Antenna array substrate defect detection method based on image recognition

Abstract

The application discloses an antenna array substrate defect detection method based on image recognition and relates to the field of image recognition.The method comprises the following steps: collecting an image of an antenna array substrate, dividing the image into an antenna image array according to an antenna array division standard, performing feature recognition on a plurality of antenna images, obtaining a plurality of antenna features, performing abnormality degree analysis on the plurality of antenna features and a plurality of standard antenna features, obtaining an antenna abnormality degree array, respectively configuring a pollution recognition coefficient according to the antenna abnormality degree array, performing pollution recognition, selecting a moving window for the antenna abnormality degree array and a pollution parameter array, performing adjacent antenna deformation verification and adjacent antenna pollution verification, processing to obtain a deformation verification coefficient and a pollution verification coefficient, obtaining a fusion defect parameter, and taking the fusion defect parameter as a defect detection result.The antenna array substrate defect detection method based on image recognition solves the problems of high false alarm rate of defect detection, low efficiency of defect detection result and insufficient accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image recognition, in particular to an antenna array substrate defect detection method based on image recognition. BACKGROUND

[0002] Image defect detection is an important research direction in the field of machine vision, mainly used for detecting scratches, color differences, defects and other defects on the surface of an article.

[0003] In the prior art, the antenna array substrate is detected by a single detection method, which relies on a manually designed feature extraction method, and it is difficult to accurately extract image features, and the scene adaptability is low to complex environmental conditions and dynamic scenes, and there is a high false positive rate of image defect detection, which leads to poor detection effect. SUMMARY

[0004] The present application provides an antenna array substrate defect detection method based on image recognition, which is used to solve the problem of low accuracy and poor detection effect of image defect detection in the prior art.

[0005] In view of the above problems, the present application provides an antenna array substrate defect detection method based on image recognition, which comprises:

[0006] According to a preset image acquisition standard, the image of the antenna array substrate is acquired, and an antenna image array is obtained according to an antenna array division standard;

[0007] Feature recognition is performed on a plurality of antenna images in the antenna image array to obtain a plurality of antenna features, and abnormality analysis is performed on the plurality of standard antenna features to obtain an antenna abnormality array;

[0008] According to the antenna abnormality array, a pollution recognition coefficient is configured respectively, pollution recognition is performed, and a pollution parameter array is obtained;

[0009] The antenna abnormality array and the pollution parameter array are selected by a moving window, and the selected moving window is verified by adjacent antenna deformation and adjacent antenna pollution, and a deformation verification coefficient and a pollution verification coefficient are obtained by processing, and the antenna array substrate fusion defect parameter is calculated and obtained by combining the antenna abnormality array and the pollution parameter array, as a defect detection result.

[0010] One or more technical solutions provided in the present application have at least the following technical effects or advantages:

[0011] The application provides an image recognition-based video abnormal behavior detection method, which finally obtains a comprehensive defect detection result by fusing multiple parameters through standardized image acquisition, feature extraction, abnormality degree calculation, pollution parameter identification and mobile window verification. Compared with the traditional image defect detection method, the image recognition-based video abnormal behavior detection method solves the problems of low scene adaptability caused by complex environmental conditions and dynamic scenes, high false positive rate of image defect detection and poor detection effect, and improves the detection precision and efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0012] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0013] Figure 1 is a flowchart of an image recognition-based antenna array substrate defect detection method;

[0014] Figure 2 is a flowchart of fusion defect parameter calculation of an image recognition-based antenna array substrate defect detection method. DETAILED DESCRIPTION

[0015] The present application provides an image recognition-based antenna array substrate defect detection method to solve the problems in the prior art.

[0016] The technical solutions in the embodiments of the present application will be described clearly and completely with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0017] It should be noted that the terms "include" and "have" are intended to cover non-exclusive inclusion, for example, a process, method, system, product or server including a series of steps or units need not be limited to those clearly listed steps or units, but can include other steps or modules not clearly listed or inherent to the process, method, product or device.

[0018] The present application will be described in detail below with reference to the drawings.

[0019] Embodiment one, as shown in the figure, the present application provides an image recognition-based antenna array substrate defect detection method, the method comprises: Figure 1 ​

[0020] S10: collect an image of the antenna array substrate according to a preset image collection standard, and divide the image into an antenna image array according to an antenna array division standard;

[0021] The antenna array division standard is used to divide the antenna array substrate image into the antenna image array, so as to ensure the consistency and repeatability of the image quality. The antenna image array is obtained by dividing the entire substrate image into a plurality of independent antenna element images, which facilitates subsequent processing.

