Non-welding bottle type container surface defect intelligent discrimination system based on magnetic powder detection

Abstract

The invention discloses a non-welding bottle type container surface defect intelligent discrimination system based on magnetic powder detection, relates to the technical field of intelligent discrimination, and aims to improve the problem of low discrimination accuracy caused by non-surface physical defects in magnetic powder detection. Screening a plurality of defect positions according to a gray value in the image, obtaining a defect area of each defect position, constructing a support vector machine model according to the defect areas and the gray value, calculating a category index and a misjudgment rate of each defect position, setting a classification threshold value, and performing classified display on each defect position. After obtaining the pixel intensity value of each pixel point in the image, calculating a complex frequency domain signal value, performing sampling adjustment on a camera in combination with a misjudgment rate, setting a similar range, screening out classification fuzzy positions from all defect positions, and marking the classification fuzzy positions; and obtaining the probability that each mark position has defects, calculating the average gray value of each mark position, and carrying out reclassification on the display categories of the mark positions by utilizing a logistic regression algorithm.

Description

Technical Field

[0001] The present invention relates to the field of intelligent identification technology, and more specifically, to an intelligent identification system for surface defects of non-welded bottle containers based on magnetic particle detection. Background Art

[0002] Intelligent identification technology is a method that uses artificial intelligence and machine learning algorithms to analyze and identify data. When applied to the surface defect identification system of non-welded bottle containers using magnetic particle inspection, intelligent identification technology can reduce labor costs and improve identification efficiency and accuracy.

[0003] The existing technology has the following deficiencies: In the past, when judging container surface defects, the defect location was identified and judged by comparing the grayscale values ​​of each pixel in the sampled image. When the grayscale value was low, the image was black, and the locations with low grayscale values ​​were screened out in the image as the basis for defect judgment. However, when non-surface physical defects such as dust appeared on the container surface in the sampled image, the grayscale value was also low, which reduced the judgment accuracy and affected the judgment efficacy. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides an intelligent surface defect identification system for non-welded bottle containers based on magnetic particle inspection. By adjusting the image sampling clarity and combining the area of ​​the position with lower grayscale value and the pixel intensity, the display category of each alternative defect position is output and detection modulation is performed to solve the problems raised in the above-mentioned background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions: The intelligent identification system for surface defects of non-welded bottle containers based on magnetic particle inspection includes a data acquisition module, a feature extraction module, a fuzzy detection module, and a critical processing module, with signal connections between the modules; The data acquisition module is used to collect images of non-welded bottle containers and transmit them to the feature extraction module to extract image features, obtain the pixel intensity value of each pixel in the image and transmit it to the fuzzy detection module. When receiving the marked position transmitted by the critical processing module, it calls historical data to transmit the probability of defects at each marked position to the critical processing module; After receiving the image of the non-welded bottle container, the feature extraction module detects the grayscale value in the image and selects multiple defect locations and obtains the defect area based on the preset grayscale threshold. The support vector machine model is constructed based on the defect area and grayscale value in the image to calculate the category index and error rate and set the classification threshold. The defect locations are classified and displayed using the classification threshold. The calculated error rate is passed to the fuzzy detection module, and each defect location and the corresponding category index of the defect location are passed to the critical processing module. After receiving the pixel intensity value of each pixel in the image, the blur detection module sets the intensity sequence and uses Laplace transform to calculate the complex frequency domain signal value. It then adjusts the sampling of the camera based on the complex frequency domain signal intensity value and the error rate input by the feature extraction module. The critical processing module sets a similarity range, selects multiple classified fuzzy positions from all defect positions based on the similarity range, and marks them. The marked positions are sent to the data acquisition module. After receiving the probability of defects at each marked position, the average grayscale value of each marked position is calculated and the display category of the marked position is reclassified.

