0087] In this embodiment, if the review result indicates a real defect and the target defect type indicates a false defect, it can be determined that the review result and the target defect type do not match; or, if the review result indicates a false defect and the target defect type indicates a real defect, it can be determined that the review result and the target defect type do not match.

[0088] Image region labels can be created manually or automatically using algorithms. Image region labels can indicate whether a region contains a real defect or a false defect.

[0089] Step 204: Generate a second training sample based on the image region and its label.

[0090] In this embodiment, the second training sample can be a training sample generated based on an image region and its label. For example, the sample image in the second training sample can be an image region, and the sample label in the second training sample can be the label of the image region.

[0091] Step 205: Using the first training sample and the second training sample, perform incremental training on the pre-trained artificial intelligence model to obtain the incrementally trained artificial intelligence model.

[0092] In this embodiment, incremental training can be a training process that adjusts the model parameters of an artificial intelligence model by using new samples (including the first training sample and the second training sample) on the basis of the original artificial intelligence model.

[0093] In some optional implementations, sample images from newly added samples (including the first and second training samples) can be used as input data for the original AI model, and sample labels from the newly added samples can be used as the expected output data for the original AI model. The sample images and corresponding sample labels are input into the original AI model, and the original AI model outputs predicted labels. The error between the predicted labels and the sample labels is calculated using a loss function. The model parameters of the original AI model are updated using the backpropagation algorithm. The model is iterated and optimized repeatedly until the accuracy and recall of the model on the validation set both reach the preset indicators, thus obtaining the incrementally trained AI model.

[0094] Furthermore, an incrementally trained AI model can be used to replace the original AI model for subsequent inference.

[0095] In some optional implementations of this embodiment, the learning rate during incremental training is less than the learning rate during pre-training.

[0096] In some alternative implementations, the learning rate during incremental training can be less than 0.001, while the learning rate during pre-training can be 0.001.

[0097] Understandably, by using a smaller learning rate during incremental training compared to pre-training, it's possible to retain most of the general features and knowledge already learned by the pre-trained AI model while making small, targeted adjustments to the model weights to adapt to new samples. This avoids overfitting on new data or forgetting previously learned important knowledge due to an excessively large learning rate, thus enabling more stable and efficient iterative updates of the AI ​​model.

[0098] It should be noted that, in addition to the contents described above, this embodiment may also include... Figure 1 The corresponding technical features described in the corresponding embodiments, thereby achieving Figure 1 For details on the technical effectiveness of the defect detection method for the metal mask shown, please refer to [link / reference needed]. Figure 1 The relevant descriptions are presented concisely and will not be elaborated upon here.

[0099] Based on the embodiments of this disclosure, by acquiring quality feedback images and their labels from the production line and generating a first training sample, the inference performance of the artificial intelligence model can be adjusted based on new problems encountered in actual production. When the verification results do not match the AI ​​model's judgment, the correct label is determined and a second training sample is generated, allowing for targeted correction of the AI ​​model's errors. Incremental training of the pre-trained AI model using these two types of new samples can improve the model's inference accuracy. Thus, continuous autonomous iterative improvement of the AI ​​model can be achieved, enabling it to adapt to new production scenarios and defects.

[0100] Figure 3 This is a schematic flowchart illustrating another defect detection method for a metal mask provided in this embodiment of the disclosure. Figure 3 As shown, the method includes: Step 301: Obtain the target image, where the target image represents the metal mask after the wire mesh stretching process.

[0101] In this embodiment, step 301 and Figure 1 Step 101 in the corresponding embodiment is basically the same, and will not be repeated here.

[0102] Step 302: Defect detection is performed on the target image using an automatic optical inspection device to obtain the image region in the target image that has defects.

[0103] In this embodiment, step 302 and Figure 1 Step 102 in the corresponding embodiment is basically the same, and will not be repeated here.

[0104] Step 303: Using a pre-trained artificial intelligence model, determine the target defect type of the image region and the confidence level corresponding to the target defect type, where the target defect type represents a real defect or a false defect.

[0105] In this embodiment, step 303 and Figure 1 Step 103 in the corresponding embodiment is basically the same, and will not be repeated here.

[0106] Step 304: If the target defect type represents a real defect, or if the confidence level is less than or equal to a preset threshold, then determine to perform defect verification on the image region.

[0107] In this embodiment, the preset threshold can be a pre-set value. In some alternative implementations, the preset threshold can be adjusted according to the production line quality requirements. For example, the higher the production line quality requirements, the larger the preset threshold can be.

[0108] In some alternative implementations, the preset threshold can be determined in the following ways: First, a threshold sequence is set. The thresholds in the sequence are arranged in ascending order, such as from largest to smallest or smallest to largest. Each threshold in the sequence corresponds to a product quality level. The product quality level indicates the quality of the metal mask. For example, the threshold sequence may include a strict threshold (e.g., 99.5%), a standard threshold (e.g., 99%), and a lenient threshold (e.g., 98.5%).