[0022] The step S10 in the method provided in the embodiments of the present application includes:

[0023] The image of the antenna array substrate is collected according to the preset image collection standard.

[0024] An antenna array division standard is obtained, wherein the antenna array division standard includes a plurality of standard regions of a plurality of antenna images.

[0025] The images of a plurality of antennas in the antenna array substrate image are divided according to the antenna array division standard, and an antenna image array is obtained.

[0026] Specifically, the preset image collection standard includes the orientation and distance of the camera, the intensity and number of the light source, etc. The image of the antenna array substrate is collected according to the preset image collection standard. The image is collected under the same standard, which can avoid the problem of non-standard image collection and the interference of other factors on the image collection, affect the definition of the collected image, and cause inaccurate data.

[0027] For example, the camera is placed vertically downward and parallel to the antenna array substrate, has a resolution of 1080P pixels, and has a preset distance. The image is collected under a standard light intensity of 100 lux.

[0028] Further, the antenna array division standard is obtained, wherein the antenna array division standard includes a plurality of standard regions of a plurality of antenna images. The antenna array division standard includes the accurate position of each antenna in the image of the antenna array substrate, which is usually from a design drawing or a standard sample.

[0029] For example, the standard region of the plurality of antenna images can be the coordinate range of the coordinate pixels in the boundary box of each antenna.

[0030] Further, due to the optical distortion of the lens, the size and shape of the antenna images at different positions in the antenna array substrate image are different. Therefore, the antenna array substrate image is cropped and divided into a plurality of standard regions according to the antenna array division standard, and an antenna image array is obtained.

[0031] For example, the antenna array substrate image is cropped according to the coordinates using a programming tool, and an antenna image array of 100 standard regions is obtained.

[0032] In the embodiment of the present application, the antenna array division standard is obtained in step S10, wherein the antenna array division standard comprises a plurality of standard regions of a plurality of antenna images, comprising:

[0033] An antenna array substrate image without defects is obtained according to the preset image acquisition standard as a standard antenna array substrate image.

[0034] The region pixel coordinate set of the image of a plurality of antennas in the standard antenna array substrate image is divided as a plurality of standard regions to obtain the antenna array division standard.

[0035] Specifically, an antenna array substrate image without defects is obtained according to the preset image acquisition standard as a standard antenna array substrate image.

[0036] Further, the division standard is derived from the standard antenna array substrate image without defects, and the region pixel coordinate set of the image of a plurality of antennas similar to [(x1, y1), (x2, y2)] in the standard antenna array substrate image is obtained by manual or annotation software boundary box annotation as a plurality of standard regions to obtain the antenna array division standard. In actual application, if the standard antenna array substrate image changes, the division standard needs to be updated and re-divided.

[0037] For example, from the manually annotated standard antenna array substrate image, the region pixel coordinate set [(200, 300), (200, 301)] of the image of an antenna in the standard antenna array substrate image is obtained.

[0038] In the embodiment of the present application, the image is collected under the same standard, which can avoid interference of other factors on image collection, affect the definition of the collected image, and cause inaccurate data. At the same time, based on the standard antenna array substrate image without defects, the division standard is established to ensure the consistency of the division standard.

[0039] S20: performing feature recognition on a plurality of antenna images in the antenna image array to obtain a plurality of antenna features, performing abnormality analysis on the plurality of antenna features and a plurality of standard antenna features to obtain an antenna abnormality array;

[0040] In the embodiment of the present application, the plurality of antenna images in the antenna image array are subjected to feature recognition to obtain antenna features such as shape, texture, and area, the antenna structure features in each antenna image are compared with the standard structure features of each antenna region, and abnormal length analysis is performed to obtain an antenna abnormality sequence.

[0041] The step S20 in the method provided by the embodiment of the present application comprises:

[0042] An antenna feature recognizer is obtained, wherein the antenna feature recognizer is constructed based on a convolutional neural network and is trained using a plurality of sample antenna image sets of a plurality of antennas and a plurality of labeled sample antenna feature sets, each sample antenna feature including structural features;

[0043] A plurality of antenna images in the antenna image array are input into the antenna feature recognizer, and a plurality of antenna features are obtained by output.