[0006] In a preferred embodiment, the categories of classification display are relevant display, non-relevant display and pseudo display. The relevant display is the magnetic trace display formed by the leakage magnetic field generated by the defect during magnetic particle inspection and the magnetic powder adsorbs the magnetic powder; the non-relevant display is the magnetic trace display formed by the leakage magnetic field generated by the cross-section change or the material magnetic permeability change during magnetic particle inspection and the magnetic powder adsorbs the magnetic powder; the pseudo display is the magnetic trace display formed by the non-leakage magnetic field adsorbing the magnetic powder.

[0007] In a preferred embodiment, the feature extraction module compares the grayscale value of each pixel in the non-welded bottle container image acquisition image with a preset grayscale threshold. When the grayscale value of adjacent pixels is lower than the preset grayscale threshold, the adjacent pixels are merged to obtain the defect location. After obtaining multiple defect locations, the defect area is calculated based on the pixel size and the number of pixels corresponding to the defect location. The average value of the grayscale values ​​of each defect position is calculated as the defect grayscale value of the corresponding defect position, and the defect grayscale value and defect area of ​​each defect position are merged into a grayscale value data set and an area data set respectively.

[0008] In a preferred embodiment, the feature extraction module integrates the gray value dataset and the area dataset to construct a support vector machine model to calculate the category index of each defect location. The specific steps are as follows: Step A1: Input the grayscale value dataset and the area dataset as analysis features into the support vector machine; Step A2: Select the Sigmoid function as the kernel function to transform the input features. The Sigmoid function is: ,in, is the kernel function result after feature conversion, is the hyperbolic tangent function, is a hyperparameter, c is a control constant, x and y are two data corresponding to the defect location in the gray value dataset and the area dataset respectively; the kernel function result obtained by transforming the corresponding data features in the two datasets of each defect location is used as the category index of the corresponding defect location; Step A3: Use the percentile method to set two adjustment coefficients to converge and classify the kernel function results. The formula is: , ,in, and are two adjustment coefficients, and are the convergence results of the two kernel functions, is the sum average of all category indices; the two kernel function results after convergence are used as classification thresholds; Step A4: Determine the defect category. Compare the category index of each defect location with the two classification thresholds. Select the maximum value between the two classification thresholds and mark it as a. Select the minimum value between the two classification thresholds and mark it as b. Defect locations with a category index exceeding a are marked as 1. Defect locations with a category index between a and b are marked as 2. Defect locations with a category index below b are marked as 3. Step A5: Output the judgment result. When the defect position is marked as 1, the defect position is marked as a related display and output; when the defect position is marked as 2, the defect position is marked as a non-related display and output; when the defect position is marked as 3, the defect position is marked as a pseudo display and output; the category index, display category and classification threshold of each defect position are output.

[0009] In a preferred embodiment, the feature extraction module accesses the image library to obtain the magnetic trace results of each defect position in the image, and uses the magnetic trace results of each defect position as a reference data set. The specific steps for calculating the misjudgment rate using the support vector machine model based on the display categories of each defect location and the reference data set are as follows: Step B1: Data preprocessing: sort and label the display categories of each defect location according to the data in the reference data set, and merge the sorted display categories into a discriminant data set; Step B2: Set the evaluation rules and compare the data in the discrimination dataset with the corresponding data in the control dataset according to the labels. If the displayed categories are the same, the defect location is marked as correct; if the displayed categories are different, the defect location is marked as incorrect. Step B3: Calculate the false positive rate. Randomly divide the discrimination data set and the control data set into n subsets of equal size and correspond them one to one according to the serial number. Select one subset as the test set each time, and the remaining n-1 subsets as the training set. Count the number of incorrect labels in the test set, and take the ratio of the number of incorrect labels to the total amount of test set data as the error rate of the test set. After calculating the error rate of the test set, randomly select a subset from the training set and calculate the error rate before removing it. Take the average of all the calculated error rates and output it as the false positive rate.

