Ceramic substrate defect intelligent classification method based on central Gaussian kernel distribution

Abstract

The invention provides a ceramic substrate defect intelligent classification method based on central Gaussian kernel distribution. The method comprises the following steps: constructing a ceramic substrate defect data set; acquiring a defect cutting image and a Gaussian kernel distribution identification image corresponding to the defect cutting image; building and training a ceramic substrate defect intelligent classification network model; evaluating a ceramic substrate defect intelligent classification network model; and carrying out defect classification on the ceramic substrate defective products. According to the ceramic substrate defect intelligent classification method based on central Gaussian kernel distribution, defect images detected by a 2D visual detection algorithm are preprocessed, the ceramic substrate defect intelligent classification network model is established, and different Gaussian kernel distributions are used as defect image feature characterization, so that the defect classification accuracy is effectively improved, and the defect classification efficiency is improved. The detection rate of interference defects is reduced, and the quality safety of the ceramic substrate is improved.

Description

Technical Field

[0001] The invention relates to the technical field of ceramic substrate production, and in particular to an intelligent classification method for ceramic substrate defects based on central Gaussian kernel distribution. Background Art

[0002] PCB, DCB, AMB, DPC and other coated ceramic substrates are important carriers in the semiconductor industry. Their quality and safety reliability directly determines the stability of product operation. During the surface coating process of different base substrates, the interference of air and other uncontrollable factors affects the formation of ceramic substrate products. There is a certain probability of forming solder residue defects on the surface of the coated products, or short circuits between the coating modules due to the uneven corrosion process. In serious cases, it directly affects the stability and reliability of product cost operation. Based on the 2D visual matching detection algorithm, the structural and appearance defects of ceramic substrates can be effectively located and identified with high precision. However, the Blob feature analysis cannot perform secondary feature extraction for multiple defect forms with similar morphology (such as AMB ceramic substrate solder residue defects and interference defects such as impurities in the air, the defects are all blackened flaky areas) to distinguish between real defects and interference defects, resulting in serious over-inspection of solder residue defects.

[0003] Therefore, it is necessary to design an intelligent classification method for ceramic substrate defects based on central Gaussian kernel distribution to solve the above problems. Summary of the invention

[0004] The purpose of the present invention is to provide a ceramic substrate defect intelligent classification method based on central Gaussian kernel distribution, which can effectively distinguish real defects from interference defects and reduce the over-inspection rate of ceramic substrate solder residual defects.

[0005] To achieve the above object, the present invention adopts the following technical solution: a ceramic substrate defect intelligent classification method based on central Gaussian kernel distribution, which comprises the following steps:

[0006] Step 1: construct a ceramic substrate defect dataset, wherein the ceramic substrate defect dataset includes ceramic substrate defect images of different types of defects, and the types of defects include n types;

[0007] Step 2: Obtain the defect cropped image and its corresponding Gaussian kernel distribution identification image;

[0008] Step 3, model construction: using the defect cropped image and Gaussian kernel distribution identification image as two inputs of the neural network model, providing a feature extraction module and a feature extraction supervision module to extract features from the defect cropped image and Gaussian kernel distribution identification image, providing a feature prediction classification module to predict the defect type and output the predicted probability value of the current defect being judged as different types;

[0009] Step 4, model training: the Gaussian kernel distribution identification image corresponding to the ceramic substrate defect image is divided into a training defect data set for model training and a verification defect data set for model verification according to the same preset ratio; the number of single calculation images, the loss function threshold and the maximum number of iterations are set for the training defect data set. When the model training reaches the set maximum number of iterations, or the model calculation loss value reaches the set loss function threshold, the weight information parameters and the network model framework in the current model training state are solidified and exported as a prediction model;

[0010] Step 5, model evaluation: manually mark the defect types of the images in the verification defect data set, call the prediction model for the marked verification defect data set, cyclically obtain the predicted defect types, record a correct prediction as an accurate prediction, record an incorrect prediction as an incorrect prediction, and statistically calculate the accuracy of the current model prediction; select the prediction model with the best evaluation result as the ceramic substrate defect intelligent classification network model;

[0011] Step 6: Apply the ceramic substrate defect intelligent classification network model to classify defects of ceramic substrate products.