[0044] A plurality of standard antenna features of a plurality of antenna images in a defect-free antenna array substrate image are obtained.

[0045] The deviation amplitudes of the plurality of antenna features and the plurality of standard antenna features are calculated respectively as antenna abnormality degrees, and an antenna abnormality degree array is obtained.

[0046] Specifically, first, an antenna feature recognizer is obtained, wherein the antenna feature recognizer is constructed based on a convolutional neural network and is trained using a plurality of sample antenna image sets of a plurality of antennas and a plurality of labeled sample antenna feature sets, each sample antenna feature including structural features. The convolutional neural network (CNN) is a deep learning model that is good at processing image-related tasks including image classification, object detection, and image segmentation.

[0047] For example, the antenna feature recognizer can be trained as follows: data preparation: divide the plurality of sample antenna image sets and the plurality of labeled sample antenna feature sets into a training set, a validation set, and a test set in a ratio of 7:2:1. Model construction: mainly composed of an input layer, a convolutional layer, a pooling layer, and an output layer, wherein the input layer inputs the sample antenna image; the convolutional layer performs preliminary feature extraction on the sample antenna image set; the pooling layer reduces the spatial size of the data through downsampling operation and retains important features in the bounding box; the output layer one-to-one maps the image features of the sample antenna image to the labeled plurality of sample antenna feature sets, summarizes the extracted features, and outputs the recognition result. Model training: using the sample antenna image set as input and the difference between the labeled plurality of sample antenna feature sets and the supervised label as target, using the Adam optimizer, the initial learning rate is 1e -4 , the MSE loss function is selected to calculate the error between the labeled plurality of sample antenna feature sets and the plurality of antenna images, the prediction result is obtained by forward propagation, the parameters are updated by back propagation, the performance is evaluated every 1 round of training with the validation set to avoid overfitting, and when the MSE loss of the training set decreases by less than 1e -5, the antenna feature recognizer is determined to be converged, the training is stopped and the final network parameters are saved to obtain the antenna feature recognizer. Test cycle: select a device (GPU or CPU), move the antenna feature recognizer to the device, perform 10 training cycles, and perform a test after each cycle.

[0048] Further, after the antenna feature recognizer is constructed, multiple antenna images in the antenna image array are input into the antenna feature recognizer, and multiple antenna features of multiple antennas are obtained.

[0049] For example, 100 antenna images are input into the antenna feature recognizer to obtain 100 corresponding antenna features such as shape, texture, area, etc.

[0050] Further, features are extracted from the defect-free antenna array substrate image as the standard for antenna array substrate defect detection, i.e., multiple standard antenna features of multiple antenna images in the defect-free antenna array substrate image are obtained, such as a standard antenna shape or the area of an antenna region formed in the image.

[0051] For example, features are extracted from the defect-free antenna array substrate image to obtain multiple standard antenna features such as standard shape features, standard texture features, and standard area features.

[0052] Further, the deviation amplitudes of the multiple antenna features and the multiple standard antenna features are calculated respectively as antenna abnormality degrees to obtain an antenna abnormality array. The difference between the multiple antenna features and the multiple standard antenna features is obtained as a deviation, and the deviation amplitude is the ratio of the difference between the multiple standard antenna features to the multiple standard antenna features, and the calculation formula is: [(standard antenna feature difference) / standard antenna feature]. The larger the value, the larger the deviation amplitude. The deviation amplitude is used as the antenna abnormality degree, and the antenna abnormality array is obtained by integration. At the same time, the antenna abnormality degree can reflect whether the antenna image may have an abnormality. The larger the antenna abnormality degree, the greater the probability of an abnormality in the antenna image, the greater the possibility of an abnormality in the antenna array substrate corresponding to the antenna image, and the more likely there is a defect.

[0053] For example, the antenna features of the input antenna images have antenna areas of 0.12 cm 2 , 0.15 cm 2 , 0.15 cm 2 , 0.16 cm 2 , and 0.18 cm 2 , and the standard antenna features have areas of 0.14 cm 2 , and the deviation amplitudes of the areas, i.e., the abnormality degrees of the areas, are 0.214, 0.067, 0.067, 0.14, and 0.285, respectively, and the antenna abnormality array is obtained by integration.