[0010] In a preferred embodiment, after receiving the pixel intensity of each pixel in the image, the blur detection module sets an intensity sequence for the pixel points in the defect area and calculates the complex frequency domain signal value using Laplace transform and performs normalization and numerical compression to obtain the complex frequency domain signal compression value; The geometric mean method is used to calculate the adjustment correction value based on the comprehensive complex frequency domain signal compression value and the error rate: , where T is the adjustment correction value, f is the complex frequency domain signal compression value, and q is the error rate.

[0011] In a preferred embodiment, the adjustment correction value is compared with a preset adjustment threshold. When the adjustment correction value exceeds the adjustment threshold, no adjustment is performed. When the adjustment correction value is lower than the adjustment threshold, the camera exposure time and rotation speed are adjusted to the corresponding pre-set precise shooting gear parameters. The inverse of the adjustment correction value is taken to obtain an adjustment ratio t, and the shooting area of ​​the shooting camera is calculated based on the adjustment ratio: ,in, is the adjusted sampling area, is the defect area at the defect location.

[0012] In a preferred embodiment, the critical processing module sets a similarity range, filters the category index of each defect area according to the similarity range, and thus filters the defect location, selects the minimum value from the two classification thresholds as the analysis threshold, processes the analysis threshold by a preset comparison difference to obtain a similarity range, adds or subtracts a comparison difference from the analysis threshold to obtain a similarity range, marks the defect area whose category index is within the similarity range, and sends the marked position to the data acquisition module; when the data acquisition module receives the marked position, it calls the historical data to pass the probability of defects occurring at each marked position into the critical processing module.

[0013] In a preferred embodiment, after receiving the probability of a defect occurring at a marked position, the criticality processing module counts the grayscale values ​​of each marked area and calculates the average value to obtain the average grayscale value of each area. The display category of the marked position is reclassified using a logistic regression algorithm based on the probability of a defect occurring at each marked position and the average grayscale value, and the logistic regression coefficient is calculated using the logistic regression algorithm: , where L is the logistic regression coefficient of the mark position, e is the natural base, z is the logistic regression parameter, and z is the sum of the probability of a defect occurring at the mark position and the average grayscale value; When the logistic regression coefficient exceeds the preset false display discrimination threshold, the suspicious mark is classified as non-relevant display; when the logistic regression coefficient is lower than the preset false display discrimination threshold, the suspicious mark is classified as false display.

[0014] The technical effects and advantages of the intelligent identification system for surface defects of non-welded bottle containers based on magnetic particle detection of the present invention are as follows: The present invention collects images of non-welded bottle containers, screens multiple defect locations according to the grayscale value in the image and obtains the defect area of ​​each defect location, calculates the category index and the error rate of each defect location according to the defect area and the grayscale value, and sets a classification threshold, and classifies and displays each defect location according to the classification threshold. The classification display can improve the discrimination accuracy and discrimination efficacy, and facilitates subsequent processing according to different classification displays. After obtaining the pixel intensity value of each pixel point in the image, the complex frequency domain signal value is calculated, and the camera is sampled and adjusted based on the complex frequency domain signal value and the error rate of each defect location. A similar range is set, and classified fuzzy locations are screened out from all defect locations for marking, the probability of defects occurring at each marked location is obtained, and the average grayscale value of each marked location is calculated to reclassify the display category of the marked location, thereby improving the defect classification accuracy of non-welded bottle containers and performing timely feedback processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a schematic diagram of the intelligent surface defect identification system for non-welded bottle containers based on magnetic particle testing according to the present invention.

[0016] Figure 2 This is a flow chart of the intelligent identification system for surface defects of non-welded bottle containers based on magnetic particle testing of the present invention. DETAILED DESCRIPTION

[0017] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0018] The present invention collects images of non-welded bottle containers, screens multiple defect locations according to the grayscale value in the image and obtains the defect area of ​​each defect location, calculates the category index and the error rate of each defect location according to the defect area and the grayscale value, sets a classification threshold, classifies and displays each defect location according to the classification threshold, obtains the pixel intensity value of each pixel point in the image and then calculates the complex frequency domain signal value, adjusts the camera sampling based on the complex frequency domain signal value and the error rate of each defect location, sets a similarity range, screens out classification fuzzy locations from all defect locations and marks them, obtains the probability of defects occurring at each marked location, calculates the average grayscale value of each marked location, and reclassifies the display category of the marked location, thereby improving the defect classification accuracy of non-welded bottle containers and performing timely feedback processing.