[0012] As a further improved technical solution of the present invention, in step 1, more than 300 ceramic substrate defect images are obtained, and the defect types are 2, including OK type defects and NG type defects. The OK type defects include bright spot and white spot defects and erasable dirt defects, and the NG type defects are solder residue defects.

[0013] As a further improved technical solution of the present invention, in step 2, the defect image of the ceramic substrate is cropped according to a preset size to obtain a defect cropped image, and the center point of the defect cropped image is the center point of the defect area; the defect cropped image is marked with a central Gaussian distribution kernel to obtain a Gaussian kernel distribution mark image of all defect cropped images;

[0014] The Gaussian distribution function formula is:

[0015]

[0016] Where (x, y) is the image pixel coordinate, which obeys the normal distribution of constants (σ1, σ2, μ1, μ2, ρ);

[0017] Set k×k (k=1,3,5,……) Gaussian distribution kernel to characterize the image center area in NG type defect image, σ1=σ2, μ1=μ2, and diffuse toward the edge along the radial direction with the center as the highest probability point; set k1×k2 (k1=1,3,5,……; k2=3,5,7,……) Gaussian distribution kernel to characterize the image center area of ​​OK type defect, and diffuse toward the edge along the major and minor axis directions with the center as the highest probability point.

[0018] As a further improved technical solution of the present invention, the ceramic substrate is an AMB ceramic substrate, the probability value corresponding to the center point of the Gaussian kernel is set to a maximum of 1.0, and the probability of the edge of the defective Gaussian kernel is set to 0.5; the ratio of the loss value calculation of the central Gaussian kernel identification area and the loss value calculation of the remaining areas is set to 5:1.

[0019] As a further improved technical solution of the present invention, the parameter ratio of the feature extraction weight variable of the central Gaussian kernel marked area to the feature extraction weight variable of the non-Gaussian kernel marked area is set to 7:3.

[0020] As a further improved technical solution of the present invention, in step 3, the feature extraction module is composed of three 7×7 convolutional layers and one 3×3 pooling layer, and the channel attention mechanism and the semantic attention mechanism are introduced as the feature extraction supervision module. The output from the first convolutional layer to the last convolutional layer is used as the action layer of the feature extraction supervision module, and the feature information of the defect cropped image and the Gaussian kernel distribution identification image is extracted at the same time; the feature prediction classification module is composed of a 3×3 convolutional layer, a BN layer, a fully connected layer and a Softmax layer. The Softmax layer outputs the predicted probability value of the current defect being judged as different types, and the type index corresponding to the maximum probability is taken as the prediction result of the currently input ceramic substrate defect image.

[0021] As a further improved technical solution of the present invention, the Softmax layer outputs different types of prediction probability values, which are described by formula (II):

[0022]

[0023] Where i = 1, 2, 3, ... n corresponds to n types;

[0024] Set n=2, the Softmax layer operation outputs the probability prob1 that the ceramic substrate defect image is classified as an OK type defect and the probability prob2 that it is classified as an NG type defect, and prob1+prob2=1.0.

[0025] As a further improved technical solution of the present invention, in step 4, the Gaussian kernel distribution identification image corresponding to the ceramic substrate defect image is divided into a training defect data set for model training and a verification defect data set for model verification in a ratio of 4:1.

[0026] As a further improved technical solution of the present invention, in step 4, the number of training defect data sets is M, the number of verification defect data sets is N, the number of model training cycles is set to 10×M, the model calculation loss threshold is 1e-4, the number of images calculated at a single time is 24, the corresponding training defect image size is 256×256, the model training learning rate is initialized to 0.01, and the warm-up strategy and weight-decay strategy are added in the model training process.

[0027] As a further improved technical solution of the present invention, in step 6: a number of new sample finished product images are newly collected, and the defect area image is obtained after calling the 2D visual detection template matching detection algorithm, and the defect area image is directly input into the ceramic substrate defect intelligent classification network model, and the defects judged by the model as OK type defects or the defects whose predicted probability values ​​do not reach the preset detection threshold are automatically not displayed, and only the coordinate information of the defect area after intelligent classification is displayed, and then the defect quantification level evaluation is performed on the detected defects.

[0028] It can be seen from the above technical scheme that the intelligent classification method for ceramic substrate defects based on the central Gaussian kernel distribution of the present invention pre-processes the defect images detected by the 2D visual detection algorithm, and builds a ceramic substrate defect intelligent classification network model. Different Gaussian kernel distributions are used as defect image feature representations to effectively improve the accuracy of defect classification, reduce the detection rate of interference defects, and improve the quality and safety of ceramic substrates. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 The figure is a flow chart of a ceramic substrate defect intelligent classification method based on central Gaussian kernel distribution in one embodiment of the present invention.

