TWI834426B - Evaluation method, evaluation device and training method for pcb defect detection model - Google Patents

Abstract

The present disclosure provides an evaluation method, an evaluation device and a training method of a PCB defect detection model, wherein the evaluation method comprising: pre-establishing a test image set, comprising a plurality of test images with defect category labels; inputting a plurality of test images into a defect detection model to be evaluated to obtain a defect prediction results corresponding to the test images, wherein the defect prediction result comprises a defect type and corresponding probability value; classifying the defect prediction results according to the defect types in the obtained defect prediction results; outputting and displaying test images predicted to be the same defect type according to the classification result, wherein the defect category label of each test image and the probability value in the corresponding defect prediction result are configured to be viewable; based on the defect category labels and classification results of the test images, evaluating the ability of the defect detection model to identify various defect types, and comparing the probability values corresponding to the test images under the same classification to evaluate the convergence of the defect detection model to identify this type of defect.

Description

Evaluation method, evaluation device and training of PCB defect detection model method

The invention relates to the field of model quality assessment, and in particular to an assessment method, assessment device and training method for a PCB defect detection model.

In a typical AI training process, defective images are fed into the AI model for identification and classification. Recognized images are assigned a score that indicates the likelihood that they are a defect or a non-defect (a false positive defect). If this score If the score is higher than the preset AI threshold, the image is considered to have real defects. If the score is lower than the AI threshold, the image is considered to have false positive defects.

The reasonable setting of the AI threshold is related to the trade-off between underkill and overkill. If the AI threshold is set to a low value, a false positive defect may be mistakenly identified as a real defect, resulting in excessive lethality and overkill; if the AI threshold is set to a low value, If it is too high, the actual defects may be mistakenly identified as false positive defects, resulting in insufficient lethality and an underkill.

It can be said that the setting of AI threshold is closely related to the accuracy of model output results.

The disclosure of the above background technology content is only used to assist in understanding the inventive concepts and technical solutions of the present invention. It does not necessarily belong to the prior art of this patent application, nor does it necessarily provide technical guidance; in the absence of clear evidence that the above content is If the patent application has been published before the filing date, the above background technology should not be used to evaluate the novelty and inventiveness of the application.

The purpose of the present invention is to provide an evaluation method, evaluation device and training method for a PCB defect detection model to effectively evaluate the recognition ability and recognition convergence of the model.

In order to achieve the above purpose, the technical solution adopted by the present invention is as follows: an evaluation method of a PCB defect detection model, which is used to evaluate the defect prediction ability of a pre-trained defect detection model on PCB images. The defect prediction results include defect types and corresponding probability values. The evaluation method includes: pre-establishing a test image set, which includes multiple test images, each test image has a defect category label; Multiple test images are input into the defect detection model to be evaluated, and defect prediction results corresponding to the test images are obtained; according to the defect types in the obtained defect prediction results, the defect prediction results are Classify the results; according to the classification results, output and display the test images predicted to be the same defect type, and the defect category label of each test image and the probability value in the corresponding defect prediction result are configured to be viewable; according to the test image The defect category labels and classification results of the image are used to evaluate the ability of the defect detection model to identify various defect types, and the probability values corresponding to the test images under the same category are compared to evaluate the convergence of the defect detection model in identifying this type of defect. Spend.

Further, the classification accuracy rate is calculated based on the defect category label and classification result of the test image. If the classification accuracy rate is lower than a preset accuracy threshold, the defect detection model's ability to identify this type of defect is unqualified; Or, if the difference between the maximum value and the minimum value of the probability value corresponding to the test image under the same category is greater than the preset difference threshold, or if the variance of the probability value corresponding to the test image under the same category is greater than the preset difference threshold If the variance threshold is exceeded, the defect detection model's convergence in identifying this type of defect is unqualified.

Further, the evaluation method of the PCB defect detection model also includes: if the evaluation result is unqualified, retraining the defect detection module, and focusing the learning attention on the identification ability or convergence of the defect detection model. Qualified defects of this type.