[0019] Example, an intelligent identification system for surface defects of non-welded bottle containers based on magnetic particle inspection, such as Figure 1As shown, it includes a data acquisition module, a feature extraction module, a fuzzy detection module and a critical processing module, and the signals between the modules are connected; The functions of each module are as follows: The data acquisition module is used to collect images of non-welded bottle containers and transmit them to the feature extraction module to extract image features, obtain the pixel intensity value of each pixel in the image and transmit it to the fuzzy detection module. When receiving the marked position transmitted by the critical processing module, it calls historical data to transmit the probability of defects at each marked position to the critical processing module; After receiving the image of the non-welded bottle container, the feature extraction module detects the grayscale value in the image and selects multiple defect locations and obtains the defect area based on the preset grayscale threshold. The support vector machine model is constructed based on the defect area and grayscale value in the image to calculate the category index and error rate and set the classification threshold. The location is classified into three display categories using the classification threshold. The calculated error rate is passed to the fuzzy detection module, and each defect location and the corresponding category index of the defect location are passed to the critical processing module. After receiving the pixel intensity value of each pixel in the image, the blur detection module sets the intensity sequence and uses Laplace transform to calculate the complex frequency domain signal value. It then adjusts the sampling of the camera based on the complex frequency domain signal intensity value and the error rate input by the feature extraction module. The critical processing module sets a similarity range, selects multiple classified fuzzy positions from all defect positions based on the similarity range, and marks them. The marked positions are sent to the data acquisition module. After receiving the probability of defects at each marked position, the average grayscale value of each marked position is calculated and the display category of the marked position is reclassified.

[0020] It should be noted that the three display categories are relevant display, irrelevant display and pseudo display. The relevant display is the magnetic trace display formed by the leakage magnetic field generated by defects during magnetic particle inspection, which absorbs magnetic powder, for example, surface cracks in non-welded bottle containers; the irrelevant display is the magnetic trace display formed by the leakage magnetic field generated by cross-section changes or material magnetic permeability changes during magnetic particle inspection, which absorbs magnetic powder, for example, the magnetic permeability between folds is inconsistent with the normal state due to high-temperature spinning during manufacturing; the pseudo display is the magnetic trace display formed by the non-leakage magnetic field which absorbs magnetic powder, for example, foreign matter such as wool and soil absorbed on the bottle body.

[0021] The data acquisition module saves the non-welded bottle container image uploaded by the camera and transmits it to the feature extraction module. The data acquisition module imports the saved non-welded bottle container image into the open source library to obtain the pixel intensity value of each pixel in the image and transmits it to the blur detection module. It should be noted that an open source library refers to a software library that is open to the public for users to view, use, modify, and distribute. In this example, it is used to obtain the pixel intensity value of each pixel in the image of a non-welded bottle container.

[0022] After receiving the non-welded bottle container image, the feature extraction module uses the edge detection algorithm to obtain the grayscale value of each pixel in the image. When processing the non-welded bottle container image, the grayscale value of each pixel is compared with the preset grayscale threshold. When the grayscale value of adjacent pixels is lower than the preset grayscale threshold, the adjacent pixels are merged to obtain the defect position. After obtaining multiple defect positions, the defect area is calculated based on the pixel size and the number of pixels corresponding to the defect position.

[0023] It should be noted that the grayscale threshold is set after professionals in this field confirm the display effect of the magnetic trace features in the image, and will not be elaborated here.