[0030] Figure 2 It is a defect image of an OK type defect and a central Gaussian kernel distribution identification effect diagram in one embodiment of the present invention.

[0031] Figure 3 It is a defect image of an NG type defect and a central Gaussian kernel distribution identification effect diagram in one embodiment of the present invention.

[0032] Figure 4 Schematic diagram of a feature extraction module and a feature extraction supervision module in one embodiment of the present invention. DETAILED DESCRIPTION

[0033] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0034] The present invention provides a ceramic substrate defect intelligent classification method based on central Gaussian kernel distribution, please refer to Figure 1 As shown, it includes the following steps:

[0035] Step 1, constructing a ceramic substrate defect dataset, wherein the ceramic substrate defect dataset includes ceramic substrate defect images of different types of defects, and the types of defects include n types;

[0036] Step 2, obtaining a defect cropped image and its corresponding Gaussian kernel distribution identification image;

[0037] Step 3, model building: using the defect cropped image and the Gaussian kernel distribution identification image as two inputs of the neural network model, providing a feature extraction module and a feature extraction supervision module to extract features from the defect cropped image and the Gaussian kernel distribution identification image, providing a feature prediction classification module to predict the defect type and output the predicted probability value of the current defect being judged as different types, and taking the type index corresponding to the maximum probability as the prediction result of the currently input ceramic substrate defect image;

[0038] Step 4, model training: the Gaussian kernel distribution identification image corresponding to the ceramic substrate defect image is divided into a training defect data set for model training and a verification defect data set for model verification according to the same preset ratio; the number of single calculation images, the loss function threshold and the maximum number of iterations are set for the training defect data set. When the model training reaches the set maximum number of iterations, or the model calculation loss value reaches the set loss function threshold, the weight information parameters and the network model framework in the current model training state are solidified and exported as a prediction model;

[0039] Step 5, model evaluation: manually mark the defect types of the images in the verification defect data set, call the prediction model for the marked verification defect data set, cyclically obtain the predicted defect types, record a correct prediction as an accurate prediction, record an incorrect prediction as an incorrect prediction, and statistically calculate the accuracy of the current model prediction; select the prediction model with the best evaluation result as the ceramic substrate defect intelligent classification network model;

[0040] Step 6: Apply the ceramic substrate defect intelligent classification network model to classify the defects of the ceramic substrate products.

[0041] Specifically, this embodiment is illustrated by taking AMB ceramics as an example. In step 1, more than 300 defect images of AMB ceramic substrates are obtained. There are two types of defects, including OK type defects (i.e. interference defects, including bright spots and white spots defects and erasable dirt defects) and NG type defects (i.e. defects to be identified, which are solder residue defects in this embodiment, and may include scratches and other types of defects in other embodiments).

[0042] In step 2, the defect image of the ceramic substrate is cropped according to a preset size to obtain a defect cropped image, and the preset size is preferably 256×256. The center point of the defect cropped image is the center point of the defect area; the defect cropped image is marked with a central Gaussian distribution kernel to obtain a Gaussian kernel distribution mark image of all defect cropped images. The Gaussian distribution function formula is:

[0043]

[0044] Where (x, y) is the image pixel coordinate, which obeys the normal distribution of constants (σ1, σ2, μ1, μ2, ρ). Set k×k (k=1,3,5,……) Gaussian distribution kernel to characterize the image center area in NG type defect image. Gaussian distribution is a circular distribution (see Figure 3 As shown in the figure, σ1=σ2, μ1=μ2 is satisfied, and the center is the point with the highest probability, and diffuses to the edge along the radius direction. Set k1×k2 (k1=1,3,5,……;k2=3,5,7,……) Gaussian distribution kernel to characterize the central area of ​​the image of OK type defects. The Gaussian distribution is an elliptical distribution (see Figure 2 As shown in the figure, the center is the point with the highest probability and diffuses to the edge along the long axis and short axis. The probability value corresponding to the center point of the Gaussian kernel is set to the highest 1.0, and the probability of the edge of the defective Gaussian kernel is set to 0.5; the ratio of the loss value calculation of the central Gaussian kernel identification area and the loss value calculation of the remaining areas is set to 5:1.