[0109] Then, based on at least one of the following: customer quality requirements, production line stage, real-time quality status of the production line, and actual performance of the artificial intelligence model, the currently used preset threshold is matched from the threshold sequence.

[0110] For example, high-end customer orders are forcibly locked at the strict threshold, which remains fixed and cannot be lowered; standard customer orders use the standard threshold by default; and general customer orders can be switched to the lenient threshold or a preset threshold above it based on the production line status.

[0111] For example, strict thresholds are mandatory during the introduction of new metal masks and small-batch trial production stages; standard thresholds are used during the mass production stabilization stage.

[0112] For example, if the production line's real-time quality status is continuously for a preset duration (e.g., 7 days) and the production line yield is greater than or equal to the preset yield value (e.g., ≥99.8%) and there are no batch defect incidents, the threshold can be lowered to the lenient level; if there is one or more customer quality complaints or production line batch yield incidents, the threshold will be immediately raised to the strict level.

[0113] For example, if the actual performance of the artificial intelligence model is as follows: in the most recent preset time period (e.g., 7 days) of manually reviewed data, the actual defect omission rate exceeds the upper limit of the corresponding level control (each threshold in the threshold sequence corresponds to an upper limit of the level control), the threshold will be increased by a preset amount (e.g., 0.2) percentage points; if the manual review rate exceeds the upper limit of the corresponding level control, the threshold will be decreased by a preset amount (e.g., 0.2) percentage points.

[0114] In some optional implementations, the triggering conditions and constraint rules for threshold adjustment can be set in the following ways: Regular adjustments: Threshold calibration is automatically performed once a day at midnight, and the effective threshold is updated based on the production line data of the previous 7 days.

[0115] Event-triggered adjustment: The threshold will be adjusted immediately when any of the following events occur: switching to a mask for production of different models / substrates, completing an incremental training of an artificial intelligence model, or a quality complaint or batch yield incident occurs.

[0116] Constraints: The single threshold adjustment range shall not exceed ±0.5 percentage points, and the final threshold shall not be lower than the lenient threshold or higher than the strict threshold; all threshold adjustment records shall be fully stored in the production database for traceability.

[0117] In some optional implementations, threshold verification and threshold rollback can be performed in the following ways: All automatically adjusted thresholds must first be verified on a small batch of continuously produced photomasks, with a minimum of 50 (e.g., a preset number). If the verification results show that the false negative rate or false positive rate exceeds the control range of the corresponding level, the threshold will be automatically rolled back to the stable threshold of the previous version and an alarm will be sent to the quality inspectors.

[0118] Step 305: If the target defect type indicates a false defect and the confidence level is greater than a preset threshold, then it is determined that no defect verification will be performed on the image region.

[0119] Step 306: If it is determined that the image region will be subject to defect verification, then the verification result of the image region is determined, and the target detection result of the image region is obtained.

[0120] In this embodiment, step 306 and Figure 1 Step 105 in the corresponding embodiment is basically the same, and will not be repeated here.

[0121] It should be noted that, in addition to the contents described above, this embodiment may also include... Figure 1 or Figure 2 The corresponding technical features described in the corresponding embodiments, thereby achieving Figure 1 or Figure 2 For details on the technical effectiveness of the defect detection method for the metal mask shown, please refer to [link / reference needed]. Figure 1or Figure 2 The relevant descriptions are presented concisely and will not be elaborated upon here.

[0122] Based on the embodiments of this disclosure, by setting the aforementioned defect review rules, it can be ensured that all image regions judged as real defects and image regions judged as false defects by the artificial intelligence model but with low confidence are subject to defect review, thereby guaranteeing the detection rate of real defects. Furthermore, image regions judged as false defects with high confidence by the artificial intelligence model can be directly filtered out. Thus, while ensuring product quality, the number of defect images requiring manual review can be minimized, balancing detection efficiency and accuracy.

[0123] The following describes the embodiments of this disclosure by way of example. However, it should be noted that the following content is only used to understand the technical solutions of the embodiments of this disclosure and does not constitute a limitation on the protection scope of the embodiments of this disclosure.

[0124] The OLED display industry has become a core development direction in the field of new displays. Metal masks (including OpenMask, CVD Mask, etc.) are key components in OLED evaporation and film formation processes. Their surface morphology and defect state directly determine the film formation accuracy, display yield, and lifespan of OLED devices. In the complete production process of metal masks, after the mesh stretching process is completed, it is usually necessary to use automated optical inspection equipment to perform high-precision scanning and inspection on the entire surface of the metal mask to identify various defects such as hole defects, edge burrs, scratches, foreign matter residue, and mesh stretching deformation. This is a crucial and essential step to ensure the quality of metal masks leaving the factory and to control the yield of downstream production.

