Review method and device of image data detection result, electronic equipment and medium

Abstract

The application discloses a kind of image data detection result's review method, device, electronic equipment and medium, the method includes: constructing BP neural network classifier, initializes BP neural network classifier, trains BP neural network classifier, corrects the connection weight between each node, until BP neural network classifier reaches convergence state;For the original feature of image data detection result that visual inspection equipment outputs is enhanced processing, obtains three-dimensional feature vector, wherein, for area weight factor, for energy density feature, area weight coefficient, original feature includes: the area value of out-of-tolerance point, energy value and the region number of coordinate position location;The above three-dimensional feature vector is input into the BP neural network classifier trained to realize the review of image data detection result that visual inspection equipment outputs, and the scheme can greatly filter out false detection, reduce the false detection rate of detection result.

Description

Technical Field

[0001] This invention relates to the field of image processing, and more specifically, to a method, apparatus, electronic device, and medium for verifying image data detection results. Background Technology

[0002] Currently, in the manufacturing processes of high-end printed materials, specialty printed materials, PCBs, and LCDs, machine vision-based automated inspection equipment is widely used for online product inspection, thereby better serving their quality management systems. Typically, manufacturers use the inspection results from these devices to guide subsequent rejection, re-inspection, or other quality control measures.

[0003] However, these test results often include false positives or results that do not meet the quality rejection criteria. In other words, not all test results are truly defective; some are good products, which are usually referred to as false positives. A high false positive rate will reduce the efficiency of quality management using these test results or increase production costs.

[0004] Therefore, how to automatically perform secondary verification calculations on the relevant information of these test results in order to filter out false detections to a great extent and reduce the false detection rate of test results has become an urgent problem to be solved. Summary of the Invention

[0005] The main objective of this invention is to disclose a method, apparatus, electronic device, and medium for verifying image data detection results, so as to at least solve the problems in related technologies, such as how to perform automatic secondary verification calculations on relevant information of detection results, so as to filter out false detections to a great extent and reduce the false detection rate of detection results, and the lack of corresponding technical solutions.

[0006] According to one aspect of the present invention, a method for verifying image data detection results is provided.

[0007] The image data detection result verification method provided by the present invention includes: constructing a backpropagation (BP) neural network classifier, initializing the BP neural network classifier, training the BP neural network classifier, correcting the connection weights between nodes until the BP neural network classifier reaches a convergence state; and enhancing the original features of the image data detection result output by the visual inspection device to obtain a three-dimensional feature vector. ,in, Area weighting factor As a characteristic of energy density, Regional weighting coefficients, the aforementioned original features include: the area value of out-of-poor points. Energy value and the area number where the coordinates are located. The above three-dimensional feature vectors The data is input into the trained BP neural network classifier to verify the detection results of the image data output by the visual inspection device.

[0008] The above construction of the BP neural network classifier, and the initialization of the above BP neural network classifier, include: Determine the weights from the input layer to the hidden layer ,in The number of nodes in the input layer. This represents the number of hidden layer nodes. Generate random numbers in the range [0, 1], where λ is the weight adjustment coefficient; Determine the weights from the hidden layer to the output layer ,in This represents the number of nodes in the output layer. Perform bias term initialization and determine the input additional node bias. Hidden layer bias Output layer bias μ is the bias fine-tuning coefficient.

[0009] The number of hidden layer nodes was calculated using the following method.

[0010] ; in, Input the number of layers. Number of output layer nodes The total number of training samples, This means that when the sample size exceeds 10,000, it will be calculated as 10,000. Set the maximum number of hidden layer nodes to 128.

[0011] The training of the aforementioned neural network classifier includes: Update the learning rate in the following ways : ,in, , This represents the change in error between adjacent iterations. For the first The total error of the iterations; Update weights in the following ways : ,in For the error gradient, For the input vector, Momentum factor It is a weak regularization coefficient; The cross-entropy loss function is used to measure the difference between the model's predictions and the true labels. Where L is the value of the cross-entropy loss function. For real labels, This is the model's predicted output.