Further, when the defect detection model has qualified convergence in identifying this type of defect, the boundary for defining real defects and false alarm defects is determined based on the probability value corresponding to the test image under this classification and the preset rules. Score.

Further, the demarcation score used to define real defects and false alarm defects is determined by finding the minimum value of the probability value corresponding to the test image under the classification, and calculating the minimum value with the preset difference threshold. , obtain the demarcation score; or, calculate the average value of the probability values corresponding to the test images under the classification, and operate the average value with the preset difference threshold to obtain the demarcation score; or, for the Sort the average probability values corresponding to the test images under the classification, exclude the first several probability values and the last several probability values, calculate the average value of the remaining probability values, and compare the average value of the remaining probability values with the preset The difference threshold is calculated to obtain the demarcation score; wherein the preset difference threshold is a positive number, a negative number or zero.

Further, the defect category label of each test image and the probability value in the corresponding defect prediction result are configured to be viewable in such a way that the probability value in the defect prediction result is displayed on the corresponding test image. a local area of , right-click on the corresponding test image, and place the cursor on one or more of the corresponding test images.

An evaluation device for a PCB defect detection model according to the present invention includes the following modules: a test sample module, which is configured to establish a test image set, which includes a plurality of test images, each test image having a defect category label; a prediction module configured to input a plurality of test images in the test image set into a defect detection model to be evaluated, and obtain a defect detection model corresponding to the test image One-to-one corresponding defect prediction results; a classification module configured to classify the defect prediction results according to the defect types in the defect prediction results of the prediction module; a display module configured to classify according to the classification module As a result, test images predicted to be of the same defect type are output and displayed, and the defect category label of each test image and the probability value in the corresponding defect prediction result can be displayed; the recognition capability evaluation module is configured to perform the test according to the test image The defect category labels and classification results of the image are used to evaluate the ability of the defect detection model to identify various defect types; the convergence evaluation module is configured to compare the probability values corresponding to test images under the same classification to evaluate the defects. Check the convergence of the model in identifying this type of defect.

Further, the recognition ability evaluation module calculates the classification accuracy rate based on the defect category label and classification result of the test image. If the classification accuracy rate is lower than the preset accuracy threshold, the evaluation result of the recognition ability evaluation module is The defect detection model's ability to identify this type of defect is unqualified; the convergence evaluation module determines whether the difference between the maximum value and the minimum value of the probability value corresponding to the test image under the same category is greater than the preset difference threshold, or , determine whether the variance of the probability values corresponding to the test images under the same category is greater than the predetermined If the variance threshold is set, then the evaluation result of the convergence evaluation module is that the convergence of the defect detection model in identifying this type of defect is unqualified.

Further, the evaluation device of the PCB defect detection model further includes a demarcation score determination module, which is configured to provide a qualified convergence degree for the defect detection model in identifying such defects when the evaluation result of the convergence evaluation module is Next, based on the probability value corresponding to the test image under this category and the preset rules, the dividing score used to define real defects and false positive defects is determined.

According to yet another aspect of the present invention, a method for training a PCB defect detection model is provided. The evaluation method as described above is used to evaluate the currently trained defect detection model. If the ability of the defect detection model to identify all types of defects is evaluated, and the convergence of identifying all types of defects is qualified, then the defect detection model ends training; otherwise, the defect detection model is retrained, and the defect detection model focuses its learning attention on the identification ability or the convergence is unqualified of such defects.

Further, if the ability of the defect detection model to identify this type of defect is evaluated to be unqualified, the learning sample library of this type of defect is updated, and the updated learning sample library is used to retrain the defect detection model; If the convergence of the defect detection model in identifying this type of defect is unqualified, the scoring mechanism of the defect detection model will be updated or adjusted.

Further, retraining the defect detection model includes: determining test images in which the defect category label does not match the classification result, and the defect detection model uses a deep learning algorithm to test the defect category label that does not match the classification result. Feature learning is performed on the image to obtain an optimized model.

