A ceramic substrate defect intelligent detection method and system

Abstract

The present invention relates to the field of image processing technology, and in particular to a method and system for intelligent detection of ceramic substrate defects, the method comprising: obtaining an image of a ceramic substrate to be detected; inputting the image into a trained target detection model, and outputting the defect type and defect position; the training process comprises: obtaining images of multiple defective ceramic substrates, inputting the training set into the target detection model for training to obtain a trained target detection model; during the training process, determining the weight of the features of each network layer according to the similarity between the circumscribed rectangle of a defect area in the ceramic substrate image and each network layer, and performing feature fusion according to the weight. The present invention improves the accuracy of ceramic substrate defect detection.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology and to a method and system for intelligently detecting defects on a ceramic substrate. Background Art

[0002] Ceramic substrate is an ultra-thin composite substrate with large current carrying capacity, and is the basic material for high-power power electronic circuit structure. In the production process of ceramic substrates, defects such as gold plating damage, excess gold on the edge, ceramic deficiency, oil stains, and foreign matter often occur. Quickly and accurately detecting these defective ceramic substrates is an important link in ensuring production quality.

[0003] Moreover, timely detection of defects can reduce scrap rates and reduce rework and scrap costs, thereby improving production efficiency and reducing overall costs.

[0004] Furthermore, defects may cause ceramic substrates to fail in extreme environments such as high temperature and high pressure. Detecting defects helps improve the reliability and stability of ceramic substrates in practical applications.

[0005] Relevant prior art, such as the patent application document with application publication number CN118429327A, is entitled A method and system for surface defect detection of ceramic substrates based on deep learning, comprising: obtaining a surface image of the substrate to be tested, inputting the surface image into a detection model obtained through training for processing, and obtaining a detection result of the surface image; the detection model is trained in the following manner; obtaining training samples, preprocessing each sample image, and obtaining a preprocessed grayscale image; segmenting the substrate area and background area of ​​the grayscale image; segmenting the defect area and other areas of the first image; and inputting the second image into a preset model based on deep learning for training, and obtaining a trained detection model.

[0006] Considering the variable shapes of ceramic substrate defects, if the above method is used to obtain the defect area only by image segmentation and then train the model, the accuracy of ceramic substrate defect detection may be low. Based on this, how to improve the accuracy of ceramic substrate defect detection is particularly important. Summary of the invention

[0007] The present invention proposes a ceramic substrate defect intelligent detection method and system to improve the accuracy of detecting defects of the ceramic substrate. To this end, the present invention provides solutions in the following aspects.

[0008] In a first aspect of an embodiment of the present invention, a method for intelligently detecting defects of a ceramic substrate is provided, comprising:

[0009] Acquire an image of a ceramic substrate to be inspected;

[0010] Inputting the image into a trained target detection model to output the defect type and defect location of the ceramic substrate; the target detection model adopts a convolutional neural network;

[0011] The training process includes: fusing the features extracted from multiple network layers. The feature fusion formula is: , is the fused feature, m is the total number of network layers, It is used to characterize the similarity between the bounding rectangle of the t-th defect area in the ceramic substrate image and the s-th network layer. is the feature extracted by the sth network layer.

[0012] The present invention determines the similarity between the bounding rectangle of the defect area and the candidate frame of the sth network layer, and then determines the weight of the feature corresponding to the sth network layer according to the size of the similarity, wherein the closer the size and shape of the bounding rectangle of the defect area and the candidate frame of the sth network layer are, the greater the similarity and the greater the weight. The greater the similarity, the more the features extracted from the candidate frame of the sth network layer can reflect the defect type. Based on this, the accuracy of ceramic substrate defect detection can be improved through the above-mentioned feature fusion method.

[0013] Optionally, the The calculation formula is:

[0014] , where n is the total number of candidate boxes of the s-th network layer, is the height of the circumscribed rectangle of the t-th defect area, is the width of the circumscribed rectangle of the t-th defect area, is the height of the kth candidate box of the sth network layer, is the width of the kth candidate box of the sth network layer, is the maximum and minimum normalized function of negative correlation.

[0015] The larger the value is, the greater the difference between the bounding rectangle of the t-th defect area and the candidate box of the s-th network layer is, and the weight of the feature corresponding to the s-th network layer should be smaller, so that the fused feature is more accurate.