[0012] The above-mentioned training of the BP neural network classifier, adjusting the connection weights between nodes until the BP neural network classifier reaches convergence, also includes: repeatedly using each training sample in the sample set, employing a stratified random sampling method. Based on the ratio r of defective to non-defective samples, samples are drawn from both classes in each iteration to ensure a stable class distribution of the input samples. Introducing sample weighting factors Dynamically adjust the contribution weights of different samples during the training process. ,in For the first The true label of each sample κ represents the model's predicted value, and κ is the weight adjustment coefficient.

[0013] The above refers to the three-dimensional feature vectors. The data is input into the trained BP neural network classifier to verify the detection results of the image data output by the visual inspection device, including: The output vector of the above BP neural network classifier is then normalized: ,in, The original output value of the output layer. The normalized predicted probability, where k is either 1 or 2; Calculate class confidence Set a baseline confidence threshold : if If so, it is determined to be a genuine defect; if If so, the image data detection result output by the aforementioned visual inspection device is determined to be a false detection; if If the sample is not found, it will be marked as a suspected sample and transferred to the manual review process.

[0014] The above refers to the three-dimensional feature vectors. The data is input into the trained BP neural network classifier to verify the detection results of the image data output by the visual inspection device, including: The area weighting factor is determined in the following way. : ,in This represents the largest area exceeding tolerance in the historical testing of this type of product. The mean area of ​​out-of-range deviation. Standard deviation ; Energy density characteristics are determined using the following methods. : ,in Avoid distortion of energy characteristics when the area is too small; The regional weight coefficients are determined in the following manner. : .

[0015] According to another aspect of the present invention, an apparatus for verifying image data detection results is provided.

[0016] The image data detection result verification device according to the present invention includes: a construction and training module for constructing a BP neural network classifier, initializing the BP neural network classifier, training the BP neural network classifier, and correcting the connection weights between nodes until the BP neural network classifier reaches a convergent state; and an enhancement processing module for enhancing the original features of the image data detection result output by the visual inspection device to obtain a three-dimensional feature vector. ,in, Area weighting factor As a characteristic of energy density, Regional weighting coefficients, the aforementioned original features include: the area value of out-of-poor points. Energy value and the area number where the coordinates are located. The verification module is used to process the above three-dimensional feature vectors. The data is input into the trained BP neural network classifier to verify the detection results of the image data output by the visual inspection device.

[0017] According to another aspect of the present invention, an electronic device is provided.

[0018] The electronic device according to the present invention includes: a processor and a memory, wherein the memory is configured to store executable instructions of the processor, wherein the processor is configured to execute the image data detection result verification method described in any one of the preceding claims by executing the executable instructions.

[0019] According to another aspect of the present invention, a computer-readable storage medium is provided.

[0020] According to the computer-readable storage medium of the present invention, the computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform the image data detection result verification method described in any of the preceding claims.

[0021] The present invention provides a method, apparatus, electronic device, and medium for verifying image data detection results. A backpropagation (BP) neural network classifier is constructed, initialized, and trained. The connection weights between nodes are adjusted until the classifier converges. The original features of the image data detection results output by the visual inspection device are enhanced by processing the aforementioned three-dimensional feature vectors. The data is input into the trained BP neural network classifier to verify the image data detection results output by the visual inspection device. This application provides an automatic secondary verification scheme for image data detection results, which can greatly filter out false detections, reduce the false detection rate of the detection results, and improve the model classification accuracy by introducing feature fusion factors and simple correction terms. Attached Figure Description

[0022] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0023] Figure 1 This is a flowchart of a method for verifying image data detection results according to an embodiment of the present invention; Figure 2 This is a structural block diagram of an image data detection result verification device according to an embodiment of the present invention; Figure 3 This is a structural block diagram of an electronic device according to a preferred embodiment of the present invention. Detailed Implementation

[0024] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] The specific implementation of the present invention will now be described in detail with reference to the accompanying drawings.