According to another aspect of the present invention, a PCB defect detection method is provided, including the following steps: inputting the PCB image to be detected into a defect detection model that has completed training, wherein the defect detection model is completed using the evaluation method as described above. Evaluation; the defect detection model outputs the defect prediction result of the PCB image to be detected, if the probability value in the defect prediction result is higher than the demarcation point determined during the evaluation process for defining real defects and false positive defects. value, then the detection result output by the defect detection model is a real defect and the defect type of the defect prediction result; otherwise, the detection result output by the defect detection model is a false positive defect.

The beneficial effects brought by the technical solution provided by the present invention are as follows: a. According to the classification of the prediction results of the test images, the defect types with OK evaluation results and the defect types with unqualified evaluation results can be screened out, and only the defect types with unqualified evaluation results can be screened out. Retrain and learn the defect types to improve the efficiency of the model in completing training; b. Comprehensively evaluate the model's ability from two aspects: defect identification ability and convergence, to ensure that the model accepts high-quality, high-demand evaluation; c. To determine overkill and underkill The demarcated AI score threshold provides an objective reference and improves the recognition accuracy of the model.

In order to more clearly explain the embodiments of the present application or the technical solutions in the prior art, the drawings needed to be used in the description of the embodiments or the prior art will be briefly described below. It is obvious that the drawings in the following description are only some embodiments recorded in this application. For those of ordinary skill in the art, other embodiments can be obtained based on these drawings without exerting creative efforts. Picture attached.

Figure 1 is a schematic flow chart of a PCB defect detection model evaluation method provided by an exemplary embodiment of the present invention; Figure 2 is a schematic diagram of the first type of interface for classifying defect prediction results provided by an exemplary embodiment of the present invention; Figure 3 is a schematic diagram of a second type of interface for classifying defect prediction results provided by an exemplary embodiment of the present invention.

In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only These are some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts should fall within the scope of protection of the present invention.

It should be noted that the terms "first", "second", etc. in the description of the invention, the patent scope of the invention, and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. . It is to be understood that the data so used are interchangeable under appropriate circumstances so that the embodiments of the invention described herein are capable of being practiced in sequences other than those illustrated or described herein. Furthermore, the term "package "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, apparatus, product or equipment that includes a series of steps or units need not be limited to those steps or units expressly listed , but may include other steps or elements not expressly listed or inherent to such processes, methods, products or devices.

The present invention proposes an evaluation tool. In the software of the evaluation tool, defect images assigned AI scores (predicted probability values) are displayed, as shown in Figures 2 and 3. The tool can be simulated by adjusting the AI threshold. result. Not only can this software obtain statistics on underkill and overkill, it can also look at specific misidentified images and create a new set of training data that specifically targets the problematic (misidentified) images. . Such a collection can be used to retrain appropriately tuned AI models for more accurate recognition.

In one embodiment of the present invention, a method for evaluating a PCB defect detection model is provided, which is used to evaluate the defect prediction ability of a pre-trained defect detection model on PCB images. The defect detection model can predict the defects of PCB images. The prediction results include defect types and corresponding probability values, as shown in Figure 1. The evaluation method includes: pre-establishing a test image set, which includes multiple test images, each test image has a defect category label; Multiple test images in the image set are input into the defect detection model to be evaluated, and defect prediction results (including defect types and corresponding probability values) corresponding to the test images are obtained; according to the obtained defect prediction results Defect type, defect prediction result Classify the results; according to the classification results, output and display the test images predicted to be the same defect type, and the defect category label of each test image and the probability value in the corresponding defect prediction result are configured to be viewable; according to the test image The defect category labels and classification results of the image are used to evaluate the ability of the defect detection model to identify various defect types, and the probability values corresponding to the test images under the same category are compared to evaluate the convergence of the defect detection model in identifying this type of defect. Spend.