[0024] The feature extraction module calculates the average value of the grayscale value of each defect position as the defect grayscale value of the corresponding defect position, and merges the defect grayscale value and defect area of ​​each defect position into a grayscale value data set and an area data set respectively; The feature extraction module integrates the grayscale value dataset and the area dataset to construct a support vector machine model to calculate the category index of each defect location. The specific steps are as follows: Step A1: Input the grayscale value dataset and the area dataset as analysis features into the support vector machine; Step A2: Select the Sigmoid function as the kernel function to transform the input features. The Sigmoid function is: ,in, is the kernel function result after feature conversion, is the hyperbolic tangent function, is a hyperparameter, c is a control constant, x and y are two data corresponding to the defect location in the gray value dataset and the area dataset respectively; the kernel function result obtained by transforming the corresponding data features in the two datasets of each defect location is used as the category index of the corresponding defect location; Step A3: Use the percentile method to set two adjustment coefficients to converge and classify the kernel function results. The formula is: , ,in, and are two adjustment coefficients, and are the convergence results of the two kernel functions, is the sum average of all category indices; the two kernel function results after convergence are used as classification thresholds; Step A4: Determine the defect category. Compare the category index of each defect location with the two classification thresholds. Select the maximum value between the two classification thresholds and mark it as a. Select the minimum value between the two classification thresholds and mark it as b. Defect locations with a category index exceeding a are marked as 1. Defect locations with a category index between a and b are marked as 2. Defect locations with a category index below b are marked as 3. Step A5: Output the judgment result. When the defect position is marked as 1, the defect position is marked as a related display and output; when the defect position is marked as 2, the defect position is marked as a non-related display and output; when the defect position is marked as 3, the defect position is marked as a pseudo display and output; the category index, display category and classification threshold of each defect position are output.

[0025] It should be noted that the Sigmoid function is used to convert input features into data that is easy to analyze and process. The hyperparameters are used to adjust the values ​​of the grayscale value dataset and to modify the grayscale values ​​so that they are correctly correlated with the output results. The control constants can be set according to actual conditions such as defect identification requirements, and will not be elaborated here.

[0026] The feature extraction module accesses the image library to obtain the magnetic trace results of each defect position in the image, and uses the magnetic trace results of each defect position as a reference data set. It should be explained that the magnetic trace results determined in the historical image are the display categories corresponding to the magnetic trace categories identified by professionals.

[0027] The specific steps for calculating the misjudgment rate using the support vector machine model based on the display categories of each defect location and the reference data set are as follows: Step B1: Data preprocessing: sort and label the display categories of each defect location according to the data in the reference data set, and merge the sorted display categories into a discriminant data set; Step B2: Set the evaluation rules and compare the data in the discrimination dataset with the corresponding data in the control dataset according to the labels. If the displayed categories are the same, the defect location is marked as correct; if the displayed categories are different, the defect location is marked as incorrect. Step B3: Calculate the false positive rate. Randomly divide the discriminant data set and the control data set into n subsets of equal size and correspond them one to one according to the serial number. Select one subset each time as the test set, and the remaining n-1 subsets as the training set. Count the number of incorrect labels in the test set, and take the ratio of the number of incorrect labels to the total amount of data in the test set as the error rate of the test set. After calculating the error rate of the test set, randomly select a subset from the training set and calculate the error rate before removing it. Take the average of all the calculated error rates and output it as the false positive rate. It should be noted that when calculating the false positive rate, the number of divisions n is not unique and can be set by yourself, for example, setting n to 5, etc., which will not be analyzed here.

[0028] After receiving the pixel intensity of each pixel in the image, the blur detection module sets an intensity sequence for the pixel points in the defect area and uses Laplace transform to calculate the complex frequency domain signal value. It then performs normalization and numerical compression to obtain the complex frequency domain signal compression value. The geometric mean method is used to calculate the adjustment correction value based on the complex frequency domain signal compression value and the false positive rate: , where T is the adjustment correction value, f is the complex frequency domain signal compression value, and q is the false positive rate; The adjustment correction value is compared with the preset adjustment threshold. If the adjustment correction value exceeds the adjustment threshold, no adjustment is performed. If the adjustment correction value is lower than the adjustment threshold, the camera exposure time and rotation speed are adjusted to the corresponding precise shooting gear parameters. The inverse of the adjustment correction value is taken to obtain the adjustment ratio, which is marked as t. The shooting area of ​​the shooting camera is calculated based on the adjustment ratio: ,in, is the adjusted sampling area, is the defect area at the defect location.

