Machine vision-based data security management method and system, computer device and storage medium

Abstract

This application proposes a data security governance method, system, computer device, and storage medium based on machine vision. The method includes: acquiring a first sample image set, obtained by scanning screen content data with a scanner at different scanning angles and / or different scanning fields; augmenting the first sample image set to obtain a second sample image set; performing image denoising on the second sample image set to obtain a third sample image set; segmenting each image in the third sample image set to obtain several sub-images; establishing semantic associations between the several sub-images based on association rules to obtain several semantic feature vectors; comparing the feature vectors of the image to be detected with pre-trained semantic feature vectors of known categories, and determining the category of the image to be detected according to the maximum membership principle based on similarity. This application can efficiently and accurately predict gene expression. This application improves detection accuracy.

Description

Technical Field

[0001] This application relates to the field of cybersecurity technology, and in particular to a data security governance method, system, computer equipment, and storage medium based on machine vision. Background Technology

[0002] Machine vision-based network security monitoring enables effective 24 / 7 monitoring and alerts. By continuously analyzing acquired images through an intelligent monitoring system based on computer technology, it completely replaces previous methods relying on manual on-site inspections or pseudocode-based detection and analysis. Visual image processing technology allows for continuous, round-the-clock monitoring, alerts, and alarms for network content. This allows network security administrators to understand the situation more efficiently and promptly, criticize, manage, and stop irregular behaviors and security risks, ensuring a healthy network environment, effectively enhancing the positive energy of network content dissemination, preventing the rampant spread of malicious behavior, and greatly improving security management capabilities. Machine vision-based security governance methods can improve the accuracy of security detection and reduce unnecessary security leaks.

[0003] Traditional manual monitoring is highly susceptible to human error and interference, directly impacting monitoring quality. Machine vision-based security governance methods, in addition to their superior image processing capabilities, utilize intelligent analysis algorithms. This allows security personnel to define the characteristics of violations more precisely and accurately, enabling them to identify malicious content, discover vulnerabilities, locate them, and analyze and dispel rumors related to anomalies on the network. This effectively reduces the frequency of false alarms and missed alarms, minimizes the generation of useless data, and reduces the risk of attacks by individuals with low levels of education, weak security awareness, and a strong sense of complacency, thus lowering the incidence of security incidents. Summary of the Invention

[0004] Therefore, it is necessary to provide a data security governance method, system, computer equipment, and computer-readable storage medium based on machine vision to address the above-mentioned technical problems.

[0005] In a first aspect, embodiments of the present invention propose a data security governance method based on machine vision, the method comprising:

[0006] A first sample image set is obtained by scanning screen content data with a scanner at different scanning angles and / or different scanning fields.

[0007] The first sample image set is augmented to obtain the second sample image set;

[0008] Denoising the second sample image set yields the third sample image set;

[0009] Each image in the third sample image set is segmented to obtain several sub-images. Semantic associations are established between the several sub-images based on association rules to obtain several semantic feature vectors.

[0010] The feature vector of the image to be detected is compared with the semantic feature vector of the known categories that have been trained, and the category of the image to be detected is determined according to the principle of maximum membership based on the similarity.

[0011] In some embodiments, augmenting the first sample image set to obtain a second sample image set includes:

[0012] The first sample image set is augmented using a trained adversarial network (GAN) to obtain a second sample image set; the GAN includes a discriminator D and a generator G.

[0013] The discriminator D and the generator G implement the following game discrimination functions:

[0014]

[0015] in, This represents the generated image with noise. Let D be the probability that the discriminator classifies it as a real image. To generate the probability that the image discriminator D classifies it as a real image, Representing an image Distribution of real images The mean of the following, Indicates the generated image In random noise Distribution The mean of the values ​​below.