[0054] In the embodiment of the present application, the antenna feature recognizer is constructed and obtained through the convolutional neural network. The plurality of antenna images in the antenna image array are input into the antenna feature recognizer to obtain a plurality of antenna features. The plurality of antenna images in the antenna image array are subjected to feature recognition and abnormality analysis with a plurality of standard antenna features of a plurality of antenna images in a defect-free antenna array substrate image to obtain an antenna abnormality array. The CNN can automatically learn features, and the antenna feature recognizer constructed and obtained through the convolutional neural network has accurate feature recognition capability. The abnormality array quantifies the deviation degree of each antenna, and provides input features for subsequent steps of feature recognition, abnormality analysis and feature extraction. By comparing the actual features with the standard features, the abnormality degree is quantified, and the basic situation of the antenna array substrate defect is preliminarily reflected to obtain the antenna abnormality array of antenna feature abnormality recognition.

[0055] S30: According to the antenna abnormality array, a pollution recognition coefficient is configured respectively to perform pollution recognition and obtain a pollution parameter array;

[0056] In the embodiment of the present application, the antenna feature abnormality recognition obtains the antenna abnormality array, but the recognized antenna feature abnormality may be deformation or may be a recognition error caused by rust pollution. The pollution recognition coefficient of integrated recognition is configured respectively to recognize the rust area and obtain the pollution parameter array. The pollution parameter sequence can be further analyzed, which is beneficial to improve the accuracy of antenna array substrate defect detection.

[0057] In the embodiment of the present application, step S30 comprises:

[0058] A plurality of antenna pollution recognition network groups corresponding to a plurality of antennas are obtained;

[0059] According to each antenna abnormality in the antenna abnormality array, a plurality of pollution recognition coefficients are obtained, and the total number of antenna pollution recognition networks in each antenna pollution recognition network group is combined to calculate a plurality of pollution recognition network numbers;

[0060] A plurality of antenna pollution recognition networks of the plurality of pollution recognition network numbers are randomly selected respectively, a plurality of antenna images in the antenna image array are input respectively, a plurality of pollution parameter sets are output and obtained, and mean value processing is performed to obtain a pollution parameter array.

[0061] Specifically, a plurality of antenna pollution recognition network groups corresponding to a plurality of antennas are obtained. The pollution recognition network group can be used for detecting the pollution degree and performing pollution recognition.

[0062] Further, according to each antenna anomaly in the antenna anomaly array, as a plurality of pollution identification coefficients. According to the pollution identification coefficient, the number of antenna pollution network groups is calculated, and the number of antenna pollution identification networks is obtained. The total number of antenna identification networks in each antenna pollution identification network group is the total number of antenna identification networks.

[0063] The product of the total number of antenna identification networks in the antenna pollution identification network group and the plurality of pollution identification coefficients is calculated respectively and rounded, as a plurality of pollution identification network numbers. The higher the antenna anomaly, the more network numbers are used to improve the accuracy of identification detection and avoid identification errors caused by the lack of network numbers, which affects the accuracy of antenna defect detection. The lower the antenna anomaly, the fewer network numbers are used to improve the detection efficiency.

[0064] For example, one antenna pollution identification network group is 20, the pollution identification coefficient is 0.7, and the total number of antenna identification networks is 20x0.7=14.

[0065] Further, a random selection method is used to randomly select a plurality of antenna pollution identification networks of a plurality of pollution identification network numbers for antenna pollution identification. A plurality of antenna images in the antenna array are input into the selected antenna pollution identification network, and a plurality of pollution parameter sets are obtained by identification output. The pollution parameter is the pollution probability, and the larger the pollution parameter, the greater the possibility of pollution. The pollution parameter set is processed by mean value to obtain the pollution parameter of a certain antenna, and the pollution parameter array of a plurality of antennas is obtained by integration. Through mean value processing, the random error of a single network is reduced, the stability of defect detection is improved, and the accuracy of defect detection is ensured.

[0066] For example, a group of antenna images are input, and 5 pollution identification networks are randomly selected for identification, and the obtained pollution parameters are 0.05, 0.2, 0.4, 0.3, and 0.15. The mean value processing is (0.05+0.2, 0.4+0.3+0.15) / 5=0.22, and the pollution parameter of a certain antenna is 0.22.