[0125] In some technologies, AOI defect detection of metal masks is based on traditional machine vision algorithms such as grayscale threshold comparison, contour template matching, and edge feature extraction. The core detection logic is "non-compliant graphics are defects." While this approach can detect all abnormal features without omission, it cannot functionally differentiate detected anomalies and cannot accurately identify false defects that do not affect the mask's performance. False defects specifically refer to slight reflective artifacts on the mask surface after mesh stretching, non-conductive dust adhesion, inherent non-functional textures of the substrate, noise in the AOI imaging system, and slight traces generated during metal mask processing that do not affect the evaporation accuracy. These features do not adversely affect OLED evaporation or film formation processes, but are uniformly judged as defects by traditional AOI algorithms, resulting in a large amount of invalid detection data.

[0126] The aforementioned solutions fail to automatically distinguish between real and false defects. False defects constitute a very high percentage of the AOI detection data after mesh fabrication, necessitating a large number of professional quality inspectors for manual review and verification. This significantly increases the labor costs of metal mask production, lengthens the production and inspection cycle, and severely reduces the overall production efficiency of metal masks. Furthermore, manual review relies on the experience and subjective judgment of operators. Different personnel cannot fully agree on the criteria for judging false defects, easily leading to missed detections of real defects and misjudgments of false defects. Missed detections could cause batch yield problems in downstream OLED production, while misjudgments could result in the scrapping of qualified metal masks, significantly increasing production losses and overall costs. Moreover, some machine vision algorithms use fixed parameter settings for defect filtering rules, making them unsuitable for various metal masks of different models, mesh parameters, and substrates, such as Open Masks and CVD Masks. They also cannot continuously optimize the judgment accuracy based on new data generated during production, hindering the iterative improvement of false defect filtering effects and failing to meet the production demands of the rapidly iterating technology and frequent product updates in the OLED industry.

[0127] In view of this, this solution utilizes an AI-based OLED metal mask defect filtering system to automatically determine the authenticity of AOI-detected defects and filter out false defects, reducing labor costs, improving judgment accuracy and production efficiency, and enabling continuous iterative optimization of the model's filtering capabilities. The system includes a data access module, a defect sample preprocessing module, an AI model training module, a defect reasoning and judgment module, a false defect filtering module, a sample iterative update module, and a production database. Specifically: Data access module: Connects to AOI inspection equipment, and accesses the original AOI inspection image (i.e., the target image mentioned above) of the OLED metal mask after the mesh is stretched, the coordinate information of the detected defects, the partial screenshot of the defects, and the preliminary judgment data of the defect type; OLED metal masks include Open Mask, CVD Mask, etc.; Defect sample preprocessing module: Performs standardized preprocessing on the defect sample data obtained by the data access module, including image size normalization, grayscale calibration, noise reduction and enhancement, and data augmentation, and outputs a standardized sample set that meets the input requirements of the artificial intelligence model; Artificial intelligence model training module: Based on the preprocessed labeled sample set, a defect classification artificial intelligence model is built and trained, and the model's weight iteration, accuracy verification and fixed output are completed; the artificial intelligence model adopts a convolutional neural network architecture, and ResNet is selected as the backbone network for industrial defect detection; Defect reasoning and judgment module: Load the solidified artificial intelligence model, reason about the defect samples (i.e. the above image area) detected by AOI in real time, and output the classification result of the defect as a real defect / false defect (i.e. the above target detection type) and the corresponding confidence level; False defect filtering module: Based on the output of the defect reasoning and judgment module, the module automatically filters out defect data that are judged as false defects and whose confidence level meets the preset threshold, and only retains the real defect data and the false defect data whose confidence level does not meet the preset threshold to push to the subsequent process (i.e. defect review). Sample Iteration Update Module: Collects the defect judgment results after manual review (i.e., review results) and the real defect misjudgment / missed judgment data fed back from the production line, supplements them to the labeled sample set, triggers the incremental training of the artificial intelligence model, and realizes the continuous improvement of the filtering capability of the artificial intelligence model; Production database: Used to store raw AOI inspection data, defect sample data, labeled datasets, artificial intelligence model weight files, defect judgment results and iterative update data.

[0128] Specifically, AOI equipment can be a fully automated optical inspection device used in metal mask production lines to complete the full-surface defect scanning and inspection of OLED metal masks after mesh stretching, and output the original inspection image (i.e. the target image mentioned above) and defect detection data.

[0129] The data access module communicates with the AOI device via industrial Ethernet, and receives the detection data (including defect detection data) output by the AOI device in real time. It extracts local screenshots, coordinate information, and preliminary type judgment data of the detected defects, generates a defect sample to be processed (i.e., the image of the above area), and transmits it to the defect sample preprocessing module.

[0130] The defect sample preprocessing module receives the defect samples to be processed and completes standardized preprocessing such as sample size normalization, gray value calibration, and noise reduction and enhancement to generate standardized samples that conform to the input specifications of the artificial intelligence model. Among them, the samples used for model training (i.e. defect images) simultaneously complete data augmentation processing, and the standardized samples are transmitted to the artificial intelligence model training module and the defect inference and judgment module respectively.

[0131] The artificial intelligence model training module builds a convolutional neural network defect classification model with a ResNet50 backbone network based on the standardized sample set with annotations (i.e., the training sample set mentioned above). It completes the iterative training, accuracy verification and weight solidification of the model. The solidified model is then deployed to the defect inference and judgment module, and the model-related data is stored in the production database.