Specifically, for example, the test image set includes 1000 test images labeled short circuit and 800 test images labeled open circuit. These 1800 test images are input into the defect detection model to obtain 1800 defect prediction results, such as The defect prediction results corresponding to the 990 test images labeled short circuit are short circuit defect types and their respective probability values. Among the remaining 10 images, 6 were identified as open circuit defect types, and 4 were identified as other defect types; for example Among them, the defect prediction results corresponding to 750 test images labeled as open circuits are open circuit defect types and their respective probability values. Among the remaining 50 images, 15 were identified as short circuit defect types, and 35 were identified as other defect types.

Then, for example, the defect prediction results are classified according to the short circuit defect type in the preset results, and the 990 test images labeled as short circuit and the 15 test images identified as the short circuit defect type and labeled as open circuit are obtained. The display interface is as shown in Figures 2 and 3: the probability value in the defect prediction result is displayed in a local area on the corresponding test image (the upper left corner area in Figures 2 and 3); the test The defect category label of the test image is configured to be displayed in a pop-up window that appears when triggered (as shown in Figure 2). The triggering operation of the pop-up window includes clicking the corresponding test image and multiple-clicking the corresponding test image. One or more of the following: image, right-click on the corresponding test image, or hovering the cursor on the corresponding test image.

This way of visually displaying the probability value on the test image can visually see the overall situation of the predicted probability value, so as to determine an appropriate dividing score for defining real defects and false positive defects. The specific determination method is as follows The evaluation of convergence is explained in detail below.

97.5%。若以98%為預設的正確率閾值，則次缺陷檢測模型識別短路類型的缺陷的能力不合格，若以97%為預設的正確率閾值，則次缺陷檢測模型識別短路類型的缺陷的能力合格。 In this embodiment, the evaluation indicators include the ability to identify defects and the convergence of identifying defects: First, the ability to identify defects is introduced. Taking the above example, the classification accuracy rate is calculated based on the defect category label and classification result of the test image: There are 10 test images labeled as short circuit and 15 test images labeled as open circuit that were misidentified. Therefore, the calculation formula of the classification accuracy can be: 100%-(10+15)/(990+15)

97.5%. If 98% is used as the preset accuracy threshold, the secondary defect detection model’s ability to identify short-circuit type defects is unqualified. If 97% is used as the preset accuracy threshold, the secondary defect detection model’s ability to identify short-circuit type defects is insufficient. Qualified ability.

As for the convergence of identifying defects, taking the above example, for example, in the defect prediction results of 990 test images that were correctly classified as short circuits, the maximum value of the probability value was 0.9 and the minimum value was 0.5, although they were all correctly identified. The defect type is short circuit, but the overall estimated probability value is too different, or the variance of these 990 probability values is too large, indicating that the overall estimated probability value distribution is relatively discrete, then the defect detection model The convergence of the model to identify this type of defect is unqualified.

The purpose of evaluating the PCB defect detection model is to optimize the model: the embodiment of the present invention classifies the defect prediction results, and can effectively evaluate the model's ability to identify defects and the degree of convergence in identifying defects under each classification. In one embodiment , only if the evaluation results of both are qualified, the model is considered to have passed the evaluation under this classification (defect type). If the evaluation result is unqualified, the defect detection module is retrained, and the defect detection model focuses its learning attention on the type of defects whose recognition ability or convergence is unqualified. This benefit is huge: for example, the ability to identify short-circuit type defects and the degree of convergence are up to standard, but the ability and/or convergence degree to identify open-circuit type defects are not up to standard, then the strategy for optimizing the model is to focus the learning attention on identification. Regarding the image features of circuit break defects, the following method can be adopted: if the evaluation model's ability to identify circuit break defects is unqualified, the learning sample library for identifying circuit break defects is updated, and the updated learning sample library is used to re-use the defect detection model. training; or extract the test images whose defect category labels do not match the classification results during the above evaluation process, and the defect detection model uses a deep learning algorithm to perform feature learning on the test images whose defect category labels do not match the classification results. , to obtain the optimized model. That is, look at specific misrecognized images and create a new set of training data that is specific to the problematic (misrecognized) images. Such a collection can be used to retrain appropriately tuned AI models for more accurate recognition.