[0045] Please refer to Figure 4As shown, in step 3, the defect cropped image and the Gaussian kernel distribution identification image are used as two inputs of the convolutional neural network, and a feature extraction module and a feature extraction supervision module are provided to extract features from the defect cropped image and the Gaussian kernel distribution identification image. The feature extraction module consists of three 7×7 convolutional layers and one 3×3 pooling layer. At the same time, the channel attention mechanism and the semantic attention mechanism are introduced as the feature extraction supervision module. The output from the first convolutional layer to the last convolutional layer is used as the action layer of the feature extraction supervision module, and the feature information of the defect cropped image and the Gaussian kernel distribution identification image is extracted at the same time. Among them, the central core identification area of ​​the Gaussian kernel distribution identification image is the defect occurrence area, and the parameter ratio of the feature extraction weight variable of the Gaussian kernel identification area to the feature extraction weight variable of the non-Gaussian kernel identification area is set to 7:3. A feature prediction and classification module is provided to predict the defect type and output the prediction probability value of the current defect being judged as different types. The feature prediction and classification module consists of a 3×3 convolution layer, a BN layer, a fully connected layer and a Softmax layer. The prediction probability value of the current defect being judged as different types is output through the Softmax layer, and the type index corresponding to the maximum probability is taken as the prediction result of the currently input ceramic substrate defect image. The prediction probability values ​​of different types output by the Softmax layer are described by formula (II):

[0046]

[0047] Where i=1,2,3,…n corresponds to n types; set n=2, the Softmax layer operation outputs the probability prob1 that the defect image of the ceramic substrate is classified as an OK type defect and the probability prob2 that it is classified as an NG type, and prob1+prob2=1.0.

[0048] In step 4, the Gaussian kernel distribution identification image corresponding to the ceramic substrate defect image is randomly divided into a training defect data set (Defect_Data_Train) for model training and a verification defect data set (Defect_Data_Valid) for model verification according to the same preset ratio, preferably the ratio is set to 4:1. The number of single calculation images, loss function threshold and maximum number of iterations are set for the training defect data set. Preferably, the number of training defect data sets is M, the number of verification defect data sets is N, the number of model training cycles is set to 10×M, and the model calculation loss threshold is 1e -4 , the number of images calculated at a time on the NVIDIA RTX 1660Ti graphics card device is 24, the corresponding training defect image size is 256×256, the model training learning rate is initialized to 0.01, and the warm-up strategy and weight-decay strategy are added in the model training process.

[0049] It is worth noting that the learning rate gradually increases at the beginning of the training phase to increase the depth of image feature learning, and gradually stabilizes as the model training progresses. Within 1 / 3 of the cycle after the model training ends, the learning rate begins to gradually decrease to avoid the problem of local fitting in the training caused by over-depth feature extraction. When the model training reaches the set maximum number of iteration cycles, or the model calculation loss value reaches the set threshold, the weight information parameters and network model framework in the current model training state are solidified and exported as a prediction model.

[0050] In step 5, the defect types of the images in the verification defect data set are manually identified, the prediction model is called on the identified verification defect data set, and the predicted defect types are obtained in a loop. After comparison with the known defect types, a correct prediction is recorded as an accurate prediction, and an incorrect prediction is recorded as an incorrect prediction. The accuracy of the current model prediction is statistically calculated, and the prediction model with the best evaluation result is selected as the intelligent classification network model for ceramic substrate defects.

[0051] Table 1 shows the evaluation statistical results of the intelligent classification network model for ceramic substrate defects in this embodiment in the classification of three groups of different types of defects (solder residue vs air impurities, porcelain stains vs air impurities, scratches vs hair impurities). Among them, the error recognition rate of the intelligent classification network model for ceramic substrate defects for OK type defects (that is, the ratio of identifying OK type defects as NG type defects) is the pass rate, and the correct recognition rate for NG type defects is the detection rate. It can be seen from Table 1 that the intelligent classification network model for ceramic substrate defects in this embodiment is applied to the above three groups of the same type of defect classification, the OK pass rate is lower than 0.5%, the NG detection rate is higher than 99.9%, and the evaluation results are better.