[0132] The defect reasoning and judgment module loads the solidified defect classification artificial intelligence model, performs reasoning calculations on the standardized reasoning samples input in real time, outputs a binary classification result of the sample as a real defect or a fake defect and the corresponding confidence level, and transmits it to the fake defect filtering module.

[0133] The false defect filtering module automatically filters out samples that are judged to be false defects and whose confidence level meets the preset filtering conditions, and only retains real defect data. False defects that do not meet the confidence level are pushed to the subsequent manual review / revision process. All judgment results are synchronously stored in the production database.

[0134] The sample iteration update module collects the defect judgment results that have been manually reviewed and corrected, as well as the production line quality feedback data. After completing the annotation and verification, it is added to the training sample set, triggering the artificial intelligence model training module to perform incremental training, completing the update and optimization of the model weights, and realizing the continuous iteration of the model's filtering capabilities.

[0135] The production database communicates with each of the above modules and is used to store AOI raw inspection data, defect sample data, labeled datasets, artificial intelligence model weight files, defect judgment results and iterative update data.

[0136] This plan includes the following steps: Step 1: Complete the screen fabrication process of the OLED metal mask and transport it to the AOI equipment. Start the AOI equipment to perform full-surface defect scanning and detection on the metal mask, and output detection data including local defect screenshots, coordinate information, and preliminary type judgment data.

[0137] Step two: The data access module completes the real-time connection and access of AOI detection data, extracts local screenshots, coordinate information, and preliminary type judgment data of detected defects, and generates defect samples to be processed.

[0138] Step 3: The defect sample preprocessing module performs standardized preprocessing on the defect samples to be processed, generating standardized inference samples that conform to the input specifications of the artificial intelligence model. For example, the defect local screenshots are first scaled to a fixed pixel size of 256×256 to complete size normalization; then the image pixel values ​​are normalized to the 0-1 range to complete grayscale calibration; finally, median filtering is used to remove imaging noise in the image to generate standardized inference samples that conform to the input specifications of the artificial intelligence model. Optionally, the standardization preprocessing in step three may specifically include: I. Normalize the size of the defect local screenshots and uniformly scale them to the fixed pixel size required by the input of the artificial intelligence model; II. Perform grayscale calibration on the normalized image to normalize the image pixel values ​​to the 0-1 range and eliminate the interference caused by the difference in imaging brightness of different AOI devices. III. Use Gaussian filtering and median filtering to denoise the image and remove salt-and-pepper noise and random noise during the AOI imaging process. IV. Perform data augmentation on the preprocessed samples, including random flipping, rotation, brightness and contrast fine-tuning, cropping and splicing, to enrich the sample set and avoid model overfitting; the augmentation is only used in the model training stage and is not performed in the real-time inference stage.

[0139] Step 4: The defect reasoning and judgment module loads the pre-trained and fixed defect classification artificial intelligence model, performs reasoning calculations on the standardized reasoning samples, and outputs the binary classification result of the sample as a real defect / false defect and the corresponding confidence level. The specific methods for constructing and pre-training the defect classification AI model are as follows: I. Dataset Construction: Collect 100,000 AOI defect samples of Open Mask and CVD Mask from historical production. The samples were labeled by senior quality inspection engineers. Among them, there were 40,000 real defect samples and 60,000 fake defect samples. The labeled dataset was divided into training set, validation set and test set in a ratio of 8:1:1. II. Model Architecture: ResNet50 is used as the backbone network, with a global average pooling layer, two fully connected layers and a Softmax classification layer connected to the backend. The output is a binary classification result and confidence score. The cross-entropy loss function and Adam optimizer are used. The initial learning rate is set to 0.001, the batch size is 32, and the maximum number of iterations is 100. III. Model Training: Input the training set samples into the artificial intelligence model for iterative training. After each round of training, calculate the model classification accuracy using the validation set. When the validation set accuracy no longer improves after 10 consecutive rounds, terminate the training early and save the weight file with the highest validation accuracy. IV. Model Validation and Consolidation: The generalization ability of the trained model was tested using a test set. The test results showed that the model classification accuracy was ≥99.6%, precision was ≥99.5%, and recall was ≥99.7%, which met the threshold requirements for production line use. The model weights were then consolidated and deployed to the defect inference and judgment module.

[0140] Step 5: Determine whether the defect classification result and confidence level meet the false defect filtering conditions: The preset filtering condition is that the classification result is a false defect and the confidence level is ≥99%; if the condition is not met, it is determined to be a real defect, the defect data is retained and pushed to the subsequent manual review and revision process; if the condition is met, it is determined to be a false defect and proceed to the next step. Here, the specific conditions for filtering false defects in step five are: the defect classification result is a false defect, and the classification confidence level is greater than or equal to the preset confidence threshold. The confidence threshold can be dynamically adjusted according to the production line quality requirements and is executed by the threshold management unit built into the false defect filtering module, without the need for manual modification of parameters one by one.