If the convergence degree of the defect detection model in identifying this type of defect is not satisfactory, grid, then the scoring mechanism of the defect detection model is updated or adjusted, which is equivalent to adjusting the model parameters or weight value allocation proportion, so that the dispersion of the predicted probability value is reduced.

When the defect detection model has qualified convergence in identifying this type of defect, the demarcation line for defining real defects and false positive defects can be determined based on the probability value corresponding to the test image under this classification and the preset rules. value, specifically the following methods can be used: find the minimum value of the probability value corresponding to the test image under the classification, and operate the minimum value with the preset difference threshold to obtain the demarcation score; or, calculate the demarcation score under the classification The average value of the probability values corresponding to the test images, and the average value is calculated with the preset difference threshold to obtain the demarcation score; or, the average value of the probability values corresponding to the test images under the classification is calculated. Sort, exclude the first several probability values and the last several probability values, calculate the average value of the remaining probability values, and calculate the average value of the remaining probability values with the preset difference threshold to obtain the demarcation score; wherein , the preset difference threshold is a positive number, a negative number or zero.

An evaluation device for a PCB defect detection model according to the present invention includes the following modules: a test sample module configured to establish a test image set, which includes a plurality of test images, each test image having a defect category label; prediction A module configured to input multiple test images in the test image set into a defect detection model to be evaluated, and obtain defect prediction results corresponding to the test images one-to-one; a classification module, which is configured to classify the defect prediction results according to the defect types in the defect prediction results of the prediction module; a display module, which is configured to output and display the prediction results according to the classification results of the classification module. Test images of the same defect type, and can display the defect category labels of each test image and the probability values in the corresponding defect prediction results; a recognition capability evaluation module configured to display the defect category labels and classification results of the test images , to evaluate the ability of the defect detection model to identify various defect types; a convergence evaluation module configured to compare the probability values corresponding to test images under the same category to evaluate the convergence of the defect detection model in identifying this type of defect Spend.

Further, the recognition ability evaluation module calculates the classification accuracy rate based on the defect category label and classification result of the test image. If the classification accuracy rate is lower than the preset accuracy threshold, the evaluation result of the recognition ability evaluation module is The defect detection model's ability to identify this type of defect is unqualified; the convergence evaluation module determines whether the difference between the maximum value and the minimum value of the probability value corresponding to the test image under the same category is greater than the preset difference threshold, or , determine whether the variance of the probability values corresponding to the test images under the same category is greater than the preset variance threshold. If so, the evaluation result of the convergence evaluation module is that the convergence of the defect detection model in identifying this type of defect is unqualified.

Further, the evaluation device of the PCB defect detection model further includes a demarcation score determination module, which is configured to determine the evaluation result of the convergence evaluation module. When the defect detection model has qualified convergence in identifying this type of defect, the demarcation score for defining real defects and false positive defects is determined based on the probability value corresponding to the test image under this classification and the preset rules.

The evaluation device embodiment of this PCB defect detection model has the same concept as the above-mentioned evaluation method embodiment, and the entire content of the above-mentioned evaluation method embodiment is incorporated into this evaluation device embodiment by reference.

According to yet another aspect of the present invention, a method for training a PCB defect detection model is provided. The evaluation method as described above is used to evaluate the currently trained defect detection model. If the ability of the defect detection model to identify all types of defects is evaluated, and the convergence of identifying all types of defects is qualified, then the defect detection model ends training; otherwise, the defect detection model is retrained, and the defect detection model focuses its learning attention on the identification ability or the convergence is unqualified of such defects.

If the ability of the defect detection model to identify this type of defect is not qualified, update the learning sample library of this type of defect, and use the updated learning sample library to retrain the defect detection model; or, determine the defect category label and For test images whose classification results do not match, the defect detection model uses a deep learning algorithm to perform feature learning on test images whose defect category labels do not match the classification results to obtain an optimized model.

If the evaluation of the defect detection model's convergence in identifying such defects is unsatisfactory, the scoring mechanism of the defect detection model is updated or adjusted.