[0026] According to an embodiment of the present invention, a method for verifying image data detection results is provided.

[0027] Figure 1 This is a flowchart of a method for verifying image data detection results according to an embodiment of the present invention. Figure 1 As shown, the verification method for the image data detection results includes: Step S101: Construct a BP neural network classifier, initialize the BP neural network classifier, train the BP neural network classifier, and correct the connection weights between each node until the BP neural network classifier reaches a convergent state. Step S102: Enhance the original features of the image data detection results output by the visual inspection device to obtain a three-dimensional feature vector. ,in, Area weighting factor As a characteristic of energy density, Regional weighting coefficients, the aforementioned original features include: the area value of out-of-poor points. Energy value and the area number where the coordinates are located. ; Step S103: Transfer the above three-dimensional feature vectors The data is input into the trained BP neural network classifier to verify the detection results of the image data output by the visual inspection device.

[0028] use Figure 1 The method for verifying the image data detection results shown involves constructing a BP neural network classifier, initializing the BP neural network classifier, training the BP neural network classifier, and correcting the connection weights between nodes until the BP neural network classifier reaches convergence. The original features of the image data detection results output by the visual inspection device are enhanced by processing the aforementioned three-dimensional feature vectors. The data is input into the trained BP neural network classifier to verify the image data detection results output by the visual inspection device. This application provides an automatic secondary verification scheme for image data detection results, which can greatly filter out false detections, reduce the false detection rate of the detection results, and improve the model classification accuracy by introducing feature fusion factors and simple correction terms.

[0029] In step S101 above, constructing and initializing the BP neural network classifier may further include the following processing: Determine the weights from the input layer to the hidden layer ,in The number of nodes in the input layer. This represents the number of hidden layer nodes. Generate random numbers in the range [0, 1], where λ is the weight adjustment coefficient; Determine the weights from the hidden layer to the output layer ,in This represents the number of nodes in the output layer. Perform bias term initialization and determine the input additional node bias. Hidden layer bias Output layer bias μ is the bias fine-tuning coefficient.

[0030] Among them, the number of hidden layer nodes mentioned above It can be calculated in the following way: ; in: Input the number of layers. Number of output layer nodes The total number of training samples, This means that when the sample size exceeds 10,000, it will be calculated as 10,000. Set the maximum number of hidden layer nodes to 128.

[0031] In the optimized implementation process, a three-layer BP neural network classifier model can be established. The traditional random initialization method is improved by adopting an adaptive weight initialization strategy: Input layer to hidden layer weights : ,in This represents the number of nodes in the input layer. For example, if the number of nodes in the input layer is 3, This represents the number of hidden layer nodes. Generate random numbers in the range [0,1], where λ is the weight adjustment coefficient. For example, λ can be set to 1.1.

[0032] Hidden layer to output layer weights : ,in This represents the number of nodes in the output layer. For example, the number of nodes in the output layer is 2.

[0033] Bias term initialization: Input layer additional node bias Hidden layer bias Output layer bias μ is the bias fine-tuning coefficient. For example, μ can be set to 1.05.

[0034] The number of hidden layer nodes can be adaptively selected, and the number of hidden layer nodes is calculated using an optimization empirical formula:

[0035] For example, (Number of nodes in the input layer) (Number of output layer nodes) The total number of training samples; This means that when the number of samples exceeds 10,000, it is calculated as 10,000. This avoids excessively amplifying the number of nodes when the sample size is large; outer layer Setting the maximum number of hidden layer nodes to 128 ensures that the model has sufficient fitting ability while avoiding excessive nodes that would lead to low training efficiency and increased risk of overfitting, thus adapting to the training needs of hundreds of thousands of samples.