[0029] It should be noted that Laplace transform is a method applied in signal processing, system analysis and image processing. It is used to quantify image data and analyze image blur. The larger the complex frequency domain signal value, the larger the adjustment correction value and the clearer the image. The error rate reflects the accuracy of model resolution. The higher the error rate, the smaller the adjustment correction value, the blurrier the image and the greater the need to expand the sampling area. The precise sampling gear parameters corresponding to the camera exposure time and rotation speed are set by professionals according to actual conditions.

[0030] The critical processing module sets a similarity range and filters the category index of each defect area based on the similarity range to screen the defect location. The minimum value is selected from the two classification thresholds as the analysis threshold. The analysis threshold is processed using a preset comparison difference to obtain a similarity range. The analysis threshold is then added or subtracted by a comparison difference to obtain a similarity range. Defect areas with category indices within the similarity range are marked and the marked locations are sent to the data acquisition module. When the data acquisition module receives the marked locations, it calls historical data to transmit the probability of defects at each marked location to the critical processing module. It should be noted that, if the defects mentioned above are magnetic marks classified as non-pseudo-display, then the probability of defects occurring at each marking position is the probability of the marking position being a magnetic mark classified as non-pseudo-display, for example, the probability of the appearance of a magnetic mark classified as non-pseudo-display at the bottle mouth.

[0031] After receiving the probability of a defect at the marked position, the critical processing module counts the grayscale values ​​of each marked area and calculates the average value to obtain the average grayscale value of each area. The display category of the marked position is reclassified using the logistic regression algorithm based on the probability of a defect at each marked position and the average grayscale value. The logistic regression coefficient is calculated using the logistic regression algorithm: , where L is the logistic regression coefficient of the mark position, e is the natural base, z is the logistic regression parameter, and z is the sum of the probability of a defect occurring at the mark position and the average grayscale value; When the logistic regression coefficient exceeds the preset false display discrimination threshold, the suspicious mark is classified as non-relevant display; when the logistic regression coefficient is lower than the preset false display discrimination threshold, the suspicious mark is classified as false display.

[0032] It should be noted that the historical data is the probability of each mark in the historical database being suspected of having defects, and is used to reclassify the display category of the suspicious mark. The larger the average grayscale value, the lower the brightness of the image, the more likely magnetic marks will appear at the mark position, the greater the probability of defects at the mark position or the larger the average grayscale value, the larger the logistic regression coefficient, and the greater the probability that the mark position is a non-false display.

[0033] The above embodiments may be implemented in whole or in part through software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments may be implemented in whole or in part in the form of a computer program product.

[0034] Those skilled in the art will appreciate that the modules and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application of the technical solution and the invention constraints. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0035] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0036] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0037] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. Intelligent identification system of surface defects of non-welded bottle containers based on magnetic particle detection, characterized in that: It includes a data acquisition module, a feature extraction module, a fuzzy detection module and a critical processing module, and the signals between the modules are connected; The data acquisition module is used to collect images of non-welded bottle containers and pass them to the feature extraction module to extract image features, obtain the pixel intensity value of each pixel in the image and pass it to the fuzzy detection module. When receiving the mark position passed by the critical processing module, the historical data is called to pass the probability of defects at each mark position to the critical processing module; After receiving the non-welded bottle container image, the feature extraction module detects the gray value in the image, and selects multiple defect locations and obtains the defect area according to the preset gray threshold. The defect area and gray value in the image are combined to build a support vector machine model to calculate the category index and the misjudgment rate and set the classification threshold. The defect location is classified and displayed using the classification threshold, and the calculated misjudgment rate is passed to the fuzzy detection module. Each defect location and the category index of the corresponding defect location are passed to the critical processing module. After receiving the pixel intensity value of each pixel in the image, the blur detection module sets the intensity sequence and calculates the complex frequency domain signal value using Laplace transform, and adjusts the sampling of the sampling camera based on the complex frequency domain signal intensity value and the error rate transmitted by the feature extraction module; The critical processing module sets a similarity range, selects multiple classified fuzzy positions from all defect positions according to the similarity range, and marks them. The marked positions are sent to the data acquisition module. After receiving the probability of defects occurring at each marked position, the average grayscale value of each marked position is calculated and the display category of the marked position is reclassified.