[0016] Optionally, the The calculation formula is:

[0017] , where n is the total number of candidate boxes of the s-th network layer, is the height of the circumscribed rectangle of the t-th defect area, is the width of the circumscribed rectangle of the t-th defect area, is the height of the kth candidate box of the sth network layer, is the width of the kth candidate box of the sth network layer, is the absolute value, is the maximum and minimum normalized function of negative correlation.

[0018] The larger the value is, the greater the difference between the bounding rectangle of the t-th defect area and the candidate box of the s-th network layer is, and the weight of the feature corresponding to the s-th network layer should be smaller, so that the fused feature is more accurate.

[0019] Optionally, the defect area acquisition process includes:

[0020] For a pixel point of a ceramic substrate image, determine the pixel points in the same position as the pixel point in the image in multiple images of defect-free ceramic substrates, calculate their means and variances, and then obtain a Gaussian model, input the grayscale of the pixel point into the Gaussian model to obtain the probability that the pixel point is a background pixel point, and in response to the probability being less than a set threshold, the pixel point is a pixel point in a defective area;

[0021] Then, it is determined whether other pixels in the ceramic substrate image are pixels in the defective area, and the defective area is obtained.

[0022] Since the grayscales of pixels in the images of multiple non-defective ceramic substrates that are located at the same position as the one pixel in the images conform to a normal distribution, the smaller the probability is, the greater the possibility that the one pixel is a pixel in a defective area.

[0023] Optionally, the defect area acquisition process includes:

[0024] For a pixel point of the ceramic substrate image, determine the pixel points in the same position as the pixel point in the image in multiple images of defect-free ceramic substrates, calculate their mean and standard deviation, if the grayscale of the pixel point satisfies: or , where r is the grayscale of the pixel, is the mean value, is the coefficient, greater than 0, is the standard deviation, then the one pixel is a pixel in the defect area;

[0025] Then, it is determined whether other pixels in the ceramic substrate image are pixels in the defective area, and the defective area is obtained.

[0026] When the grayscale of a pixel satisfies: or , it is highly likely that the one pixel is a pixel in the defective area.

[0027] Optionally, the loss function of the target detection model is a cross entropy loss function.

[0028] Optionally, the target detection model is a YOLOV5 model.

[0029] In a second aspect of an embodiment of the present invention, a ceramic substrate defect intelligent detection system is provided, comprising: a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, a ceramic substrate defect intelligent detection method according to any one of the first aspects of the embodiment of the present invention is implemented.

[0030] Beneficial effects of the present invention:

[0031] The present invention first determines the bounding rectangle of the defect area in the ceramic substrate image, and for a bounding rectangle of a defect area, calculates the similarity between the bounding rectangle and the candidate box of each layer of the network, and then determines the weight of the features extracted by each layer of the network, and then performs feature fusion according to the weight. For the s-th layer of the network, the greater the similarity between the bounding rectangle and the candidate box of the s-th layer of the network, that is, the smaller the difference between the bounding rectangle and the candidate box of the s-th layer of the network, the more the features extracted by the s-th layer of the network can reflect the defect type, and the weight of the features extracted by the s-th layer of the network should be greater, so as to improve the accuracy of ceramic substrate defect detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] By reading the following detailed description with reference to the accompanying drawings, the above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily understood. In the accompanying drawings, several embodiments of the present invention are shown in an exemplary and non-limiting manner, and the same or corresponding reference numerals represent the same or corresponding parts, wherein:

[0033] Figure 1 is a flow chart schematically showing a method for intelligently detecting defects of a ceramic substrate according to an embodiment of the present invention;

[0034] Figure 2 The figure is a schematic diagram schematically showing a structural block diagram of a ceramic substrate defect intelligent detection system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0035] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0036] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0037] Figure 1 FIG. 1 is a flow chart schematically showing a method for intelligently detecting defects of a ceramic substrate according to an embodiment of the present invention. Figure 1 The specific instructions are as follows:

[0038] S1: Train the target detection model to obtain a trained target detection model.

[0039] In one embodiment, the target detection model is a YOLOV5 model.

[0040] The YOLOV5 model is a convolutional neural network model (CNN model), which includes multiple convolution kernels. Each convolution kernel serves as a network layer. Each network layer extracts corresponding features. The features of multiple network layers are fused and then pooled and output.