[0036] Before training the aforementioned BP neural network classifier, training samples need to be manually labeled. Specifically, each detection result is manually judged as either a defect or a non-defect. First, the detection results from the equipment over a period of time, including images and related information, need to be presented to the judges one by one. The judges then determine, based on the actual judgment criteria, whether the detection result is a true defect or a false detection (non-defect). It should be noted that the more data and the broader the coverage, the more effective the training of the neural network model will be.

[0037] Machine vision inspection equipment determines the presence of defects by comprehensively considering factors such as the area, energy, and coordinate position of the out-of-tolerance points. Generally, the larger the out-of-tolerance area, the higher the out-of-tolerance energy value, and the more likely the coordinate position is to be within a key quality monitoring area of ​​the product, the more likely it is to be identified as a defect. When these inspection data exceed a set threshold, the inspection equipment will determine the existence of the defect and store the corresponding defect information and defect thumbnail for future use.

[0038] In performing secondary verification calculations on defective image data, this application first selects the area value of the out-of-tolerance points. Energy value The area number where the coordinates are located These three dimensions serve as the basic features. To improve the model's classification accuracy, the basic features need to be enhanced by introducing a feature fusion factor and a simple correction term: Area weighting factor : Dynamically adjusted based on historical product data. ,in, This represents the largest area exceeding tolerance in the historical testing of this type of product. The mean area of ​​out-of-range deviation. Standard deviation (Avoid denominators of 0).

[0039] Energy density characteristics Combining the relationship between area and energy, ,in, Avoid distortion of energy characteristics when the area is too small.

[0040] Regional weighting coefficient : Set according to the regional quality requirement level.

[0041]

[0042] The final optimized input feature vector is: The feature dimension remains 3-dimensional, and it is compatible with the original model structure.

[0043] Therefore, during training, the number of nodes in the input layer is the same as the feature dimension of the samples participating in training; the number of nodes in the output layer is the same as the number of classifications. Each detection result is a training sample. Each training sample, after feature enhancement, has a feature dimension of 3, forming the input vector. Therefore, the number of nodes in its input layer is 3. Since the classification result is either defect or non-defect, i.e., the number of categories is 2, the number of nodes in the output layer is also 2. The manual judgment result of this detection result forms an output vector, where defect is... Non-defect is .

[0044] In step S101 above, training the neural network classifier may further include the following processing: Update the learning rate in the following ways : ,in, , This represents the change in error between adjacent iterations. For the first The total error of the iterations; Update weights in the following ways : ,in For the error gradient, For the input vector, Momentum factor It is a weak regularization coefficient; The cross-entropy loss function is used to measure the difference between the model's predictions and the true labels. Where L is the value of the cross-entropy loss function. For real labels, This is the model's predicted output.

[0045] In the optimal implementation process, an adaptive learning rate BP algorithm with a driving term can be used, with the addition of a simple regularization term: Learning rate update formula: ,in , This represents the change in error between adjacent iterations. For the first The total error of the iterations.

[0046] Weight update formula: ,in For the error gradient, For the input vector, Momentum factor It is a weak regularization coefficient to suppress overfitting.

[0047] Error function: Cross-entropy loss function is used. ,in For real labels, This is the model's predicted output.

[0048] In step S101, the BP neural network classifier is trained, and the connection weights between nodes are adjusted until the BP neural network classifier reaches convergence. The following processing may also be included: Each training sample in the sample set is used multiple times in a stratified random sampling manner: Based on the ratio r of defective to non-defective samples, samples are drawn from both classes in each iteration to ensure a stable class distribution of the input samples. Introducing sample weighting factors Dynamically adjust the contribution weights of different samples during the training process. ,in For the first The true label of each sample κ represents the model's predicted value, and κ is the weight adjustment coefficient.

