TWI747686B - A defect detection method and a defect detection device - Google Patents

Abstract

The present invention relates to a defect detection method and a defect detection device. The method includes three detection steps. The first detection step is to setting a defect template in an AOI system and detecting the defect type of the defect template to obtain a first detection result. If the first detection result is accurate, then output, otherwise proceed to the second detection step. The second detection step is to transmit the defect template to the AI model to detect the defect type of the defect template and to obtain the second detection result, and to calculate the detection accuracy of the AI model. If the result is accurate and the detection accuracy meets the set threshold, the second detection result is output and the AI model is trained. If the detection accuracy is less than the set threshold, the third detection step is performed. In the third detection step, the defect template is transmitted to the re-inspection system and a matching detection between the defect template and the preset template is performed to obtain the third detection result, and calculate the re-inspection accuracy rate. If the re-inspection accuracy rate meets the set threshold, output the third detection result and train the AI model.

Description

Defect detection method and detection device

The present invention relates to the technical field of pattern detection, in particular to a detection method and a detection device for defect detection.

At present, defect detection technology, such as chip defect detection, traditionally includes two detection methods. One is to rely on manual visual inspection, due to the limited viewing angle of people (for example, it is impossible to effectively observe the unsoldering between small SMD components, reverse polarity, etc. ), there are shortcomings of slow detection speed, low accuracy, and poor detection effect, so this method is gradually used less and less. Another way is to judge through the AOI system (automatic optical inspection, which is based on optical principles to detect common defects encountered in welding production); although the AOI system uses optical principles, optical lenses are used instead of human eyes, and the Zooming in during the shooting process can obtain clearer equipment images. However, the AOI system has the disadvantage that the method of judging whether the detection point is faulty is manually based on the standard digital image stored in the AOI system and the actual detection. For comparison and discrimination of the images in the image, which means that human visual contrast detection is also required, so there are also the shortcomings of slow detection speed, missed detection, and low accuracy.

On this basis, with the development of artificial intelligence technology, artificial intelligence deep learning methods have also been applied to detection technology. However, the current artificial intelligence in detection, classification, and detection requires long-term large-scale sample collection, training and repeated testing, which makes it difficult for artificial intelligence to let users see the results and benefits in the short term or quickly, and artificial intelligence cannot handle it. Quantitative, positioning and logical relationships. In addition, most of the current artificial intelligence detection methods use neural network learning methods to detect chip defects. The neural network achieves the detection purpose by training the hidden layer, but the training and detection process of the hidden layer is similar to a "black box". , The judging process is unknowable, resulting in uncontrollable test results.

In order to solve the above problems, the present invention provides a defect detection method and detection device. Based on the use of AOI system to obtain product pictures, a large number of product defect pictures are used as a picture database to automatically establish artificial intelligence models in real time, and continuously improve artificial intelligence model defect detection At the same time, according to the desired detection accuracy and the actual detection accuracy of the artificial intelligence model, a custom detection method based on AOI detection logic is introduced as appropriate to make the detection result more controllable.

Specifically, the present invention provides a defect detection method, including: In the first detection step, a defect template is set in the AOI system. The AOI system detects the defect type of the defect template and obtains a first detection result. If the first detection result is accurate, the first detection result is output, Otherwise, proceed to the second detection step; The second detection step is to transmit the defect template to an artificial intelligence model, the artificial intelligence model detects the defect type of the defect template, obtains a second detection result, and calculates the detection accuracy of the artificial intelligence model, If the second detection result is accurate and the detection accuracy rate meets a set threshold, output the second detection result and train the artificial intelligence model; if the detection accuracy rate is less than the set threshold, perform The third detection step; In the third inspection step, the defect template is transmitted to a re-inspection system, and the re-inspection system performs matching inspection on the defect template and a preset template, obtains the third inspection result, and calculates the re-inspection accuracy rate, if If the recheck accuracy rate meets the set threshold, the third detection result is output and the artificial intelligence model is trained.

According to the detection method, wherein the training step of the artificial intelligence model includes: Collect defective pictures of products, and classify the defective pictures according to product category and defect type; Separately count the number of defective pictures with various types of defects for each product. When the number of defective pictures with a type of defect for any product exceeds a preset value, a preset method is used to train the artificial intelligence model for each Describe the defect pictures for inspection; The training detection accuracy rate of the artificial intelligence model is calculated, and when the training detection accuracy exceeds the prediction accuracy rate of the artificial intelligence model trained by the preset method, the artificial intelligence model training for the defective picture is completed.