[0052] Serial number Real defect: Interference defect OK inspection rate NG detection rate 1 Solder Residue: Air Impurities 0.3％ 99.99％ 2 Porcelain pollution: air impurities 0.2％ 99.95％ 3 Scratches: Hair impurities 0.5％ 99.95％

[0053] Table 1 Statistics of classification results of different types of defects in ceramic substrate defect images

[0054] In step 6, after a comprehensive evaluation of its prediction and a preliminary assessment of the feasibility of the currently trained intelligent classification network model for ceramic substrate defects, more than 10 new sample finished product images are collected, and the defect area image is obtained after calling the 2D visual detection template matching detection algorithm. The defect area image is directly input into the intelligent classification network model for ceramic substrate defects, and the defects that are judged by the model as OK type defects or the predicted probability value does not reach the detection threshold are automatically not displayed, and only the coordinate information of the classified defect area is displayed, and then the detected defects are evaluated for defect quantification level, and a series of quantitative indicators such as the area and length of the defect area are calculated. Further, the one-time finished product pass rate of the inspection product can be statistically tested according to the classification results to verify the timeliness and practicality of the intelligent detection classification model. Preferably, the OK type defect image and NG type defect image generated in step 6 can be incorporated into the ceramic substrate defect image data set constructed in step 1, enriching the sample size in the ceramic substrate defect image data set, and then iteratively optimizing the model to improve the intelligent classification performance of the model.

[0055] In summary, the intelligent classification method for ceramic substrate defects based on central Gaussian kernel distribution of the present invention extracts the coordinates of the center point of the defect area for the defect data set detected by the 2D visual detection algorithm, intercepts the defect image of a fixed size, and builds a ceramic substrate defect intelligent classification network model. Different Gaussian kernel distributions are used as the feature representation of the defect center area, which effectively improves the accuracy of defect classification, controls the false detection rate of interference defects while ensuring the detection rate of defects to be detected, and improves the quality safety and stability of ceramic substrates.

[0056] The above embodiments are only used to illustrate the present invention but not to limit the technical solutions described in the present invention. The understanding of this specification should be based on the technicians in the relevant technical field. Although this specification has described the present invention in detail with reference to the above embodiments, ordinary technicians in the field should understand that technicians in the relevant technical field can still modify or replace the present invention with equivalents, and all technical solutions and improvements that do not depart from the spirit and scope of the present invention should be covered within the scope of the claims of the present invention.

Claims

1. A ceramic substrate defect intelligent classification method based on central Gaussian kernel distribution, characterized in that: The following steps are involved: Step 1: construct a ceramic substrate defect dataset, wherein the ceramic substrate defect dataset includes ceramic substrate defect images of different types of defects, and the types of defects include n types; Step 2: Obtain the defect cropped image and its corresponding Gaussian kernel distribution identification image; Step 3, model construction: using the defect cropped image and Gaussian kernel distribution identification image as two inputs of the neural network model, providing a feature extraction module and a feature extraction supervision module to extract features from the defect cropped image and Gaussian kernel distribution identification image, providing a feature prediction classification module to predict the defect type and output the predicted probability value of the current defect being judged as different types; Step 4, model training: the Gaussian kernel distribution identification image corresponding to the ceramic substrate defect image is divided into a training defect data set for model training and a verification defect data set for model verification according to the same preset ratio; the number of single calculation images, the loss function threshold and the maximum number of iterations are set for the training defect data set. When the model training reaches the set maximum number of iterations, or the model calculation loss value reaches the set loss function threshold, the weight information parameters and the network model framework in the current model training state are solidified and exported as a prediction model; Step 5, model evaluation: manually mark the defect types of the images in the verification defect data set, call the prediction model for the marked verification defect data set, cyclically obtain the predicted defect types, record a correct prediction as an accurate prediction, record an incorrect prediction as an incorrect prediction, and statistically calculate the accuracy of the current model prediction; select the prediction model with the best evaluation result as the ceramic substrate defect intelligent classification network model; Step 6: Apply the ceramic substrate defect intelligent classification network model to classify the ceramic substrate defective products.

2. The intelligent classification method for ceramic substrate defects based on central Gaussian kernel distribution according to claim 1, characterized in that: In step 1, more than 300 defect images of ceramic substrates are obtained. There are two types of defects, including OK type defects and NG type defects. The OK type defects include bright spot and white spot defects and erasable dirt defects, and the NG type is solder residue defects.