[0049] In the optimal implementation process, during the training and convergence of the neural network classifier, each training sample in the sample set is used thousands or even tens of thousands of times. A stratified random sampling method can be adopted. Based on the ratio of defective to non-defective samples (set as follows): Each iteration draws samples from both classes to ensure a stable class distribution of the input samples.

[0050] Introducing sample weighting factors , The contribution weights of different samples during training are dynamically adjusted. For the first The true label of each sample The value is the model prediction, and 0.5 is the weight adjustment coefficient, which moderately strengthens the training weights of difficult-to-distinguish samples.

[0051] In step S102 above, after the neural network classifier is trained, the original features (area values) of the detection results output by the visual inspection device are used. Energy value Area code The enhanced feature vector is calculated using the same feature enhancement formula mentioned above. The input is then fed into the trained neural network classifier.

[0052] Similarly, the area weighting factor can be determined in the following way. : ,in This represents the largest area exceeding tolerance in the historical testing of this type of product. The mean area of ​​out-of-range deviation. Standard deviation ; Similarly, energy density characteristics can be determined in the following ways. : ,in Avoid distortion of energy characteristics when the area is too small; Similarly, the regional weighting coefficients can be determined in the following ways. : .

[0053] In step S103 above, the above three-dimensional feature vectors are... The input to the trained BP neural network classifier, in order to verify the detection results of the image data output by the visual inspection device, can further include the following processing: The output vector of the above BP neural network classifier is then normalized: ,in, The original output value of the output layer. The normalized predicted probability, where k is either 1 or 2; Calculate class confidence Set a baseline confidence threshold : if If so, it is determined to be a genuine defect; if If so, the image data detection result output by the aforementioned visual inspection device is determined to be a false detection; if If the sample is not found, it will be marked as a suspected sample and transferred to the manual review process.

[0054] In the preferred implementation process, this application introduces a confidence threshold and a simple dynamic adjustment mechanism: The model output vector is then normalized using Softmax. ,in The original output value of the output layer. The normalized prediction probability is represented by k, where k is either 1 or 2.

[0055] Calculate class confidence For example, a baseline confidence threshold can be set. : like If so, it is determined to be a real defect (the output vector is equivalent to...). ); like If the image data output by the visual inspection device is then determined to be a false detection (the output vector is equivalent to...), then the detection result is considered a false detection. ); like If the sample is not found, it will be marked as a suspected sample and transferred to the manual review process.

[0056] Since the BP neural network classifier is a non-linear classifier, combined with feature enhancement, improved training algorithms and confidence decision mechanisms, it can effectively remove a large proportion of false detection information from machine vision online inspection equipment, significantly reduce the false detection rate, and improve the efficiency and effectiveness of quality management in the production process.

[0057] According to an embodiment of the present invention, a device for verifying image data detection results is provided.

[0058] Figure 2 This is a structural block diagram of an image data detection result verification device according to an embodiment of the present invention. Figure 2 As shown, the verification device for the image data detection result includes: a construction and training module 20, used to construct a BP neural network classifier, initialize the BP neural network classifier, train the BP neural network classifier, and correct the connection weights between nodes until the BP neural network classifier reaches a convergence state; and an enhancement processing module 22, used to enhance the original features of the image data detection result output by the visual inspection device to obtain a three-dimensional feature vector. ,in, Area weighting factor As a characteristic of energy density, Regional weighting coefficients, the aforementioned original features include: the area value of out-of-poor points. Energy value and the area number where the coordinates are located. The verification module 24 is used to process the above three-dimensional feature vectors. The data is input into the trained BP neural network classifier to verify the detection results of the image data output by the visual inspection device.

[0059] Figure 2 The image data detection result verification device shown includes a training module 20 that constructs a BP neural network classifier, initializes the BP neural network classifier, trains the BP neural network classifier, and corrects the connection weights between nodes until the BP neural network classifier reaches convergence; an enhancement processing module 22 enhances the original features of the image data detection result output by the visual inspection device; and a verification module 24 converts the three-dimensional feature vector... The data is input into the trained BP neural network classifier to verify the image data detection results output by the visual inspection device. This application provides an automatic secondary verification scheme for image data detection results, which can greatly filter out false detections, reduce the false detection rate of the detection results, and improve the model classification accuracy by introducing feature fusion factors and simple correction terms.