According to the detection method, wherein the step of classifying the defective picture includes: Obtain the product information to be tested, and determine the category of the product; Acquiring a defect picture of the product according to the product information, and determining the defect location and defect type of the defect picture; The defect pictures are stored in a picture collection table according to the product category and the defect type.

According to the detection method, wherein the preset method includes a first training method, a second training method, and a third training method, and the first training method, the second training method, and the third training method have different results. The prediction accuracy rate.

According to the detection method, wherein the third detection step further includes: Identifying the defect template, and matching the defect template to a preset template; Extracting feature information of the defect template, where the feature information includes defect location, color, and compliance; The feature information is compared and detected with the preset template one by one.

In order to achieve another objective of the present invention, the present invention also provides a defect detection device, including: The receiving module is used to obtain the defect template to be detected; The first detection module includes an AOI system, and the AOI system is used to detect the defect type of the defect template to obtain the first detection result; The second detection module includes an artificial intelligence model that is used to perform defect detection on the defect template according to the trained detection data to obtain a second detection result; The third inspection module includes a re-inspection system, and the re-inspection system performs matching inspection on the defect template with a preset template to obtain a third inspection result; The control module is used to separately judge the first detection result, the second detection result and the third detection result, wherein, if the first detection result is accurate, output the first detection result, otherwise the The defect template is transmitted to the second inspection module; The control module is also used to calculate the detection accuracy rate of the artificial intelligence model, and if the second detection result is accurate and the detection accuracy rate meets a set threshold, output the second detection result and compare the The artificial intelligence model is trained, and if the detection accuracy rate is less than the set threshold, the defect template is transmitted to the third detection module; The control module is also used to calculate the re-inspection accuracy rate of the re-inspection system, and if the re-inspection accuracy rate meets the set threshold, output the third inspection result and train the artificial intelligence model.

According to the detection device, wherein the second detection module further includes a training unit for training the artificial intelligence model; Wherein, the training unit includes: The defective picture classification subunit is used to collect defective pictures of products and classify the defective pictures according to product category and defect type; The model training subunit is used to separately count the number of defect pictures with each type of defect for each product. When the number of defect pictures with a type of defect for any product exceeds a preset value, a preset method is used to train the The artificial intelligence model detects each of the defective pictures; The calculation subunit is used to calculate the training detection accuracy rate of the artificial intelligence model. When the training detection accuracy rate exceeds the prediction accuracy rate of the artificial intelligence model trained by the preset method, the manual operation for the defective picture is completed. Intelligent model training.

According to the detection device, wherein the defective picture classification subunit further includes: The product identification subunit is used to obtain information about the product to be tested and determine the category of the product; The defect identification subunit is used to obtain a defect picture of the product according to the product information, and determine the defect location and defect type of the defect picture; The storage subunit is used to store the defective pictures in a picture collection table according to the product category and the defect type.

According to the detection device, wherein the preset method includes a first training method, a second training method, and a third training method, and the first training method, the second training method, and the third training method have different results. The prediction accuracy rate.

According to the detection device, wherein the third detection module further includes: The template matching unit is used to identify the defective template and match the defective template to a preset template; A feature extraction unit for extracting feature information of the defect template, where the feature information includes defect location, color, and compliance; The detection unit is used to compare and detect the characteristic information with the preset template one by one.

The following describes the present invention in detail with reference to the accompanying drawings and specific embodiments, but it is not intended to limit the present invention.

The structural principle and working principle of the present invention will be described in detail below in conjunction with the accompanying drawings. The description in this part is only exemplary and explanatory, and should not have any limiting effect on the protection scope of the present invention.

Please refer to FIG. 1, which is a schematic flowchart of a defect detection method according to an embodiment of the present invention. The defect detection method of this embodiment includes the following steps.

The first detection step S1 is implemented based on the existing AOI system. A defect template is set in the AOI system. The defect template is a template or picture of a type of defect of a product to be tested. In some embodiments, the defect template may include defect images of a larger number of products. After the AOI system obtains the defect template, it distinguishes its defect type, and then detects its specific defect to obtain the first inspection result. If the first inspection result is accurate, that is, the AOI inspection is credible, the first inspection result is The output is directly performed, otherwise the second detection step S2 is performed, that is, the second detection step S2 is initiated because the AOI detection in the first detection step S1 is not credible (the accuracy of the detection does not meet the needs of the user).