[0067] In the embodiment of the application, the step S30 of obtaining a plurality of antenna pollution identification network groups corresponding to a plurality of antennas comprises:

[0068] According to the historical data of antenna pollution identification, a plurality of sample antenna image sets of a plurality of antennas are collected, and the pollution parameters in each sample antenna image are identified to obtain a plurality of sample pollution parameter sets.

[0069] Based on the convolutional neural network, a plurality of antenna pollution identification network groups are constructed respectively.

[0070] The multiple sample antenna image sets and the multiple sample pollution parameter sets are randomly divided respectively, and the multiple antenna pollution identification network groups are integrated for training to obtain the multiple antenna pollution identification network groups that are trained and tested to converge.

[0071] Specifically, historical data of antenna pollution identification is obtained within a certain historical time, and the obtained historical data is ensured to be valid data. A plurality of sample antenna image sets of a plurality of antennas are collected. The pollution parameters in each sample antenna image are labeled using a labeling tool and according to the pollution degree to obtain a plurality of sample pollution parameter sets. Among them, the pollution parameters with high pollution degree, the pollution parameters with medium pollution degree, and the pollution parameters with low pollution degree are marked as high pollution, medium pollution, and low pollution respectively.

[0072] For example, the labeling tool is used to mark the pollution degree of the pollution parameters to obtain the sample pollution parameter set: [high pollution, high pollution, medium pollution, high pollution], or the pollution percentage, such as the pollution area percentage, such as 90% or 50%.

[0073] Further, based on a convolutional neural network, a plurality of antenna pollution identification network groups are constructed respectively. Among them, the network group is composed of a plurality of CNN models, and each model can be independently trained.

[0074] For example, 10 CNN models are created, each model has 3 convolutional layers and 2 fully connected layers, and the antenna pollution identification network group can be obtained by training the following technical path: data preparation: the plurality of antenna images and the labeled plurality of sample pollution parameter sets are used as input, the data is randomly divided into 10 parts, and the 10 CNN models are trained simultaneously. Model construction: mainly composed of an input layer, a convolutional layer, a pooling layer, and an output layer, wherein the input layer divides the plurality of antenna images into 3 categories according to the pollution degree, or the corresponding pollution percentage; the convolutional layer performs preliminary feature extraction on the plurality of sample pollution parameter sets; the pooling layer reduces the spatial size of the data through the downsampling operation to retain the marked important features; the output layer one-to-one maps the image features of the plurality of antenna images to the labeled plurality of sample pollution parameter sets, and outputs the recognition result by summarizing the extracted features.

[0075] Further, the multiple sample antenna image sets and the multiple sample pollution parameter sets are randomly divided respectively, and integrated for training to obtain the multiple antenna pollution identification network groups that are trained and tested to converge.

[0076] For example, the multiple antenna pollution identification network groups are integrated for training to ensure that the data is randomly divided for training and to ensure the diversity of the models, such as 80% of the training set and 20% of the test set.

[0077] In this embodiment, multiple antenna contamination identification network groups are constructed using convolutional neural networks. These networks output multiple contamination parameter sets, which are then used to calculate a contamination parameter array. Specifically, multiple antenna contamination identification networks are randomly selected for antenna contamination identification. Multiple antenna images within the antenna image array are input separately, and the output yields multiple contamination parameter sets. These contamination parameters are, for example, contamination level or contamination percentage; higher contamination parameters indicate a greater likelihood and degree of contamination. The contamination parameter sets are averaged to obtain the contamination parameters for a specific antenna, which are then integrated to obtain the contamination parameter array for multiple antennas. Averaging reduces the random error of individual models, improving the stability and accuracy of defect detection. Ensemble training enhances model stability, avoids bias from a single model, and comprehensively covers various contamination scenarios with historical data, ensuring model diversity.

[0078] S40: Select a moving window for the antenna anomaly array and the contamination parameter array, and perform adjacent antenna deformation verification and adjacent antenna contamination verification on the selected moving window. Process to obtain deformation verification coefficient and contamination verification coefficient. Combine the antenna anomaly array and the contamination parameter array to calculate the fusion defect parameters of the antenna array substrate as the defect detection result.

[0079] In this embodiment, if external force causes deformation or corrosion, the magnitude of anomalies and contamination in nearby antennas should be approximately the same. The moving window selects 2×2 antennas each time, analyzes the similarity of anomaly and contamination parameters, performs proximity verification, and finally calculates the mean to obtain the deformation verification coefficient and contamination verification coefficient. Combined with the antenna anomaly array and contamination parameter array, the fusion defect parameters of the antenna array substrate are calculated.