[0141] The specific method for dynamically adjusting the confidence threshold based on production line quality requirements is as follows: I. Establish a three-tiered basic threshold benchmark system: Based on the core control objectives of the production line (quality first / quality and efficiency balance / efficiency first), three fixed-level basic thresholds are preset, and the corresponding quality control indicators and applicable scenarios for each level are clearly defined: Strict Threshold (Quality Priority): Confidence level ≥ 99.5%, control indicator is actual defect false negative rate ≤ 0.1%, suitable for high-end customer customized products, new product trial production stage, and filtering of high-risk defect types; Standard threshold (balanced mode): Confidence level ≥ 99%, control indicators are actual defect missed rate ≤ 0.3% and manual review rate ≤ 5%, which is the system default threshold and is suitable for standard products in the mass production stage; Relaxed threshold (efficiency priority): Confidence level ≥ 98.5%, control indicator is manual review rate ≤ 2%, applicable to general consumer products, production line capacity is tight and there are no quality abnormalities for 7 consecutive days.

[0142] II. Determining the core basis and priority for threshold adjustment: The threshold adjustment comprehensively considers the following four dimensions of production line data, sorted from highest to lowest priority: 1. Customer quality requirements and product grade (highest priority): High-end customer orders are forcibly locked at the strict threshold and cannot be lowered; standard customer orders use the standard threshold by default; general customer orders can be switched to the lenient threshold according to the production line status. 2. Production line production stage: Strict thresholds are mandatory during new product introduction and small-batch trial production; standard thresholds are used during mass production stabilization. 3. Real-time quality status of the production line: If the production line yield is ≥99.8% for 7 consecutive days and there are no batch defect incidents, the threshold can be lowered to the lenient level; if there is 1 or more customer quality complaints or production line batch yield incidents, the threshold will be immediately raised to the strict level. 4. Actual model performance: Based on the manual review data of the past 7 days, if the actual defect miss rate exceeds the upper limit of the corresponding level of control, the threshold will be increased by 0.2 percentage points; if the manual review rate exceeds the upper limit of the corresponding level of control, the threshold will be decreased by 0.2 percentage points.

[0143] III. Set the triggering conditions and constraint rules for threshold adjustment: Regular adjustments: Threshold calibration is automatically performed once a day at midnight, and the effective threshold is updated based on the production line data of the previous 7 days; Event-triggered adjustment: The threshold will be adjusted immediately when any of the following events occur: switching to a mask for production of different models / substrates, completing an incremental training of an artificial intelligence model, or a quality complaint or batch yield incident occurs. Constraints: The single threshold adjustment range shall not exceed ±0.5 percentage points, and the final threshold shall not be lower than the lenient threshold or higher than the strict threshold; all threshold adjustment records shall be fully stored in the production database for traceability.

[0144] IV. Threshold Validation and Rollback Mechanism: All automatically adjusted thresholds must first be verified in a small batch on no fewer than 50 continuously produced mask sheets. If the verification results show that the false negative rate or false positive rate exceeds the control range of the corresponding level, the thresholds will be automatically rolled back to the stable threshold of the previous version and an alarm will be sent to the quality inspectors.

[0145] Step six: The fake defect filtering module automatically filters and removes the fake defect sample data, excluding it from the subsequent defect review and processing flow, and simultaneously stores the filtering results in the production database.

[0146] Step 7: If there are still unprocessed defect samples in the current batch of masks, return to step 2 and process the next defect sample to be processed in a loop; if all defects in the current batch of masks have been processed, proceed to the next step.

[0147] Step 8: The sample iteration update module collects the defect judgment results that have been manually reviewed and corrected within this batch, as well as the misjudged / missed samples reported by the production line. After the quality inspection engineer completes the labeling and verification, the samples are added to the training sample set. This triggers the artificial intelligence model training module to perform incremental fine-tuning training with a small learning rate, update and optimize the model weights, and replace the original fixed model after verification. This completes the model iteration and ends the current process.

[0148] Optionally, the incremental training in step eight may specifically include: I. Manually label and verify the collected review and correction samples and quality feedback samples to confirm their classification labels for real defects / false defects; II. Add the newly labeled samples to the original training dataset, and re-divide the training and validation sets according to the updated dataset; III. Based on the original fixed model weight file, incremental fine-tuning training is carried out with a small learning rate to update the model weights and improve the model's adaptability to new sample scenarios. IV. Verify the accuracy of the incrementally trained model. Once the model meets the production line threshold requirements, replace the original fixed model and complete the model iteration update.

[0149] The image acquisition module acquires video images from the video access module, the image comparison module compares two images acquired by the image acquisition module before and after acquisition, and the image saving module saves the changed images to the database of the data server.

[0150] It should be noted that, in addition to the contents described above, this embodiment may also include the technical features described in the above embodiments, thereby achieving the technical effect of the defect detection method for the metal mask shown above. Please refer to the above description for details. For the sake of brevity, it will not be elaborated here.