According to another aspect of the present invention, a PCB defect detection method is provided The method includes the following steps: inputting the PCB image to be detected into the defect detection model that has completed training, wherein the defect detection model uses the evaluation method as described above to complete the evaluation; the defect detection model outputs the PCB to be detected Defect prediction results of the image, if the probability value in the defect prediction results is higher than the dividing score determined during the evaluation process to define real defects and false positive defects, then the detection result output by the defect detection model is true The defect and the defect type of the defect prediction result; otherwise, the detection result output by the defect detection model is a false positive defect. For example, comparing Figure 2 and Figure 3, a lower score (probability value) in Figure 2 indicates that the image in Figure 2 is more likely to have a false alarm defect, and a higher score (probability value) in Figure 3 indicates that the image in Figure 3 Like having real flaws is more likely.

It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that these entities or operations are mutually exclusive. any such actual relationship or sequence exists between them. Furthermore, the terms "comprises," "comprises" or any other variations thereof are intended to cover a non-exclusive inclusion such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also those not expressly listed other elements, or elements inherent to the process, method, article or equipment. Without further limitation, an element defined by the statement "comprises a..." does not exclude the presence of additional identical elements in a process, method, article, or device that includes the stated element.

The above are only specific embodiments of the present application. It should be pointed out that for Those of ordinary skill in the art can make several improvements and modifications without departing from the principles of the present application, and these improvements and modifications should also be regarded as the protection scope of the present application.

Claims (13)