The second detection step S2, the detection step is mainly based on a trained artificial intelligence model, which is implemented based on a neural network, for example. Specifically, after the first detection step S1 is confirmed to be untrustworthy, the defect template is transmitted to the artificial intelligence model, and the artificial intelligence model compares the defect template with the trained inspection data to detect the defect type of the defect template, and obtain The second detection result, and the detection accuracy rate of the artificial intelligence model is calculated at the same time. If the second detection result is accurate and the detection accuracy rate meets a set threshold, that is, the detection accuracy rate reaches or exceeds the set threshold, then the second detection result is output and The artificial intelligence model is trained, and if the detection accuracy rate is less than the set threshold, the third detection step S3 is activated to continue detection. In this embodiment, by performing cyclic training on the artificial intelligence detection process, the automatic establishment mechanism of the artificial intelligence model is realized, and at the same time, the artificial intelligence model is continuously trained and learned to further improve the detection accuracy.

Specifically, this embodiment provides an artificial intelligence model training method, but the present invention is not limited to this. Please refer to FIG. 2. FIG. 2 is a schematic diagram of the flow of an artificial intelligence model training method according to an embodiment of the present invention. As shown in FIG. 2, the artificial intelligence model training method includes the following steps.

S21. Collect defective pictures of products, and classify the defective pictures according to product category and defect type. The present invention does not limit the collection method of defective pictures. In this embodiment, a picture labeling system is provided to obtain pictures and store them in categories. Specifically, please refer to FIG. 3, which is a schematic flowchart of a method for classifying defective pictures according to an embodiment of the present invention. The steps for classifying defective pictures of defective pictures include: S211. Obtain product information to be tested to determine the category of the product. That is, the picture labeling system determines the category of the product according to the product barcode by scanning the product barcode. S212. Obtain the defect picture of the product according to the product information, and determine the defect location and defect type of the defect picture. That is, the picture labeling system finds the product defect picture to be classified according to the product barcode, and displays the coordinate of the product defect picture to be classified in the display area, as shown in Figure 4. The operator clicks on the corresponding coordinate point according to the actual defect position, and then the corresponding coordinate point will pop up the defect picture and defect code selection box to complete the labeling and classification of the defect picture. S213. Store the defect pictures in a picture collection table according to the product category and defect type. At this time, the operator selects the defect code according to the actual defect, and the picture labeling system classifies and stores the actual defect in a picture collection table according to the defect code selected. The picture collection form is a two-dimensional form indicating product types and defect types. Of course, the present invention is not limited to this, and the storage form of the picture may also be other suitable forms such as bar graph information and fan graph information.

S22. Separately count the number of defect pictures with various types of defects for each product. When the number of pictures with any type of defect in a certain product exceeds a preset value, use a preset method to train the artificial intelligence model for each defect Picture for inspection. In this embodiment, when the picture labeling system collects more than 500 defective pictures of a certain defect type stored in the picture collection table, the artificial intelligence model is started to be trained.

S23. Calculate the training and detection accuracy of the artificial intelligence model. When the training detection accuracy exceeds the prediction accuracy of the artificial intelligence model trained by the preset method, complete the artificial intelligence model training for the defective image. Among them, in this embodiment, multiple training algorithms for the artificial intelligence model can be preset, and each training algorithm has a different prediction accuracy rate, for example, including the first training method, the second training method, and the third training method. The training method, the second training method and the third training method respectively have different prediction accuracy rates. In actual training, the actual detection accuracy rate of the artificial intelligence model under the training algorithm is calculated, and the actual detection accuracy rate is compared with the prediction accuracy rate of the algorithm. When the actual detection accuracy rate reaches or exceeds the prediction accuracy rate, Then activate the artificial intelligence model trained on this type of defect picture.

For the detection accuracy obtained in the second detection step S2, if the user's requirements are not yet satisfied, the third detection step S3 is further performed. That is to say, in the present invention, the second detection step S2 and the third detection step S3 are a flexible combination. The third detection step S3 is mainly based on a re-inspection system that integrates specific detection methods on the basis of AOI detection logic.