[0151] Based on the embodiments of this disclosure, an artificial intelligence deep learning model is used to achieve automatic identification and filtering of false defects detected by AOI in OLED metal masks. This eliminates the need for manual review of numerous false defects, removing over 99% of invalid false defect data, significantly reducing quality inspection labor costs, shortening the mask production inspection cycle, and significantly improving production efficiency. It also avoids the workload of repetitive manual operations, achieving fully automated operation of the inspection process. Furthermore, standardized artificial intelligence model reasoning is used to determine whether a defect is genuine or false, completely eliminating subjective judgment bias from manual review, unifying defect judgment standards, and significantly reducing the missed detection rate of genuine defects and the false detection rate of false defects. This ensures the quality of the metal masks leaving the factory, avoids batch yield accidents in downstream OLED production, reduces the problem of erroneous scrapping of qualified masks, and lowers production losses and overall costs. Through a sample iteration update module and incremental training process, the AI ​​model can be continuously optimized based on real-time verification data and quality feedback data generated on the production line. The model's false defect filtering capability can be autonomously iterated and improved. It can adapt to various metal masks such as Open Masks and CVD Masks with different models, mesh parameters, and substrates, matching the rapid iteration production needs of the OLED industry. Its generalization ability and scenario adaptability are far superior to traditional fixed-rule machine vision algorithms. Through targeted sample preprocessing and lightweight model architecture optimization, the inference speed and classification accuracy of the AI ​​model are guaranteed. It can seamlessly connect with the output rhythm of existing AOI inspection equipment without changing the core inspection process of the existing production line. The equipment modification cost is low and it is easy to deploy and apply at scale.

[0152] Figure 4 This is a schematic diagram of a defect detection device for a metal mask provided in an embodiment of the present disclosure. The defect detection device for the metal mask includes: The first acquisition unit 401 is configured to acquire a target image, wherein the target image represents a metal mask after being processed by the wire mesh stretching process; The detection unit 402 is configured to: perform defect detection on the target image via an automatic optical inspection device to obtain the image region in the target image where a defect exists; The first determining unit 403 is configured to: determine the target defect type of the image region and the confidence level corresponding to the target defect type through a pre-trained artificial intelligence model, wherein the target defect type represents a real defect or a false defect; The second determining unit 404 is configured to: determine whether to perform defect verification on the image region based on the target defect type and confidence level; The third determining unit 405 is configured to: if it is determined that a defect verification of the image region is to be performed, then determine the verification result of the image region and obtain the target detection result of the image region.

[0153] In some possible implementations, the artificial intelligence model is trained in the following manner: Obtain a training sample set; wherein the training samples in the training sample set include sample images and sample labels; the sample images include image regions with defects as determined by an automated optical inspection device; the sample labels represent real defects or false defects; real defects include at least one of the following: hole defects, edge burrs, scratches, foreign matter residue, and mesh deformation; false defects include at least one of the following: reflection artifacts, dust adhesion, inherent texture of the substrate, imaging noise, and non-functional traces; The artificial intelligence model is trained by using machine learning algorithms, taking the sample images included in the training sample set as input data and the sample labels included in the training sample set as the expected output data.

[0154] In some possible implementations, the sample images included in the training sample set are generated in the following manner: The second acquisition unit (not shown in the figure) is configured to: acquire the image region with defects as determined by the automatic optical inspection equipment, and obtain a defect image; The first processing unit (not shown in the figure) is configured to: preprocess the defect image to obtain the processed image, wherein the preprocessing includes at least one of the following: size normalization processing, gray value calibration, and noise reduction processing; The second processing unit (not shown in the figure) is configured to perform augmentation processing on the processed image to obtain a sample image, wherein the augmentation processing includes at least one of the following: flipping, rotating, brightness adjustment, contrast adjustment, cropping, and stitching.

[0155] In some possible implementations, after determining the verification result of the image region and obtaining the target detection result of the image region if it is determined to perform defect verification on the image region, the apparatus further includes: The third acquisition unit (not shown in the figure) is configured to acquire a quality feedback image of the metal mask and a label for the quality feedback image. The first generation unit (not shown in the figure) is configured to generate the first training sample based on the quality feedback image and the label of the quality feedback image; The fourth determining unit (not shown in the figure) is configured to: determine the label of the image region if the verification result does not match the target defect type; The second generation unit (not shown in the figure) is configured to generate a second training sample based on the image region and the label of the image region. The training unit (not shown in the figure) is configured to incrementally train the pre-trained artificial intelligence model using the first training sample and the second training sample to obtain the incrementally trained artificial intelligence model.

[0156] In some possible implementations, the learning rate during incremental training is smaller than the learning rate during pre-training.

[0157] In some possible implementations, the artificial intelligence model has a convolutional neural network architecture; the loss function of the artificial intelligence model includes the cross-entropy loss function.

[0158] In some possible implementations, determining whether to perform defect verification on the image region based on the target defect type and confidence level includes: If the target defect type represents a real defect, or if the confidence level is less than or equal to a preset threshold, then the image region is determined to undergo defect verification. If the target defect type indicates a false defect and the confidence level is greater than a preset threshold, then it is determined that no defect verification will be performed on the image region.