[0041] In the present invention, the core improvement lies in the fusion of multiple network layers. Other technical details are common sense in deep learning technology. Therefore, the following focuses on the fusion process and the related training process.

[0042] The defects of ceramic substrates vary in shape and size, and there are many smaller ceramic defects corresponding to the defects of ceramic substrates. For defects with smaller areas, when using deep learning models to identify the defects, as the number of network layers increases, the receptive field of the features extracted by the network on the original image becomes larger and larger. A larger receptive field is more conducive to extracting the features of defects with larger areas, and thus more conducive to identifying the types of defects with larger areas, because the larger the area of ​​the defect, the greater the proportion of the defect in the original image. For defects with smaller areas, the features extracted by a larger receptive field contain more useless information, so it is not conducive to identifying the types of defects with smaller areas.

[0043] Therefore, the present invention attempts to improve the fusion method in order to improve the final recognition accuracy.

[0044] The following is a detailed description of the sample processing process of the training model:

[0045] During the model training process, the training set (collected grayscale images of substrates) is input into the target detection model to obtain a trained target detection model. The training set includes multiple samples, each of which includes an image of a defective ceramic substrate and a label corresponding to the image (used to mark the defect type and location).

[0046] 1.1: For example, taking a substrate image as an example (hereinafter referred to as the target substrate image), the acquisition process of its defect area (hereinafter referred to as the target defect area) includes: for pixel point A in the target substrate image, first obtain multiple images of non-defective ceramic substrates, then determine the pixels in the images of the multiple non-defective ceramic substrates that are in the same position as the pixel point A in the image, calculate their mean and variance, and then obtain a Gaussian model. The grayscale of the pixel point A is input into the Gaussian model to obtain the probability that the pixel point A is a background pixel point. If the probability is less than the set threshold, the pixel point A is a pixel point in the target defect area. If the probability is greater than or equal to the set threshold, the pixel point A is not a pixel point in the target defect area.

[0047] Based on this, by traversing the entire image, it is possible to determine which pixels belong to the target defect area. All pixels in the target defect area constitute the target defect area. At this point, the target defect area is obtained.

[0048] It should be noted that the defect-free ceramic substrate and the defective ceramic substrate are from the same batch of ceramic substrates, and the circuits on the same batch of ceramic substrates are distributed in the same manner. For the image of the defect-free ceramic substrate, the grayscale values ​​of the pixels at the same position conform to the normal distribution, or in other words, the grayscale values ​​of the pixels at the same position are basically equal. For example, the grayscale values ​​of the pixels on multiple defect-free ceramic substrate images satisfy the following distribution: ,in, represents the distribution of grayscale values ​​of pixels in the i-th row and j-th column of multiple defect-free ceramic substrate images, is the mean grayscale value of the pixel points in the i-th row and j-th column, is the variance of the grayscale value of the pixel in the i-th row and j-th column, Indicates the mean And the variance is The normal distribution of .

[0049] Those skilled in the art may set the threshold value by themselves, and the threshold value may be 0.85.

[0050] In another embodiment, the process of acquiring the target defect area includes: for a pixel point B in the target substrate image, determining the pixel points D1, D2, ... in the images of the plurality of defect-free ceramic substrates that are at the same position as the pixel point B in the image, and calculating the average of the pixel points D1, D2, ... and standard deviation , if the grayscale r of the pixel point B satisfies: or ,in, is the coefficient, If the pixel B is greater than 0, it is determined that the pixel B is a pixel in the target defect area. Then, it is determined whether other pixel points in the target substrate image are pixel points in the target defect area, and the target defect area is obtained.

[0051] Technical personnel in the field can set For example, 1.5 is recommended.

[0052] 1.2: In one embodiment, for the target substrate image, the pixel values ​​in the target defect area in the target substrate image are set to 1, and the other pixel values ​​are set to 0 to obtain a binary image. A morphological operation is performed on the binary image, such as first performing an expansion operation on the binary image and then performing an erosion operation to obtain a processed binary image. The processed binary image is multiplied with the target substrate image to obtain a foreground image. The bounding rectangle of the area of ​​pixels whose grayscale values ​​in the foreground image are not 0 is used as the bounding rectangle of the target defect area. In other words, a morphological operation is performed on the binary image to obtain an image Y. The bounding rectangle of the area whose pixel value in the image Y is 1 is used as the bounding rectangle of the target defect area.