Specifically, when it is confirmed that the detection accuracy obtained in the second detection step S2 cannot meet the requirements, the defect template is transmitted to a re-inspection system, and the re-inspection system performs matching detection between the defect template and a preset template to obtain the first Three detection results, and the re-inspection accuracy rate is calculated. If the re-inspection accuracy rate reaches or exceeds the set threshold, the third detection result is output and the artificial intelligence model is trained. The set threshold can be actually set according to user needs.

Please refer to FIG. 5 for details. FIG. 5 is a schematic diagram of the detection process of the re-inspection system according to an embodiment of the present invention, which includes the following detection steps: S31. Identify the defect template, and match the defect template with a preset template in the re-inspection system; S32. Extract feature information of the defect template, where the feature information includes, for example, defect location, color, and compliance; S33. Compare and detect the feature information with the preset template one by one.

In order to achieve another objective of the present invention, based on the same inventive concept described above, the present invention also provides a defect detection device. Please continue to refer to FIG. 6, which shows a block diagram of a private defect detection device according to the present invention. As shown in FIG. 6, the defect detection device 100 of this embodiment specifically includes a receiving module 110, a first detection module 120, a second detection module 130, a third detection module 140, and a control module 150.

Among them, the receiving module is used to obtain the defect template to be detected. The first inspection module 120 includes an AOI system, which is used to detect the defect type of the defect template to obtain the first inspection result; the second inspection module 130 includes an artificial intelligence model, which is used to The trained inspection data performs defect inspection on the defect template to obtain the second inspection result; the third inspection module 140 includes a re-inspection system, and the re-inspection system performs matching inspection on the defect template with a preset template to obtain the third inspection Result; the control module 150 is used to separately determine the first detection result, the second detection result and the third detection result, wherein, if the first detection result is accurate, output the first detection result, otherwise the defect template is transmitted to The second detection module 130; the control module 150 is also used to calculate the detection accuracy of the artificial intelligence model. If the second detection result is accurate and the detection accuracy meets a set threshold, the second detection result is output and the The artificial intelligence model is trained. If the detection accuracy rate is less than the set threshold, the defect template is transmitted to the third detection module 140; the control module 150 is also used to calculate the re-inspection accuracy rate of the re-inspection system, if the re-inspection accuracy rate is If the threshold is met (that is, reached or exceeded), the third detection result is output and the artificial intelligence model is trained.

Further, the second detection module 130 further includes a training unit 131 for training the artificial intelligence model. Among them, the training unit 131 includes the following.

The defective picture classification subunit 1311 is used to collect defective pictures of products and classify the defective pictures according to product categories and defect types; the present invention does not limit the collection method of defective pictures. In this embodiment, a picture label is provided The system is used to obtain pictures and sort them into storage. Specifically, the classification subunit of the defective picture further includes: The product identification subunit A is used to obtain the information of the product to be tested and determine the category of the product. That is, the picture labeling system determines the category of the product according to the product barcode by scanning the product barcode. The defect recognition subunit B is used to obtain the defect picture of the product according to the product information, and determine the defect location and defect type of the defect picture. That is, the picture labeling system finds the product defect picture to be classified according to the product barcode, and displays the coordinate of the product defect picture to be classified in the display area, as shown in Figure 4. The operator clicks on the corresponding coordinate point according to the actual defect position, and then the corresponding coordinate point will pop up the defect picture and defect code selection box to complete the labeling and classification of the defect picture. The storage subunit C is used to store the defective pictures in a picture collection table according to the product category and the defect type. At this time, the operator selects the defect code based on the actual defect, and the picture labeling system sorts and stores the actual defect in a picture collection table according to the selected defect code. The picture collection table is a two-dimensional table indicating the product type and the defect type. Of course, the present invention is not limited to this, and the storage form of the picture may also be other suitable forms such as bar graph information and fan graph information.

The model training subunit 1312 is used to separately count the number of defective pictures with various types of defects for each product. When the number of defective pictures with a type of defect for any product exceeds a preset value, a preset method is used to train the manual The intelligent model detects each defect picture. In this embodiment, when the picture labeling system collects more than 500 defective pictures of a certain defect type stored in the picture collection table, the artificial intelligence model is started to be trained.