[0159] The metal mask defect detection device provided in this embodiment can execute the corresponding steps of the above-described metal mask defect detection methods, thereby achieving the technical effects of the above-described metal mask defect detection methods. The metal mask defect detection device and the metal mask defect detection methods can refer to and cite each other in terms of specific implementation and technical effects. For the sake of brevity, they will not be elaborated here.

[0160] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present disclosure. Figure 5 The illustrated electronic device 500 includes at least one processor 501, a memory 502, at least one network interface 504, and other user interfaces 503. The various components in the electronic device 500 are coupled together via a bus system 505. It is understood that the bus system 505 is used to implement communication between these components. In addition to a data bus, the bus system 505 also includes a power bus, a control bus, and a status signal bus. However, for clarity, ... Figure 5 The general designated all buses as Bus System 505.

[0161] The user interface 503 may include a display, keyboard, or clicking device (e.g., mouse, trackball, touchpad, or touchscreen).

[0162] It is understood that the memory 502 in this embodiment can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDRSDRAM), Enhanced Synchronous DRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), and Direct Rambus RAM (DRRAM). The memory 502 described herein is intended to include, but is not limited to, these and any other suitable types of memory.

[0163] In some implementations, memory 502 stores elements, executable units or data structures, or subsets thereof, or extended sets thereof: operating system 5021 and application program 5022.

[0164] The operating system 5021 includes various system programs, such as the framework layer, core library layer, and driver layer, used to implement various basic business functions and handle hardware-based tasks. The application program 5022 includes various applications, such as a media player and a browser, used to implement various application functions. The program implementing the method of this embodiment can be included in the application program 5022.

[0165] In this embodiment, by calling the program or instructions stored in memory 502, specifically the program or instructions stored in application program 5022, processor 501 executes the method steps provided in each method embodiment, including, for example: Acquire the target image, where the target image represents the metal mask after the wire mesh stretching process; The defect detection of the target image is performed by an automated optical inspection device to obtain the image region in the target image where defects exist; By using a pre-trained artificial intelligence model, the target defect type of the image region and the confidence level corresponding to the target defect type are determined, where the target defect type represents a real defect or a false defect. Based on the target defect type and confidence level, determine whether to perform defect verification on the image region; If it is determined that a defect verification will be performed on the image region, then the verification result of the image region will be determined, and the target detection result of the image region will be obtained.

[0166] The methods disclosed in the above embodiments of this disclosure can be applied to or implemented by processor 501. Processor 501 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the integrated logic circuit of the hardware or by instructions in the form of software in processor 501. The processor 501 may be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this disclosure. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this disclosure can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software units in the decoding processor. The software units may be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in memory 502. Processor 501 reads the information in memory 502 and, in conjunction with its hardware, completes the steps of the above method.

[0167] It is understood that the embodiments described herein can be implemented in hardware, software, firmware, middleware, microcode, or a combination thereof. For hardware implementation, the processing unit can be implemented in one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), general-purpose processors, controllers, microcontrollers, microprocessors, other electronic units for performing the functions described above, or combinations thereof.

[0168] For software implementation, the techniques described herein can be implemented by units that perform the functions described above. The software code can be stored in memory and executed by a processor. The memory can be implemented within the processor or external to the processor.

[0169] The electronic device provided in this embodiment may be as follows: Figure 5 The electronic device shown can perform all the steps of the defect detection methods for the metal masks described above, thereby achieving the technical effects of the defect detection methods for the metal masks described above. For details, please refer to the relevant descriptions above. For the sake of brevity, it will not be elaborated here.

[0170] This disclosure also provides a storage medium (computer-readable storage medium). This storage medium stores one or more programs. The storage medium may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as read-only memory, flash memory, hard disk, or solid-state drive; the memory may also include combinations of the above types of memory.

[0171] When one or more programs in the storage medium can be executed by one or more processors to implement the above-described defect detection method for metal masks executed on the electronic device side.

[0172] The processor described above is used to execute a defect detection program for a metal mask stored in memory to implement the following steps of a defect detection method for a metal mask executed on the electronic device side: Acquire the target image, where the target image represents the metal mask after the wire mesh stretching process; The defect detection of the target image is performed by an automated optical inspection device to obtain the image region in the target image where defects exist; By using a pre-trained artificial intelligence model, the target defect type of the image region and the confidence level corresponding to the target defect type are determined, where the target defect type represents a real defect or a false defect. Based on the target defect type and confidence level, determine whether to perform defect verification on the image region; If it is determined that a defect verification will be performed on the image region, then the verification result of the image region will be determined, and the target detection result of the image region will be obtained.