3. The intelligent classification method for ceramic substrate defects based on central Gaussian kernel distribution according to claim 2, characterized in that: In step 2, the defect image of the ceramic substrate is cropped according to a preset size to obtain a defect cropped image, the center point of the defect cropped image is the center point of the defect area; the defect cropped image is marked with a central Gaussian distribution kernel to obtain a Gaussian kernel distribution mark image of all defect cropped images; The Gaussian distribution function formula is: Where (x, y) is the image pixel coordinate, which obeys the normal distribution of constants (σ1, σ2, μ1, μ2, ρ); Set k×k (k=1,3,5,……) Gaussian distribution kernel to characterize the central area of ​​the image in NG type images, σ1=σ2, μ1=μ2, and diffuse toward the edge along the radial direction with the center as the highest probability point; set k1×k2 (k1=1,3,5,……; k2=3,5,7,……) Gaussian distribution kernel to characterize the central area of ​​the image of OK type defects, and diffuse toward the edge along the major and minor axis directions with the center as the highest probability point.

4. The ceramic substrate defect intelligent classification method based on central Gaussian kernel distribution according to claim 3, characterized in that: The ceramic substrate is an AMB ceramic substrate, the probability value corresponding to the center point of the Gaussian kernel is set to the highest 1.0, and the probability of the edge of the defective Gaussian kernel is set to 0.5; the ratio of the loss value calculation of the central Gaussian kernel identification area and the loss value calculation of the remaining areas is set to 5:

1.

5. The ceramic substrate defect intelligent classification method based on central Gaussian kernel distribution according to claim 3, characterized in that: The parameter ratio of the feature extraction weight variable of the central Gaussian kernel marked area to the feature extraction weight variable of the non-Gaussian kernel marked area is set to 7:

3.

6. The ceramic substrate defect intelligent classification method based on central Gaussian kernel distribution according to claim 1, characterized in that: In step 3, the feature extraction module consists of three 7×7 convolutional layers and one 3×3 pooling layer. At the same time, the channel attention mechanism and the semantic attention mechanism are introduced as the feature extraction supervision module. The output from the first convolutional layer to the last convolutional layer is used as the action layer of the feature extraction supervision module, and the feature information of the defect cropped image and the Gaussian kernel distribution identification image is extracted at the same time; the feature prediction classification module consists of a 3×3 convolutional layer, a BN layer, a fully connected layer and a Softmax layer. The Softmax layer outputs the predicted probability value of the current defect being judged as different types, and the type index corresponding to the maximum probability is taken as the prediction result of the currently input ceramic substrate defect image.

7. The intelligent classification method for ceramic substrate defects based on central Gaussian kernel distribution according to claim 6, characterized in that: The Softmax layer outputs different types of predicted probability values, which are described by formula (II): Where i = 1, 2, 3, ... n corresponds to n types; Set n=2, the Softmax layer operation outputs the probability prob1 that the ceramic substrate defect image is classified as an OK type defect and the probability prob2 that it is classified as an NG type defect, and prob1+prob2=1.

0.

8. The intelligent classification method for ceramic substrate defects based on central Gaussian kernel distribution according to claim 1, characterized in that: In step 4, the Gaussian kernel distribution identification image corresponding to the ceramic substrate defect image is divided into a training defect data set for model training and a verification defect data set for model verification at a ratio of 4:

1.

9. The intelligent classification method for ceramic substrate defects based on central Gaussian kernel distribution according to claim 1, characterized in that: In step 4, the number of training defect data sets is M, the number of verification defect data sets is N, the number of model training cycles is set to 10×M, and the model calculation loss threshold is 1e -4 The number of images calculated at a single time is 24, the corresponding training defect image size is 256×256, the model training learning rate is initialized to 0.01, and the warm-up strategy and weight-decay strategy are added in the model training process.

10. The intelligent classification method for ceramic substrate defects based on central Gaussian kernel distribution according to claim 1, characterized in that: In step 6: several new sample finished product images are newly collected, and the defect area image is obtained after calling the 2D visual detection template matching detection algorithm. The defect area image is directly input into the ceramic substrate defect intelligent classification network model. The defects judged by the model as OK type defects or the defects whose predicted probability values ​​do not reach the preset detection threshold are automatically not displayed. Only the coordinate information of the defect area after intelligent classification is displayed, and then the defect quantitative level evaluation is performed on the detected defects.

Images