[0053] At this point, the bounding rectangle of the target defect area is obtained.

[0054] 1.3: Next, the fused features corresponding to the target substrate image are calculated.

[0055] In one embodiment, the feature fusion formula is:

[0056] .

[0057] .in, is the fused feature, m is the total number of network layers, It is used to characterize the similarity between the bounding rectangle of the t-th defect area in the ceramic substrate image and the s-th network layer. is the feature extracted by the sth network layer; n is the total number of candidate boxes of the sth network layer, is the height of the circumscribed rectangle of the t-th defect area, is the width of the circumscribed rectangle of the t-th defect area, is the height of the kth candidate box in the candidate box of the sth network layer, is the width of the kth candidate box in the candidate box of the sth network layer, is the maximum and minimum normalized function of negative correlation.

[0058] For example, Denoted as the difference index A1 of the t-th defect area, the difference index A1 of each defect area is calculated, the maximum value in the maximum-minimum normalized function is the maximum value of all difference indexes A1, and the minimum value in the maximum-minimum normalized function is the minimum value of all difference indexes A1.

[0059] In the above formula, s is used to represent the count of network layers. The maximum value of s is m, which means that a Yolov5 model has m network layers. k represents the count of candidate boxes in the sth layer, up to n. t represents the count of defect areas in the input image.

[0060] It should be noted that, for the target substrate image, the type of defects and the location of defects in the target substrate image can be predicted according to the fused features corresponding to the target substrate image.

[0061] When the target detection model uses the YOLOV5 model for training and prediction, if the size of the bounding rectangle of the defect area of ​​the ceramic substrate is closer to the size of the candidate box corresponding to the smaller object, the features corresponding to the shallow network and the middle network are mainly fused through weight distribution; if the size of the bounding rectangle of the defect area of ​​the ceramic substrate is closer to the candidate box corresponding to the medium object, the features corresponding to the middle network and the features corresponding to the deep network are mainly fused through weight distribution; if the size of the bounding rectangle of the defect area of ​​the ceramic substrate is closer to the candidate box corresponding to the larger object, the features of the deep network are mainly fused through weight distribution.

[0062] It should be noted that in the object detection algorithm in deep learning, the candidate box sizes of shallow networks and deep networks are different. The candidate boxes of shallow networks are relatively small, while the candidate boxes of deep networks are relatively large. Moreover, the difference in the sizes of candidate boxes of the same layer of network is small.

[0063] In summary, the features of defects with smaller areas extracted by the shallow network are more accurate, and the features of defects with larger areas extracted by the deep network are more accurate. Therefore, in the process of feature fusion, weights are set for the features corresponding to the shallow network, the middle network, and the deep network according to the size of the defect. The larger the area of ​​the defect area, the greater the weight of the feature corresponding to the deep network should be.

[0064] Further explanation: The bigger, The smaller, The smaller.

[0065] The larger it is, the greater the difference between the bounding rectangle of the t-th defect area and the candidate box of the s-th network layer, and the smaller the weight of the feature corresponding to the s-th network layer should be, so that the fused feature is more accurate. The reason is that the closer the bounding rectangle of the defect area is to the size of the candidate box of a certain layer of network, the more the features extracted by the certain layer of network match the features of the defect area, and the higher the accuracy of using the features of the certain layer of network to identify the type of defect. The closer the size of the defect area is to the size of the candidate box of a certain layer, the greater the possibility that the defect can be identified by this layer of network, and the greater the weight of the feature corresponding to this layer of network should be when the features are fused.

[0066] In combination with the concept of the present invention, some other variant embodiments are provided: In one embodiment, the The calculation formula is:

[0067] ,in, is the maximum and minimum normalized function of negative correlation.

[0068] In another embodiment, the The calculation formula is:

[0069] ,in, is the natural exponential function, is an absolute value.

[0070] In one embodiment, the loss function of the target detection model is a cross entropy loss function.

[0071] In one embodiment, in response to the number of training times of the target detection model reaching the preset number of training times, the training is stopped. Those skilled in the art may set the preset number of training times. The preset number of training times may be 1000. It should be noted that since too few samples may result in too large a value of the loss function, the model should be trained based on more samples.