[0173] Furthermore, the computer program product provided in this disclosure embodiment may include computer-readable code that, when executed on a device, causes a processor in the device to implement the steps of a defect detection method for a metal mask executed on the electronic device side: Acquire the target image, where the target image represents the metal mask after the wire mesh stretching process; The defect detection of the target image is performed by an automated optical inspection device to obtain the image region in the target image where defects exist; By using a pre-trained artificial intelligence model, the target defect type of the image region and the confidence level corresponding to the target defect type are determined, where the target defect type represents a real defect or a false defect. Based on the target defect type and confidence level, determine whether to perform defect verification on the image region; If it is determined that a defect verification will be performed on the image region, then the verification result of the image region will be determined, and the target detection result of the image region will be obtained.

[0174] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this disclosure.

[0175] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented in hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0176] It should be understood that the terminology used herein is for the purpose of describing particular exemplary embodiments only and is not intended to be limiting. Unless the context clearly indicates otherwise, the singular forms “a,” “an,” and “described” as used herein may also include the plural forms. The terms “comprising,” “including,” “containing,” and “having” are inclusive and therefore indicate the presence of the stated features, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, elements, components, and / or combinations thereof. The method steps, processes, and operations described herein are not construed as requiring them to be performed in a particular order described or illustrated unless the order of performance is explicitly indicated. It should also be understood that additional or alternative steps may be used.

[0177] The above description is merely a specific embodiment of this disclosure, enabling those skilled in the art to understand or implement it. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this disclosure. Therefore, this disclosure is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A method for defect detection in a metal photomask, characterized in that, include: Acquire a target image, wherein the target image represents a metal mask after the wire mesh stretching process; The target image is subjected to defect detection using an automated optical inspection device to obtain the image region in the target image where defects exist; The target defect type and the confidence level corresponding to the target defect type are determined by a pre-trained artificial intelligence model, wherein the target defect type represents a real defect or a false defect. Based on the target defect type and the confidence level, determine whether to perform defect verification on the image region; If it is determined that the image region will be subject to defect verification, the verification result of the image region will be determined, and the target detection result of the image region will be obtained.

2. The method according to claim 1, characterized in that, The artificial intelligence model was trained using the following method: Obtain a training sample set; wherein the training samples in the training sample set include sample images and sample labels; the sample images include image regions with defects as determined by an automated optical inspection device; the sample labels represent real defects or false defects; real defects include at least one of the following: hole defects, edge burrs, scratches, foreign matter residue, and mesh deformation; false defects include at least one of the following: reflection artifacts, dust adhesion, inherent texture of the substrate, imaging noise, and non-functional traces; An artificial intelligence model is trained by using machine learning algorithms, taking the sample images included in the training sample set as input data and the sample labels included in the training sample set as expected output data.

3. The method according to claim 2, characterized in that, The training sample images included in the training sample set are generated in the following manner: The defective image is obtained by acquiring the image area with defects as determined by an automated optical inspection device. The defective image is preprocessed to obtain a processed image, wherein the preprocessing includes at least one of the following: size normalization, grayscale calibration, and noise reduction. The processed image is augmented to obtain a sample image, wherein the augmentation process includes at least one of the following: flipping, rotating, brightness adjustment, contrast adjustment, cropping, and stitching.

4. The method according to claim 1, characterized in that, After determining the verification result of the image region and obtaining the target detection result of the image region if it is determined to perform defect verification on the image region, the method further includes: Obtain a quality feedback image of the metal mask, and a label for the quality feedback image; Based on the quality feedback image and the label of the quality feedback image, a first training sample is generated; If the verification result does not match the target defect type, then the label of the image region is determined; A second training sample is generated based on the image region and the label of the image region; Using the first training sample and the second training sample, the pre-trained artificial intelligence model is incrementally trained to obtain the incrementally trained artificial intelligence model.

5. The method according to claim 4, characterized in that, The learning rate during the incremental training process is less than the learning rate during the pre-training process.

6. The method according to claim 1, characterized in that, The artificial intelligence model has a convolutional neural network architecture; the loss function of the artificial intelligence model includes the cross-entropy loss function.

7. The method according to any one of claims 1-6, characterized in that, The step of determining whether to perform defect verification on the image region based on the target defect type and the confidence level includes: If the target defect type represents a real defect, or if the confidence level is less than or equal to a preset threshold, then it is determined that the image region will be subject to defect verification. If the target defect type indicates a false defect, and the confidence level is greater than the preset threshold, then it is determined that the image region will not be subject to defect verification.

8. A defect detection device for a metal mask, characterized in that, include: The first acquisition unit is configured to acquire a target image, wherein the target image represents a metal mask after being processed by the wire mesh stretching process; The detection unit is configured to: perform defect detection on the target image via an automatic optical inspection device to obtain image regions in the target image where defects exist; The first determining unit is configured to: determine the target defect type of the image region and the confidence level corresponding to the target defect type through a pre-trained artificial intelligence model, wherein the target defect type represents a real defect or a false defect; The second determining unit is configured to: determine whether to perform defect verification on the image region based on the target defect type and the confidence level; The third determining unit is configured to: if it is determined that the image region will be subject to defect verification, determine the verification result of the image region and obtain the target detection result of the image region.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing a computer program stored in the memory, wherein when the computer program is executed, it implements the method described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1-7.

Images