The calculation subunit 1313 is used to calculate the training detection accuracy of the artificial intelligence model. When the training detection accuracy exceeds the prediction accuracy of the artificial intelligence model trained by the preset method, the artificial intelligence model training for the defective picture is completed. Among them, in this embodiment, multiple training algorithms for the artificial intelligence model can be preset, and each training algorithm has a different prediction accuracy rate, for example, including the first training method, the second training method, and the third training method. The training method, the second training method and the third training method respectively have different prediction accuracy rates. In actual training, the actual detection accuracy rate of the artificial intelligence model under the training algorithm is calculated, and the actual detection accuracy rate is compared with the prediction accuracy rate of the algorithm. When the actual detection accuracy rate reaches or exceeds the prediction accuracy rate, Then activate the artificial intelligence model trained on this type of defect picture.

For the detection accuracy rate obtained by the second detection module 130, if the requirements of the user are not yet satisfied, the third detection module 140 is further activated. That is to say, in the present invention, the second detection module 130 and the third detection module 140 are a flexible combination. The third detection module 140 is mainly based on a re-inspection system that integrates specific detection methods on the basis of AOI detection logic.

Specifically, when it is confirmed that the detection accuracy obtained in the second detection step S2 cannot meet the requirements, the defect template is transmitted to a re-inspection system, and the re-inspection system performs matching detection between the defect template and a preset template to obtain the first Three detection results, and the re-inspection accuracy rate is calculated. If the re-inspection accuracy rate meets (that is, reaches or exceeds) the set threshold, the third detection result is output and the artificial intelligence model is trained. Among them, the set threshold can be actually set according to user needs.

Specifically, the third detection module 140 further includes: a template matching unit 141, a feature extraction unit 142, and a detection unit 143, wherein the template matching unit 141 is used to identify a defective template and match the defective template with a preset template; The feature extraction unit 142 is used to extract feature information of the defect template, where the feature information includes defect location, color, and coincidence; the detection unit 143 is used to compare and detect the feature information with the preset template one by one.

In summary, the present invention provides a new defect detection method and device for various electronic components. In view of the low detection accuracy of the artificial intelligence model, the present invention separately increases the score card control (that is, sets the set threshold ), and in the re-inspection system, a custom inspection method based on AOI logic is adopted to re-inspect product defects, and the inspection and determination process is artificially interfered to increase the controllability of the inspection results. Further, the present invention allows users to see results and benefits quickly, and can start the detection process by setting and inputting pictures of different defect types, so that the detection effect is more efficient. The present invention collects defective pictures during the defect detection process through the picture labeling system, and continuously trains the artificial intelligence model, so that the artificial intelligence model can be automatically established, and avoids the need for long-term large-scale sample collection, training and training in the process of detection and classification. A situation of repeated testing. In addition, the defect detection method and device of the present invention can be used in parallel for a variety of objects to be determined to improve the detection efficiency.

Of course, the present invention can also have various other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art can make various corresponding changes and modifications according to the present invention, but these corresponding changes and All deformations shall fall within the protection scope of the attached patent application of the present invention.

S1~S3, S21~S23, S211~S213, S31~S33: steps 100: Defect detection device 110: receiving module 120: The first detection module 130: The second detection module 131: Training Unit 1311: Defect image classification subunit A: Product identification subunit B: Defect Recognition Subunit C: Storage subunit 1312: Model training subunit 1313: calculation subunit 140: The third detection module 141: template matching unit 142: Feature Extraction Unit 143: detection unit 150: control module

FIG. 1 is a schematic flowchart of a defect detection method according to an embodiment of the present invention. Fig. 2 is a schematic flowchart of a training method of an artificial intelligence model according to an embodiment of the present invention. FIG. 3 is a schematic flowchart of a method for classifying defective pictures according to an embodiment of the present invention. Fig. 4 is a schematic diagram of a picture labeling system according to an embodiment of the present invention. FIG. 5 is a schematic diagram of the detection process of the re-inspection system according to an embodiment of the present invention. Fig. 6 is a frame diagram of a defect detection device according to an embodiment of the present invention.