[0016] In some embodiments, performing image denoising on the second sample image set to obtain the third sample image set includes:

[0017] The threshold function is defined as follows:

[0018]

[0019] in, It is the grayscale value of the image. It is within the 3x3 template The grayscale value of each pixel, γ is the adjustment coefficient;

[0020] Image denoising using denoising methods:

[0021]

[0022] in, , , For the change threshold, The grayscale value of the image after threshold denoising. The corresponding image is the new image after noise reduction.

[0023] In some embodiments, the segmentation of each image in the third sample image set to obtain several sub-images includes:

[0024] For each image in the third sample image set, find the derivatives and the points that are collinear with the points that have the largest derivative values;

[0025] Connecting the collinear points yields the segmentation curves of the image;

[0026] The segmentation curve is used to segment each image in the third sample image set to obtain several sub-images.

[0027] In some embodiments, establishing semantic associations between the several sub-images based on association rules to obtain several semantic feature vectors includes:

[0028] Get Each expert assigned a target In sub-image and sub-image The combined approximation accuracy is as follows:

[0029]

[0030]

[0031] in, Experts For each sub-image Includes target The minimum membership value assigned; Experts For each sub-image Includes target The highest membership value assigned;

[0032] if , To represent the desired threshold, we define the sub-image. and sub-image The semantic content is related, linking two related sub-images. and sub-image Merge into a single image patch;

[0033] The image patch is then compared with other sub-images according to the above association rules until there are no more semantically related sub-images.

[0034] Then, proceed to check the remaining pairs of related sub-images in turn, until all sub-images have been traversed.

[0035] In some embodiments, comparing the feature vector of the image to be detected with a pre-trained semantic feature vector of a known category includes:

[0036] Calculate the feature vector of the image to be detected and the known first feature vector. Class semantic feature vectors in The first parameter direction The similarity of matching membership values ​​is:

[0037]

[0038] in, and They represent and The longest extension width of semantic factors relative to the peak point. u Indicates corresponding to semantic factors; express membership function, Indicates that the first... Class semantics in the first j The first feature parameter direction m One possible value; express Membership function; The feature vector of the image to be detected is represented at the th... Membership values ​​in each parameter direction; Represents the membership value The mean, Indicates that the first... Class semantic feature vectors in The first parameter direction The mean of the membership values.

[0039] In some embodiments, determining the category of the image to be detected based on the maximum membership principle according to similarity includes:

[0040] The vector norm method is used to determine the category of the image to be detected based on the maximum membership principle according to the similarity, so as to detect the semantics of the image to be detected.

[0041] Secondly, embodiments of the present invention propose a data security governance system based on machine vision, the system comprising:

[0042] An image acquisition module is used to acquire a first sample image set, which is obtained by scanning screen content data by a scanner at different scanning angles and / or different scanning fields.

[0043] An image augmentation module is used to augment the first sample image set to obtain a second sample image set.

[0044] An image denoising module is used to denoise the second sample image set to obtain a third sample image set;

[0045] The feature extraction module is used to segment each image in the third sample image set to obtain several sub-images, and establish semantic associations between the several sub-images based on association rules to obtain several semantic feature vectors.

[0046] The governance decision module is used to compare the feature vector of the image to be detected with the semantic feature vector of the known categories that have been trained, and to determine the category of the image to be detected according to the principle of maximum membership based on the similarity.

[0047] Thirdly, embodiments of the present invention provide a computer device including a memory and a processor, wherein the memory stores a computer program and the processor executes the steps described in the first aspect.

[0048] Fourthly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, wherein the processor executes the computer program to implement the steps described in the first aspect.