[0080] like Figure 2 As shown in the embodiment of this application, S40 includes:

[0081] A moving window is selected within the antenna anomaly array to obtain the first antenna anomaly window;

[0082] The deformation verification coefficients of the first window are obtained by performing adjacent antenna deformation verification on several antenna anomalies within the first antenna anomaly window.

[0083] Continue to select moving windows for the antenna anomaly array and the contamination parameter array, and perform deformation verification and contamination verification of neighboring antennas to obtain the set of window deformation verification coefficients and the set of window contamination verification coefficients. Calculate the mean values ​​of the two sets of window deformation verification coefficients and contamination verification coefficients respectively.

[0084] The maximum antenna abnormality and the maximum pollution parameter in the antenna abnormality array and the pollution parameter array are obtained, and the abnormality parameter and the pollution parameter are calculated by combining the historical average antenna abnormality and the historical average pollution parameter of the antenna.

[0085] The fusion defect parameter is calculated by weighting the abnormality parameter and the pollution parameter according to the deformation verification coefficient and the pollution verification coefficient, and is taken as the defect detection result.

[0086] Specifically, a moving window is selected in the antenna abnormality array to obtain a first antenna abnormality window. The moving window can move on the antenna abnormality array, so that each local area is processed.

[0087] For example, a 2*2 moving window is selected as the first antenna abnormality window, which contains the antenna abnormality of four antennas.

[0088] Further, the antenna abnormality in the first antenna abnormality window is subjected to adjacent antenna deformation verification, and the similarity of the antenna abnormality in the first antenna abnormality window can be taken as the deformation verification coefficient to obtain the first window deformation verification coefficient. Similarly, the pollution parameter array is subjected to moving window selection and adjacent antenna pollution verification.

[0089] For example, a 2*2 moving window containing four antenna abnormality is selected to perform adjacent antenna deformation verification to obtain the first window deformation verification coefficient.

[0090] Further, the antenna abnormality array and the pollution parameter array are subjected to moving window selection and adjacent antenna deformation verification and adjacent antenna pollution verification. The variance of the antenna abnormality in the first antenna abnormality window can be used to calculate the deformation verification coefficient, specifically the inverse of the variance. If the similar antennas are deformed or polluted, the antenna abnormality and the pollution parameter are probably similar, and the variance is small, and the deformation verification coefficient is large.

[0091] Similarly, the pollution verification coefficient can be calculated, and the window deformation verification coefficient set and the window pollution verification coefficient set are combined to calculate the average to obtain the deformation verification coefficient and the pollution verification coefficient.

[0092] Further, the maximum antenna abnormality and the maximum pollution parameter in the antenna abnormality array and the pollution parameter array are obtained, and the abnormality parameter and the pollution parameter are calculated by combining the historical average antenna abnormality and the historical average pollution parameter of the antenna in the historical time.

[0093] Meanwhile, the influence degree of the deformation verification coefficient and the pollution verification coefficient on the defect detection result is allocated a weight, the abnormal parameter and the pollution parameter are calculated by weighting, and a fusion defect parameter is obtained as the defect detection result. The deformation verification coefficient and the pollution verification coefficient are allocated a calculation weight, the ratio of the deformation verification coefficient and the pollution verification coefficient to the sum of the two is calculated as a deformation weight and a pollution weight, and the fusion defect parameter = deformation weight × abnormal parameter + pollution weight × pollution parameter. The greater the value is, the more likely there is a defect.

[0094] The skilled person can refer to the fusion defect parameter to select and use the antenna array substrate, for example, the antenna array substrate with a larger fusion defect parameter is abandoned.

[0095] In the embodiment of the present application, in step S40, the deformation of the adjacent antennas of the several antenna abnormality degrees in the first antenna abnormality degree window is verified, and a first window deformation verification coefficient is obtained, including:

[0096] The similarity of the several antenna abnormality degrees in the first antenna abnormality degree window is calculated as the first window deformation verification coefficient.

[0097] Specifically, the similarity of the several antenna abnormality degrees in the first antenna abnormality degree window is calculated as the first window deformation verification coefficient. For example, the similarity is the inverse of the variance of the several antenna abnormality degrees in the first antenna abnormality degree window. The smaller the variance is, the higher the similarity is, and the larger the first window deformation verification coefficient is.