2. According to claim 1, the intelligent identification system for surface defects of non-welded bottle containers based on magnetic particle detection is characterized in that: The categories of classified display are relevant display, non-relevant display and pseudo display. The relevant display is the magnetic trace display formed by the leakage magnetic field generated by the defect during magnetic particle inspection and the magnetic powder adsorbs the magnetic powder; the non-relevant display is the magnetic trace display formed by the leakage magnetic field generated by the cross-section change or the material permeability change during magnetic particle inspection and the magnetic powder adsorbs the magnetic powder; the pseudo display is the magnetic trace display formed by the non-leakage magnetic field adsorbing the magnetic powder.

3. The intelligent identification system for surface defects of non-welded bottle containers based on magnetic particle detection according to claim 1 is characterized in that: The feature extraction module collects the grayscale value of each pixel in the image of the non-welded bottle container and compares the grayscale value of each pixel with the preset grayscale threshold. When the grayscale value of adjacent pixels is lower than the preset grayscale threshold, the adjacent pixels are merged to obtain the defect position. After obtaining multiple defect positions, the defect area is calculated according to the pixel size and the number of pixels corresponding to the defect position. The average value of the grayscale value of each defect position is calculated as the defect grayscale value of the corresponding defect position, and the defect grayscale value and defect area of ​​each defect position are merged into a grayscale value data set and an area data set respectively.

4. The intelligent identification system for surface defects of non-welded bottle containers based on magnetic particle detection according to claim 3 is characterized in that: The feature extraction module integrates the gray value data set and the area data set to build a support vector machine model to calculate the category index of each defect location. The specific steps are as follows: Step A1: input the gray value data set and the area data set as analysis features into the support vector machine; Step A2: Select the Sigmoid function as the kernel function to transform the input features. The Sigmoid function is: ,in, is the kernel function result after feature conversion, is the hyperbolic tangent function, is a hyperparameter, c is a control constant, x and y are two data corresponding to the defect position in the gray value data set and the area data set respectively; the kernel function result obtained by transforming the corresponding data features in the two data sets where each defect position is located is used as the category index of the corresponding defect position; Step A3: Use the percentile method to set two adjustment coefficients to converge and classify the kernel function results. The formula is: , ,in, and are two adjustment coefficients, and are the convergence results of the two kernel functions, is the summed average of all category indices; the two kernel function results after convergence are used as classification thresholds; Step A4: Determine the defect category, compare the category index of each defect position with the two classification thresholds, select the maximum value of the two classification thresholds and mark it as a, select the minimum value of the two classification thresholds and mark it as b, mark the defect position with a category index exceeding a as 1, mark the defect position with a category index between a and b as 2, and mark the defect position with a category index lower than b as 3; Step A5: Output the discrimination result. When the defect position is marked as 1, the defect position is marked as a relevant display and output; when the defect position is marked as 2, the defect position is marked as a non-relevant display and output; when the defect position is marked as 3, the defect position is marked as a pseudo display and output; output the category index, display category and classification threshold of each defect position; The Sigmoid function is used to convert input features into kernel function processing data types.