[0049] The aforementioned method, system, computer equipment, and storage medium acquire a first sample image set, which is obtained by scanning screen content data with a scanner at different scanning angles and / or different scanning fields; the first sample image set is augmented to obtain a second sample image set; the second sample image set is denoised to obtain a third sample image set; each image in the third sample image set is segmented to obtain several sub-images, and semantic associations are established between the several sub-images based on association rules to obtain several semantic feature vectors; the feature vectors of the image to be detected are compared with the semantic feature vectors of known categories that have been trained, and the category of the image to be detected is determined according to the principle of maximum membership based on similarity, so as to improve the detection accuracy. Attached Figure Description

[0050] Figure 1 This is a flowchart illustrating a data security governance method based on machine vision in one embodiment;

[0051] Figure 2 This is a schematic diagram of the architecture of a Genetic Adversarial Network (GAN) in one embodiment;

[0052] Figure 3 This is a flowchart illustrating an image segmentation method in one embodiment;

[0053] Figure 4 This is a schematic diagram of the structure of a machine vision-based data security governance system in one embodiment. Detailed Implementation

[0054] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are merely some examples or embodiments of the present invention. For those skilled in the art, the present invention can be applied to other similar scenarios based on these drawings without creative effort. Unless obvious from the context or otherwise specified, the same reference numerals in the drawings represent the same structures or operations.

[0055] As indicated in this invention and the claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" do not specifically refer to the singular and may also include the plural. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0056] While this invention makes various references to certain modules in an apparatus according to embodiments of the invention, any number of different modules can be used and run on a computing device and / or processor. Modules are merely illustrative, and different aspects of the apparatus and methods may use different modules.

[0057] It should be understood that when a unit or module is described as "connected" or "coupled" to other units, modules, or blocks, it may refer to a direct connection or coupling, or communication with other units, modules, or blocks, or the presence of intermediate units, modules, or blocks, unless the context explicitly indicates otherwise. The term "and / or" as used herein may include any and all combinations of one or more of the related listed items.

[0058] like Figure 1 As shown, this embodiment of the invention provides a data security governance method based on machine vision, including the following steps:

[0059] S102, acquire a first sample image set, which is obtained by scanning screen content data by a scanner at different scanning angles and / or different scanning fields.

[0060] By capturing screen display content from different perspectives, a large amount of image data from different sources, multiple angles, and multiple fields of view is obtained.

[0061] S104, the first sample image set is augmented to obtain the second sample image set.

[0062] To increase the sample size for machine deep learning, the first sample image set is augmented to overcome the problem of insufficient data samples.

[0063] S106, perform image denoising on the second sample image set to obtain the third sample image set.

[0064] To enable the pointer to read the scanned image data more clearly and accurately, and to provide better deep learning training samples, image denoising is performed on the second sample image set.

[0065] S108, each image in the third sample image set is segmented to obtain several sub-images, and semantic associations are established between the several sub-images based on association rules to obtain several semantic feature vectors.

[0066] More accurate features can be extracted through segmentation and semantic association.

[0067] S110, compare the feature vector of the image to be detected with the semantic feature vector of the known categories that have been trained, and determine the category of the image to be detected according to the principle of maximum membership based on the similarity.

[0068] Through machine learning, a feature vector database with specified standards is established, and this database is used as the standard library. The feature vectors of the image to be detected are compared with the semantic feature vectors of known categories in the standard library.

[0069] In step 102, to address the issue of capturing screen content from multiple sources, varying magnitudes, and multiple spatiotemporal sequences of network data using cameras, a high-definition, multi-directional, and high-resolution image acquisition device with different viewing angles and fields of view is designed. Simultaneously, it is also necessary to establish the relative positions of the multi-source cameras, the screen, and the imaging plane with respect to the screen coordinate system.

[0070] In practical data acquisition, the number of images collected is usually limited due to constraints. However, accurately identifying network content using image processing algorithms and achieving security detection requires building deep learning models, which in turn require large datasets for training. Therefore, using augmented datasets for deep learning is essential. Deep learning networks can accurately classify and locate objects in images with varying sizes, positions, and angles. This is a key reason why data augmentation can improve the accuracy of model detection.

[0071] In step 104, the process of augmenting the first sample image set to obtain the second sample image set includes:

[0072] Based on adversarial networks GAN (Generative Adversarial Network) Image augmentation is a deep learning model consisting of two parts: a generator G and a discriminator D. These two parts learn through a game-like process, with the generator G producing an image similar to the original image, which is then compared to the original image in the discriminator D to ultimately determine whether the generated image is real or fake. The architecture of a Generative Adversarial Network (GAN) is as follows: Figure 2 As shown.