[0098] In the embodiment of the present application, the mobile window verification uses the spatial relationship to analyze the local part, and the similarity of the antenna abnormality degree and the pollution parameter in the mobile window is calculated. The difference between the historical average antenna abnormality degree and the historical average pollution parameter and the maximum antenna abnormality degree and the maximum pollution parameter is calculated to obtain the abnormal parameter and the pollution parameter. Finally, the abnormal parameter and the pollution parameter are used to weight and fuse the deformation verification coefficient and the pollution verification coefficient, so that the result is more reasonable.

[0099] Through the specific implementation manner described above, the embodiment of the present application achieves the following technical effects:

[0100] In the embodiments of the present application, first, images are collected under the same standard to avoid interference of other factors on image collection, affect the definition of the collected images, etc., and cause inaccurate data. At the same time, based on the image of the standard antenna array substrate without defects, a division standard is established to ensure the consistency of the division standard. Second, an antenna feature recognizer is constructed and obtained through a convolutional neural network. Multiple antenna images in the antenna image array are input into the antenna feature recognizer to obtain multiple antenna features. The multiple antenna images in the antenna image array are subjected to feature recognition and abnormality analysis with the multiple standard antenna features of the multiple antenna images in the antenna array substrate image without defects to obtain an antenna abnormality array. The CNN can automatically learn features, and the antenna feature recognizer constructed and obtained through the convolutional neural network has precise feature recognition capability. The abnormality array quantifies the deviation degree of each antenna, provides input for subsequent steps, and performs feature recognition and abnormality analysis. The features are extracted, the actual features are compared with the standard features, the abnormality degree is quantified, the basic situation of the antenna array substrate defects is preliminarily reflected, and the antenna abnormality array of the antenna feature abnormality recognition is obtained. In addition, multiple antenna pollution recognition network groups are constructed through a convolutional neural network. Through the multiple antenna pollution recognition network groups, multiple pollution parameter sets are output and obtained, and a pollution parameter array is calculated. The multiple antenna pollution recognition networks are selected in a random selection manner to perform antenna pollution recognition. The multiple antenna images in the antenna image array are input respectively, and multiple pollution parameter sets are output and obtained. The pollution parameter is a pollution probability, and the larger the pollution parameter, the greater the possibility of pollution. The pollution parameter sets are subjected to mean value processing to obtain the pollution parameter of a certain antenna, and the pollution parameter array of multiple antennas is obtained by integration. The mean value processing reduces the random error of a single model. Finally, the moving window verification utilizes the spatial relationship to distinguish the local, and the similarity of the antenna abnormality and the pollution parameter in the moving window is calculated. The difference between the historical average antenna abnormality and the historical average pollution parameter and the maximum antenna abnormality and the maximum pollution parameter is calculated respectively to obtain the abnormality parameter and the pollution parameter. Finally, the abnormality parameter and the pollution parameter are used to weight and fuse the deformation verification coefficient and the pollution verification coefficient to make the result more reasonable.

[0101] Compared with the prior art, the embodiments of the present application avoid single model deviation through image processing, feature recognition, abnormality analysis, pollution identification, and spatial context verification, and fully cover various pollution scenarios with historical data to ensure model diversity. At the same time, the spatial relationship of adjacent antennas is utilized for defect detection to reduce false positives, improve the stability of antenna array substrate defect detection, ensure the accuracy of detection, and realize efficient and accurate defect detection.

[0102] It should be noted that the sequence of the above-mentioned embodiments of the present application is only for description, and does not represent the advantages and disadvantages of the embodiments. And the above describes specific embodiments of the present application. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multi-task processing and parallel processing are also possible or can be advantageous.