[0060] It should be noted that the verification device for the above image data detection results can be found in the corresponding documentation. Figure 1 The relevant descriptions and effects in the illustrated embodiments are for understanding purposes only and will not be repeated here.

[0061] According to embodiments of the present invention, an electronic device is also provided.

[0062] Figure 3 This is a structural block diagram of an electronic device according to a preferred embodiment of the present invention. Figure 3 As shown, the electronic device according to the present invention includes: a memory 30 and a processor 32, wherein the memory 30 is used to store computer execution instructions of the processor 32; and the processor 32 is configured to execute the image data detection result verification method described above by executing the executable instructions.

[0063] Processor 32 can be a central processing unit (CPU). Processor 32 can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or combinations of the above types of chips.

[0064] The memory 30, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the program instructions / modules corresponding to the image data detection result verification method in this embodiment of the invention. The processor executes various functional applications and data processing by running the non-transitory software programs, instructions, and modules stored in the memory.

[0065] The memory 30 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created by the processor, etc. Furthermore, the memory may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory 30 may optionally include memory remotely constructed relative to the processor, which can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0066] One or more of the above modules are stored in the above memory 30, and when executed by the above processor 32, they perform the following: Figure 1 The method for verifying the image data detection results in the embodiment.

[0067] For specific details regarding the aforementioned electronic devices, please refer to the relevant documentation. Figure 1 The relevant descriptions and effects in the illustrated embodiments are for understanding purposes only and will not be repeated here.

[0068] According to embodiments of the present invention, a computer-readable storage medium is also provided. A computer program is stored thereon, which, when executed by a processor, implements the method for verifying the image data detection results provided in any of the preceding embodiments.

[0069] Computer-readable storage media can be memory, for example, volatile memory or non-volatile memory, or both. Non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDRSDRAM), Enhanced Synchronous DRAM (ESDRAM), Sync Link DRAM (SLDRAM), and Direct Rambus RAM (DRRAM).

[0070] In summary, the embodiments provided by this invention offer a scheme for verifying image data detection results. A BP neural network classifier model is constructed, and the traditional random initialization method is improved by employing an adaptive weight initialization strategy. During the secondary verification calculation of defective image data, the area value of the out-of-error points is selected. Energy value The area number where the coordinates are located Using these three dimensions as basic features, to improve the model's classification accuracy, the basic features are enhanced by introducing feature fusion factors and simple correction terms. An adaptive learning rate BP algorithm with a driving term is adopted, and a simple regularization term is added. Thus, the core improvement of the training calculation and weight correction of the BP neural network classifier lies in the aforementioned feature enhancement, initialization optimization, and improved BP algorithm. Since the neural network classifier is a non-linear classifier, combining feature enhancement, improved training algorithms, and confidence decision mechanisms can effectively remove a large proportion of false detection information from machine vision online inspection equipment, significantly reduce the false detection rate, and improve the efficiency and effectiveness of quality management in the production process.

[0071] The above specific embodiments further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solution of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for verifying image data detection results, characterized in that, include: Construct a backpropagation BP neural network classifier, initialize the BP neural network classifier, train the BP neural network classifier, and correct the connection weights between each node until the BP neural network classifier reaches a convergent state. The original features of the image data detection results output by the visual inspection device are enhanced to obtain a three-dimensional feature vector. ,in, Area weighting factor As a characteristic of energy density, Regional weighting coefficients, the original features including: area values ​​of out-of-poor points Energy value and the area number where the coordinates are located. ; The three-dimensional feature vector The data is input into the trained BP neural network classifier to verify the detection results of the image data output by the visual inspection device.