[0072] In one embodiment, when the loss is less than a preset loss threshold, the training is stopped. Those skilled in the art may set the preset loss threshold. The smaller the preset loss threshold, the greater the possibility that the accuracy of defect detection by the trained target detection model is higher.

[0073] In one embodiment, a plurality of trained target detection models are obtained by training the target detection model. The target detection model with the highest defect detection accuracy is used as the target detection model finally used for defect detection. For example, a trained target detection model can be obtained based on 1,000 images of defective ceramic substrates; another trained target detection model can be obtained based on another 1,000 images of defective ceramic substrates, and then the one with the highest defect detection accuracy is selected.

[0074] At this point, the training of the target detection model has been completed, and a trained target detection model has been obtained. Other contents related to the CNN model training are irrelevant to this case and will not be described in detail.

[0075] S2: Detect defects of ceramic substrates based on the trained target detection model.

[0076] Obtain an image of a ceramic substrate to be inspected. For any image of a ceramic substrate to be inspected, input the image into a trained target detection model, and output the defect type and defect location of the ceramic substrate to be inspected. The defect type includes no defect. That is, when the ceramic substrate to be inspected has no defect, output that the ceramic substrate has no defect.

[0077] It should be noted that, for any image of the ceramic substrate to be inspected, if the ceramic substrate is determined to have no defective area according to the above-mentioned method for obtaining the defective area, then the output is that the ceramic substrate has no defects; in other words, if the foreground image corresponding to the image of the ceramic substrate is blank, then the output is that the ceramic substrate has no defects.

[0078] For any image of the ceramic substrate to be inspected, in the process of predicting defects, or in other words, in the process of performing defect detection, the fused features corresponding to the image are obtained according to the above-mentioned method of obtaining fused features. Since the principle is the same, it will not be repeated here. In general, the defect area of ​​the image is first determined, and then the bounding rectangle of the defect area is determined, and then the difference between the size of the bounding rectangle and the size of the candidate frame of each layer is calculated. The weight of the features corresponding to each layer is determined based on the difference, and then the features are fused according to the weight of each feature to obtain the fused features. Finally, the type of defect and the location of the defect are obtained based on the fused features. Among them, the greater the difference between the size of the bounding rectangle and the size of the candidate frame of a certain layer, the smaller the weight of the feature corresponding to the network of this layer.

[0079] Similarly, by traversing other ceramic substrates to be inspected, the defect type and defect position of each ceramic substrate to be inspected can be obtained.

[0080] At this point, the defect detection of the ceramic substrate is completed.

[0081] Figure 2 FIG. 1 is a schematic diagram schematically showing a structural block diagram of a ceramic substrate defect intelligent detection system according to an embodiment of the present invention. Figure 2 As shown, 01 is a communication bus. The present invention also provides a ceramic substrate defect intelligent detection system, the system includes a processor and a memory, the memory stores computer program instructions, when the computer program instructions are executed by the processor, a ceramic substrate defect intelligent detection method according to the present invention is implemented.

[0082] The system also includes other components familiar to those skilled in the art, such as a communication bus and a communication interface, whose configuration and functions are known in the art and thus will not be described in detail here.

[0083] In the present invention, the aforementioned memory may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, apparatus or device. For example, a computer-readable storage medium may be any appropriate magnetic storage medium or magneto-optical storage medium, such as a resistive random access memory RRAM (Resistive Random Access Memory), a dynamic random access memory DRAM (Dynamic Random Access Memory), a static random access memory SRAM (Static Random-Access Memory), an enhanced dynamic random access memory EDRAM (Enhanced Dynamic Random Access Memory), a high-bandwidth memory HBM (High-Bandwidth Memory), a hybrid memory cube HMC (Hybrid Memory Cube), etc., or any other medium that can be used to store the required information and can be accessed by an application, a module, or both. Any such computer storage medium may be part of a device or accessible or connectable to a device. Any application or module described in the present invention may be implemented using computer-readable / executable instructions that may be stored or otherwise maintained by such a computer-readable medium.

[0084] In the description of this specification, "plurality" or "several" means at least two, such as two, three or more, etc., unless otherwise clearly and specifically defined.