S1~S3: steps

Claims (10)

A defect detection method, including: In the first detection step, a defect template is set in the AOI system. The AOI system detects the defect type of the defect template and obtains a first detection result. If the first detection result is accurate, the first detection result is output, Otherwise, proceed to the second detection step; The second detection step is to transmit the defect template to an artificial intelligence model, the artificial intelligence model detects the defect template, obtains a second detection result and calculates the detection accuracy of the artificial intelligence model, if the first 2. If the detection result is accurate and the detection accuracy rate meets a set threshold, output the second detection result and train the artificial intelligence model. If the detection accuracy rate is less than the set threshold, perform a third detection step ； In the third inspection step, the defect template is transmitted to a re-inspection system, and the re-inspection system performs matching inspection on the defect template and a preset template to obtain the third inspection result and calculate the re-inspection accuracy rate. If the detection accuracy rate meets the set threshold, output the third detection result and train the artificial intelligence model. The defect detection method according to claim 1, wherein the training step of the artificial intelligence model includes: Collect defective pictures of products, and classify the defective pictures according to product category and defect type; Separately count the number of defective pictures with various types of defects for each product. When the number of defective pictures with a type of defect for any product exceeds a preset value, a preset method is used to train the artificial intelligence model for each Check the defect pictures one by one; The training detection accuracy rate of the artificial intelligence model is calculated, and when the training detection accuracy rate is greater than or equal to the prediction accuracy rate of the artificial intelligence model trained by the preset method, the artificial intelligence model training for the defective picture is completed. The defect detection method according to claim 2, wherein the step of classifying the defect picture includes: Obtain the product information to be tested, and determine the category of the product; Acquiring a defective picture of the product according to the product information, and determining the defect type of the defective picture; The defect pictures are classified and stored according to the product category and the defect type. The defect detection method according to claim 2, wherein the preset method includes a first training method, a second training method, and a third training method, and the first training method, the second training method, and the third training method Each has a different prediction accuracy rate. The defect detection method according to claim 1, wherein the third detection step further includes: Identifying the defect template, and matching the defect template to a preset template; Extracting feature information of the defect template, where the feature information includes defect location, color, and compliance; The feature information is compared and detected with the preset template one by one. A defect detection device, including: The receiving module is used to obtain the defect template to be detected; The first detection module includes an AOI system, and the AOI system is used to detect the defect type of the defect template to obtain the first detection result; The second detection module includes an artificial intelligence model that is used to perform defect detection on the defect template according to the trained detection data to obtain a second detection result; The third inspection module includes a re-inspection system, and the re-inspection system performs matching inspection on the defect template with a preset template to obtain a third inspection result; The control module is used to separately judge the first detection result, the second detection result and the third detection result, wherein, if the first detection result is accurate, output the first detection result, otherwise the The defect template is transmitted to the second inspection module; The control module is also used to calculate the detection accuracy rate of the artificial intelligence model, and if the second detection result is accurate and the detection accuracy rate meets a set threshold, output the second detection result and compare the The artificial intelligence model is trained, and if the detection accuracy rate is less than the set threshold, the defect template is transmitted to the third detection module; The control module is also used to calculate the re-inspection accuracy rate of the re-inspection system, and if the re-inspection accuracy rate meets the set threshold, output the third inspection result and train the artificial intelligence model. The defect detection device according to claim 6, wherein the second detection module further includes a training unit for training the artificial intelligence model; Wherein, the training unit includes: The defective picture classification subunit is used to collect defective pictures of products and classify the defective pictures according to product category and defect type; The model training subunit is used to separately count the number of defect pictures with each type of defect for each product. When the number of defect pictures with a type of defect for any product exceeds a preset value, a preset method is used to train the The artificial intelligence model detects each of the defective pictures; The calculation subunit is used to calculate the training detection accuracy rate of the artificial intelligence model. When the training detection accuracy rate exceeds the prediction accuracy rate of the artificial intelligence model trained by the preset method, the manual operation for the defective picture is completed. Intelligent model training. The defect detection device according to claim 7, wherein the defect picture classification subunit further includes: The product identification subunit is used to obtain information about the product to be tested and determine the category of the product; The defect identification subunit is used to obtain a defect picture of the product according to the product information, and determine the defect location and defect type of the defect picture; The storage subunit is used to classify and store the defect pictures according to the product category and the defect type. The defect detection device according to claim 7, wherein the preset method includes a first training method, a second training method, and a third training method, and the first training method, the second training method, and the third training method Each has a different prediction accuracy rate. The defect detection device according to claim 6, wherein the third detection module further includes: The template matching unit is used to identify the defective template and match the defective template to a preset template; A feature extraction unit for extracting feature information of the defect template, where the feature information includes defect location, color, and compliance; The detection unit is used to compare and detect the characteristic information with the preset template one by one.

Images