2. The method according to claim 1, characterized in that, Constructing a BP neural network classifier, and initializing the BP neural network classifier includes: Determine the weights from the input layer to the hidden layer ,in The number of nodes in the input layer. This represents the number of hidden layer nodes. Generate random numbers in the range [0, 1], where λ is the weight adjustment coefficient; Determine the weights from the hidden layer to the output layer ,in This represents the number of nodes in the output layer. Perform bias term initialization and determine the input additional node bias. Hidden layer bias Output layer bias μ is the bias fine-tuning coefficient.

3. The method according to claim 2, characterized in that, The number of hidden layer nodes is calculated in the following way. ; in, Input the number of layer nodes. Number of output layer nodes The total number of training samples, This means that when the sample size exceeds 10,000, it will be calculated as 10,000. Set the maximum number of hidden layer nodes to 128.

4. The method according to claim 1, characterized in that, Training the neural network classifier includes: Update the learning rate in the following ways : ,in, , This represents the change in error between adjacent iterations. For the first The total error of the iterations; Update weights in the following ways : ,in For the error gradient, For the input vector, Momentum factor It is a weak regularization coefficient; The cross-entropy loss function is used to measure the difference between the model's predictions and the true labels. Where L is the value of the cross-entropy loss function. For real labels, This is the model's predicted output.

5. The method according to claim 1, characterized in that, Training the BP neural network classifier, adjusting the connection weights between nodes, until the BP neural network classifier reaches convergence, further includes: Each training sample in the sample set is used multiple times in a stratified random sampling manner: Based on the ratio r of defective to non-defective samples, samples are drawn from both classes in each iteration to ensure a stable class distribution of the input samples. Introducing sample weighting factors Dynamically adjust the contribution weights of different samples during the training process. ,in For the first The true label of each sample κ represents the model's predicted value, and κ is the weight adjustment coefficient.

6. The method according to claim 1, characterized in that, The three-dimensional feature vector The input is fed into the trained BP neural network classifier to verify the detection results of the image data output by the visual inspection device, including: The output vector of the BP neural network classifier is then normalized. ,in, The original output value of the output layer. The normalized predicted probability, where k is either 1 or 2; Calculate class confidence Set a baseline confidence threshold : if If so, it is determined to be a genuine defect; if If the image data output by the visual inspection device is determined to be a false detection, then the detection result is determined to be false. if If the sample is not found, it will be marked as a suspected sample and transferred to the manual review process.

7. The method according to claim 1, characterized in that, The area weighting factor is determined in the following way. : ,in This represents the largest area exceeding tolerance in the historical testing of this type of product. The mean area of ​​the out-of-range phenomenon. Standard deviation, ; Energy density characteristics are determined using the following methods. : ,in Avoid distortion of energy characteristics when the area is too small; The regional weight coefficients are determined in the following manner. : 。 8. A device for verifying image data detection results, characterized in that, include: A training module is constructed to build a BP neural network classifier, initialize the BP neural network classifier, train the BP neural network classifier, and correct the connection weights between nodes until the BP neural network classifier reaches a convergent state. The enhancement processing module is used to enhance the original features of the image data detection results output by the vision inspection device to obtain a three-dimensional feature vector. ,in, Area weighting factor As a characteristic of energy density, Regional weighting coefficients, the original features including: area values ​​of out-of-poor points Energy value and the area number where the coordinates are located. ; The verification module is used to verify the three-dimensional feature vector. The data is input into the trained BP neural network classifier to verify the detection results of the image data output by the visual inspection device.

9. An electronic device, characterized in that, include: A processor and a memory, the memory being used to store executable instructions of the processor, wherein the processor is configured to perform a method for verifying image data detection results according to any one of claims 1 to 7 by executing the executable instructions.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device on which the computer-readable storage medium is located to perform the verification method for the image data detection results as described in any one of 1 to 7.

Images