[0085] Additionally, the articles "a" and "an" as used in this disclosure and the appended claims are generally construed to mean "one or more" unless specified otherwise or clear from context to be directed to a singular form.

[0086] Where the terms "includes," "having," "having," "having," or variations thereof are used in the detailed description or claims, such terms are intended to be inclusive in a manner similar to the term "comprising."

[0087] In some of the processes described in the specification and claims of the present invention and the above-mentioned figures, multiple operations appearing in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this article or executed in parallel. The serial numbers of the operations, such as S1, S2, etc., are only used to distinguish different operations, and the serial numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that if descriptions such as "first", "second", "A1", "B1" appear in this article, their role is to distinguish different messages, devices, modules, etc., and do not represent the order of precedence, do not have any technical meaning, and do not limit "A1", "B1", "first" and "second" to different types. For example: the first feature of an image is the feature of the image. The role of adding "first" includes facilitating subsequent references to different objects, that is, facilitating indexing.

[0088] Although this specification has shown and described a number of embodiments of the present invention, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Those skilled in the art will conceive of many modifications, changes and alternatives without departing from the ideas and spirit of the present invention. It should be understood that in the practice of the present invention, various alternatives to the embodiments of the present invention described herein may be employed.

Claims

1. A ceramic substrate defect intelligent detection method, characterized in that: include: Acquire an image of a ceramic substrate to be inspected; Inputting the image into a trained object detection model, and outputting the defect type and defect location of the ceramic substrate; The target detection model adopts a convolutional neural network; The training process includes: fusing the features extracted from multiple network layers. The feature fusion formula is: , is the fused feature, m is the total number of network layers, It is used to characterize the similarity between the bounding rectangle of the t-th defect area in the ceramic substrate image and the s-th network layer. is the feature extracted by the sth network layer; Said The calculation formula is: ； Wherein, n is the total number of candidate boxes of the sth network layer, is the height of the circumscribed rectangle of the t-th defect area, is the width of the circumscribed rectangle of the t-th defect area, is the height of the kth candidate box of the sth network layer, is the width of the kth candidate box of the sth network layer, is the maximum and minimum normalized function of negative correlation.

2. The intelligent detection method for ceramic substrate defects according to claim 1, characterized in that: Said The calculation formula can also be replaced by: ； Wherein, n is the total number of candidate boxes of the sth network layer, is the height of the circumscribed rectangle of the t-th defect area, is the width of the circumscribed rectangle of the t-th defect area, is the height of the kth candidate box of the sth network layer, is the width of the kth candidate box of the sth network layer, is the absolute value, is the maximum and minimum normalized function of negative correlation.

3. The intelligent detection method for ceramic substrate defects according to claim 1, characterized in that: The process of obtaining defect areas includes: For a pixel point of a ceramic substrate image, determine the pixel points in the same position as the pixel point in the image in multiple images of defect-free ceramic substrates, calculate their means and variances, and then obtain a Gaussian model, input the grayscale of the pixel point into the Gaussian model to obtain the probability that the pixel point is a background pixel point, and in response to the probability being less than a set threshold, the pixel point is a pixel point in a defective area; Then, it is determined whether other pixels in the ceramic substrate image are pixels in the defective area, and the defective area is obtained.

4. The intelligent detection method for ceramic substrate defects according to claim 1, characterized in that: The process of obtaining defect areas includes: For a pixel point of the ceramic substrate image, determine the pixel points in the same position as the pixel point in the image in multiple images of defect-free ceramic substrates, calculate their mean and standard deviation, if the grayscale of the pixel point satisfies: or , where r is the grayscale of the pixel, is the mean value, is the coefficient, greater than 0, is the standard deviation, then the one pixel is a pixel in the defect area; Then, it is determined whether other pixels in the ceramic substrate image are pixels in the defective area, and the defective area is obtained.

5. The intelligent detection method for ceramic substrate defects according to claim 1, characterized in that: The loss function of the target detection model is a cross entropy loss function.

6. The intelligent detection method for ceramic substrate defects according to claim 1, characterized in that: The target detection model is the YOLOV5 model.

7. A ceramic substrate defect intelligent detection system, characterized in that: include: A processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, a ceramic substrate defect intelligent detection method according to any one of claims 1 to 6 is implemented.

Images