An evaluation method for PCB defect detection models, used to evaluate the defect prediction ability of a pre-trained defect detection model on PCB images. The defect prediction results of the defect detection model on PCB images include defect types and corresponding probability values, Wherein, the evaluation method includes: pre-establishing a test image set, which includes multiple test images, each test image having a defect category label; inputting multiple test images in the test image set into the defect detection to be evaluated model to obtain defect prediction results that correspond to the test images one-to-one; classify the defect prediction results according to the defect types in the obtained defect prediction results; output and display tests predicted to be of the same defect type according to the classification results images, and the defect category labels of each test image and the probability values in the corresponding defect prediction results are configured to be viewable; according to the defect category labels and classification results of the test images, the defect detection model is evaluated to identify various defects type capabilities, and compare the probability values corresponding to test images under the same category to evaluate the convergence of the defect detection model in identifying this type of defect. The evaluation method of PCB defect detection model as described in claim 1, wherein the classification accuracy rate is calculated based on the defect category label and classification result of the test image. If the classification accuracy rate is lower than the preset accuracy threshold, then The defect detection model's ability to identify such defects is unqualified; or, If the difference between the maximum value and the minimum value of the probability value corresponding to the test image under the same category is greater than the preset difference threshold, or if the variance of the probability value corresponding to the test image under the same category is greater than the preset variance threshold , then the convergence degree of the defect detection model in identifying this type of defect is unqualified. The evaluation method of PCB defect detection model as described in claim 1, which further includes: if the evaluation result is unqualified, retraining the defect detection module, and focusing the learning attention on identification Defects such as unsatisfactory ability or convergence. The evaluation method of PCB defect detection model as described in claim 1, wherein, when the convergence degree of the defect detection model in identifying this type of defect is qualified, the probability value corresponding to the test image under the classification and the preset rules to determine the cut-off score used to define real defects and false positive defects. The evaluation method of PCB defect detection model as described in claim 4, wherein the demarcation score used to define real defects and false positive defects is determined in the following manner: finding the minimum probability value corresponding to the test image under the classification value, and calculate the minimum value and the preset difference threshold to obtain the demarcation score; or, calculate the average value of the probability values corresponding to the test images under the category, and calculate the average value and the preset difference threshold. Perform operations to obtain the demarcation score; or, sort the average probability values corresponding to the test images under the classification, exclude the first several probability values and the last several probability values, and calculate the average of the remaining probability values. value, and the demarcation score is obtained by operating the average value of the remaining probability values and a preset difference threshold; wherein the preset difference threshold is a positive number, a negative number, or zero. The evaluation method of PCB defect detection model as described in claim 1, wherein the defect category label of each test image and the probability value in the corresponding defect prediction result are configured to be viewable in the following manner: the defect prediction The probability value in the result is displayed in a local area on the corresponding test image; the defect category label of the test image is configured to be displayed in a pop-up window that appears when triggered, and the triggering operation of the pop-up window includes a single One or more of the following: click the corresponding test image, multi-click the corresponding test image, right-click the corresponding test image, or stay with the cursor on the corresponding test image. An evaluation device for a PCB defect detection model, which includes the following modules: a test sample module configured to establish a test image set, which includes a plurality of test images, each test image having a defect category label; a prediction module, It is configured to input multiple test images in the test image set into a defect detection model to be evaluated, and obtain defect prediction results corresponding to the test images one-to-one; a classification module, which is configured to operate as described Defect types in the defect prediction results of the prediction module are used to classify the defect prediction results; A display module configured to output and display test images predicted to be of the same defect type according to the classification results of the classification module, and to be able to display the defect category label of each test image and the probability in the corresponding defect prediction result value; the recognition capability evaluation module, which is configured to evaluate the ability of the defect detection model to identify various defect types based on the defect category labels and classification results of the test image; the convergence evaluation module, which is configured to compare the same classification The probability value corresponding to the test image is used to evaluate the convergence degree of the defect detection model in identifying this type of defect. The evaluation device of PCB defect detection model according to claim 7, wherein the recognition capability evaluation module calculates the classification accuracy rate based on the defect category label and classification result of the test image. If the classification accuracy rate is lower than the preset the accuracy threshold, then the evaluation result of the recognition ability evaluation module is that the defect detection model's ability to identify this type of defect is unqualified; the convergence evaluation module determines the maximum value of the probability value corresponding to the test image under the same category and Whether the difference between the minimum values is greater than the preset difference threshold, or whether the variance of the probability values corresponding to the test images under the same category is greater than the preset variance threshold, if so, the evaluation result of the convergence evaluation module is as described The convergence of the defect detection model in identifying this type of defects is unqualified. The evaluation device for a PCB defect detection model as described in claim 7, further comprising a demarcation score determination module configured to use the evaluation result of the convergence evaluation module to identify the type of defect for the defect detection model. Convergence is qualified In this case, based on the probability value corresponding to the test image under the classification and the preset rules, the dividing score used to define the real defect and the false alarm defect is determined. A training method for a PCB defect detection model, wherein the currently trained defect detection model is evaluated using the evaluation method described in any one of claims 1 to 6. If the defect detection model is evaluated to identify all types of defects, If the ability and convergence of identifying all types of defects are qualified, the defect detection model will end the training; otherwise, the defect detection model will be retrained, and the defect detection model will focus its learning attention on the insufficiency of identification ability or convergence. Qualified defects of this type. The training method as described in claim 10, wherein if the ability of the defect detection model to identify this type of defect is not qualified, the learning sample library of this type of defect is updated, and the updated learning sample library is used to detect the defect. The model is retrained; if the evaluation of the defect detection model's convergence in identifying this type of defect is unqualified, the scoring mechanism of the defect detection model is updated or adjusted. The training method according to claim 10, wherein retraining the defect detection model includes: determining test images whose defect category labels do not match the classification results, and the defect detection model uses a deep learning algorithm to Test images whose category labels do not match the classification results are subjected to feature learning to obtain an optimized model. A PCB defect detection method, which includes the following steps: The PCB image to be detected is input into the defect detection model that has completed training, wherein the defect detection model uses the evaluation method as described in claim 4 or 5 to complete the evaluation; the defect detection model outputs the PCB image to be detected If the probability value in the defect prediction result is higher than the cutoff value determined during the evaluation process to define real defects and false positive defects, then the detection result output by the defect detection model is a real defect. and the defect type of the defect prediction result; otherwise, the detection result output by the defect detection model is a false positive defect.