5. The intelligent identification system for surface defects of non-welded bottle containers based on magnetic particle detection according to claim 4 is characterized in that: The feature extraction module accesses the image library to obtain the magnetic trace results of each defect position in the image, and uses the magnetic trace results of each defect position as a reference data set. The specific steps of calculating the misjudgment rate using the support vector machine model based on the display categories of each defect location and the reference data set are as follows: Step B1: data preprocessing, sorting and numbering the display categories of each defect position according to the data in the reference data set, and merging the sorted display categories into a discriminant data set; Step B2: Set the evaluation rules, compare the data in the discrimination data set with the corresponding data in the control data set according to the labels, and if the displayed categories are the same, mark the defect location as correct; If the display categories are not the same, the defect location is marked as an error; Step B3: Calculate the false positive rate, randomly divide the discrimination data set and the control data set into n subsets of equal size and correspond one to one according to the serial number, select one subset each time as the test set, and the remaining n-1 subsets as the training set, count the number of wrong labels in the test set, and take the ratio of the number of wrong labels to the total amount of test set data as the error rate of the test set; after calculating the error rate of the test set, randomly select a subset from the training set and calculate the error rate and then remove it; take the average of all the calculated error rates as the false positive rate and output it; The magnetic trace results determined in the historical image are the display categories corresponding to the identified magnetic trace categories.

6. The intelligent identification system for surface defects of non-welded bottle containers based on magnetic particle detection according to claim 5 is characterized in that: After receiving the pixel intensity of each pixel in the image, the blur detection module sets an intensity sequence for the pixel in the defect area and calculates the complex frequency domain signal value by Laplace transform and performs normalization and numerical compression to obtain the complex frequency domain signal compression value; The geometric mean method is used to calculate the adjustment correction value based on the comprehensive complex frequency domain signal compression value and the error rate: , where T is the adjustment correction value, f is the complex frequency domain signal compression value, and q is the misjudgment rate; The intensity sequence is a combination of pixel intensities of pixels in the defect area. Laplace transform is used to calculate the complex frequency domain signal value according to the pixel intensity combination. The complex frequency domain signal is a parameter that reflects the clarity of the image.

7. The intelligent identification system for surface defects of non-welded bottle containers based on magnetic particle detection according to claim 6 is characterized by: The adjustment correction value is compared with the preset adjustment threshold. When the adjustment correction value exceeds the adjustment threshold, no adjustment is performed. When the adjustment correction value is lower than the adjustment threshold, the camera exposure time and rotation speed are adjusted to the corresponding precise shooting gear parameters set in advance. The adjustment correction value is taken as the inverse to obtain the adjustment ratio, which is marked as t. The shooting area of ​​the shooting camera is calculated according to the adjustment ratio: ,in, is the adjusted sampling area, is the defect area at the defect location.

8. The intelligent identification system for surface defects of non-welded bottle containers based on magnetic particle detection according to claim 4 is characterized in that: The critical processing module sets a similarity range, and screens the category index of each defect area according to the similarity range to screen the defect location. The minimum value is selected as the analysis threshold from the two classification thresholds, and the analysis threshold is processed by a preset comparison difference to obtain a similarity range. The analysis threshold is added or subtracted by a comparison difference to obtain a similarity range, and the defect area with a category index within the similarity range is marked, and the marked position is sent to the data acquisition module; When the data acquisition module receives the marked position, it calls the historical data to pass the probability of defects occurring at each marked position to the critical processing module.

9. The intelligent identification system for surface defects of non-welded bottle containers based on magnetic particle detection according to claim 8 is characterized in that: After receiving the probability of defects at the marked position, the critical processing module counts the grayscale values ​​of each marked area and calculates the average grayscale value of each area. The display category of the marked position is reclassified using the logistic regression algorithm based on the probability of defects at each marked position and the average grayscale value, and the logistic regression coefficient is calculated using the logistic regression algorithm: , where L is the logistic regression coefficient of the marking position, e is the natural base, z is the logistic regression parameter and z is the sum of the probability of a defect occurring at the marking position and the average gray value; When the logistic regression coefficient exceeds the preset false display discrimination threshold, the suspicious mark is classified as a non-relevant display; when the logistic regression coefficient is lower than the preset false display discrimination threshold, the suspicious mark is classified as a false display.

Images