[0103] The above only describes the preferred embodiments of the present application and does not limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. An antenna array substrate defect detection method based on image recognition, characterized by, The method comprises: According to the preset image acquisition standard, the image of the antenna array substrate is collected, and the antenna image array is obtained according to the antenna array division standard; Feature recognition is performed on a plurality of antenna images in the antenna image array to obtain a plurality of antenna features, and abnormality analysis is performed on the plurality of standard antenna features to obtain an antenna abnormality array; According to the antenna abnormality array, a pollution recognition coefficient is configured respectively, pollution recognition is performed, and a pollution parameter array is obtained; The antenna abnormality array and the pollution parameter array are selected by a moving window, and adjacent antenna deformation verification and adjacent antenna pollution verification are performed on the selected moving window to obtain a deformation verification coefficient and a pollution verification coefficient, comprising: The first antenna abnormality window is obtained by selecting a moving window in the antenna abnormality array; Adjacent antenna deformation verification is performed on a plurality of antenna abnormalities in the first antenna abnormality window to obtain a first window deformation verification coefficient, wherein the similarity of a plurality of antenna abnormalities in the first antenna abnormality window is calculated as the first window deformation verification coefficient; The window deformation verification coefficient set and the window pollution verification coefficient set are obtained by continuing to select a moving window on the antenna abnormality array and the pollution parameter array and performing adjacent antenna deformation verification and adjacent antenna pollution verification, and the deformation verification coefficient and the pollution verification coefficient are obtained by calculating the average value respectively; The fusion defect parameter of the antenna array substrate is calculated according to the antenna abnormality array and the pollution parameter array, and the defect detection result is obtained. 2.The image recognition based antenna array substrate defect detection method according to claim 1, wherein, According to the preset image acquisition standard, the image of the antenna array substrate is collected, and the antenna image array is obtained according to the antenna array division standard, comprising: According to the preset image acquisition standard, the image of the antenna array substrate is collected; An antenna array division standard is obtained, wherein the antenna array division standard comprises a plurality of standard regions of a plurality of antenna images; According to the antenna array division standard, the images of a plurality of antennas in the antenna array substrate image are divided to obtain the antenna image array. 3.The image recognition based antenna array substrate defect detection method of claim 2, wherein, An antenna array division standard is obtained, wherein the antenna array division standard comprises a plurality of standard regions of a plurality of antenna images, comprising: A standard antenna array substrate image is obtained by collecting a defect-free antenna array substrate image according to the preset image acquisition standard; The region pixel coordinate set of the images of a plurality of antennas in the standard antenna array substrate image is divided to obtain a plurality of standard regions as the antenna array division standard. 4.The image recognition based antenna array substrate defect detection method of claim 1, wherein, Feature recognition is performed on a plurality of antenna images in the antenna image array to obtain a plurality of antenna features, and abnormality analysis is performed on a plurality of standard antenna features to obtain an antenna abnormality array, comprising: An antenna feature recognizer is obtained, wherein the antenna feature recognizer is constructed based on a convolutional neural network and is trained using a plurality of sample antenna image sets and a plurality of labeled sample antenna feature sets of a plurality of antennas, and each sample antenna feature comprises a structural feature; The plurality of antenna images in the antenna image array are input into the antenna feature recognizer, and a plurality of antenna features are output; A plurality of standard antenna features of a plurality of antenna images in a defect-free antenna array substrate image are obtained; The deviation amplitudes of the plurality of antenna features and the plurality of standard antenna features are respectively calculated as antenna anomaly degrees to obtain an antenna anomaly degree array. 5.The image recognition based antenna array substrate defect detection method of claim 1, wherein, According to the antenna anomaly degree array, a pollution identification coefficient is respectively configured to perform pollution identification to obtain a pollution parameter array, including: Obtaining a plurality of antenna pollution identification network groups corresponding to a plurality of antennas; According to each antenna anomaly degree in the antenna anomaly degree array as a plurality of pollution identification coefficients, and combining the total number of antenna pollution identification networks in each antenna pollution identification network group, a plurality of pollution identification network numbers are calculated and obtained; Respectively randomly selecting a plurality of antenna pollution identification networks of the plurality of pollution identification network numbers, inputting a plurality of antenna images in the antenna image array respectively, and outputting a plurality of pollution parameter sets and performing mean value processing to obtain a pollution parameter array. 6.The image recognition based antenna array substrate defect detection method according to claim 5, wherein, Obtaining a plurality of antenna pollution identification network groups corresponding to a plurality of antennas, including: According to historical data of antenna pollution identification, a plurality of sample antenna image sets of a plurality of antennas are collected, and a plurality of sample pollution parameter sets are obtained by identifying the pollution parameters in each sample antenna image; Based on a convolutional neural network, a plurality of antenna pollution identification network groups are respectively constructed; The plurality of sample antenna image sets and the plurality of sample pollution parameter sets are respectively randomly divided, and the plurality of antenna pollution identification network groups are integrated and trained to obtain a plurality of antenna pollution identification network groups that are trained and tested to converge.

Images