[0073] First, fix the generator G and train the discriminator D to better distinguish whether the input image is a real sample or a fake sample; then iteratively train the discriminator D... k After each training iteration, the learning rate parameter of the discriminator D is updated. This process is repeated during the training of the discriminator D. k After one training iteration, a smaller learning rate is chosen to train the generator network, minimizing the difference between the generated images and the original images; training k After the discriminator D and the generator G, the discriminator D is unable to determine whether the input image is a real image or a generated image.

[0074] Since the key to data augmentation is the continuous optimization of the parameters in the discriminant function, the discriminant function for G and D in a GAN (Generative Adversarial Network) to implement the game is defined as follows:

[0075] (1)

[0076] here, or , The maximum value that the generated image is judged as the real image by the discriminator D is a given expected value; The maximum value that the discriminator D determines to be a real image is a given expected value. x Representing real images; z The generator network G represents the input. x The generated random noise, This represents the generated image with noise. Let D be the probability that the discriminator classifies it as a real image. The probability that the generated image is judged as a real image by the discriminator D. Representing an image x Distribution of real images The mean below, similarly, Indicates the generated image In random noise z Distribution The mean of the values ​​below. For generator G, it hopes that the images it generates are close to real images and can be judged as real images by the judge. That is, hope =0, at this time For the discriminator network D, the goal is to accurately distinguish between genuine and fake images generated by the generator. As large as possible That is, hope At this time The probabilities here refer to the ratio of the number of judgments to the total number of judgments, that is:

[0077] , , (2)

[0078] here, Indicates the judgment of discriminator D. x The number of times a real image is used; Indicates the judgment of discriminator D. The number of times a real image is used; This represents the total number of judgments made by the discriminator D.

[0079] The discriminant function of the adversarial network consists of perceptual probability and adversarial probability, used to measure the difference between the generated image and the real image, and the realism of the generated image. By alternately optimizing the discriminator and generator, the network parameters are continuously adjusted, ultimately achieving high-quality super-resolution image generation. Finally, the real high-resolution image and the virtual high-resolution image are input into the next layer of the feature extraction network system to extract their features.

[0080] The main task of the discriminator network is to determine the realism of the input and output images. It primarily consists of several repeated convolutional layers, activation functions, and a normalization process. When an input image is fed to the discriminator, its image processing is similar to that of a generator network, which performs convolution, normalization, activation function enhancement filtering, and ultimately outputs a judgment on the image's realism. The goal is for the discriminator to classify the generated image as a real image, meaning the label output by the discriminator should be close to the real image. This indicates that the discriminator considers the newly generated image to be true; conversely, the closer the result is to reality, the less likely it is to be true. At this point, the discriminator deems the generated image unacceptable. Images deemed authentic by the discriminator are then added to a standard database to achieve the data augmentation required for deep learning.

[0081] In traditional models, mean squared error is often used as the loss function for training networks. However, this may result in a high signal-to-noise ratio and unsatisfactory image performance after the operation, and may lead to the loss of some important feature information of the network data. This embodiment enhances the realism of the restored image by optimizing the first term, perceptual probability, and the second term, adversarial probability, in formula (1). The main purpose is to generate high-resolution images and increase the number of samples by preserving the feature information of the original image.

[0082] The generator mainly processes the received low-resolution images. x As input, a high-resolution image is generated using a 3×3 convolutional template in the generator. The specific steps are as follows: First, the low-resolution image is input, processed by a 3×3 convolutional template, and then activated by an activation function. ,here The input image is used. The second step involves performing two convolution operations in the next layer's network structure with two convolution operations: first, a template weight update operation, and then a convolution operation calculating the difference between the generated and real images. Simultaneously, a normalization function from the regression layer is used. Perform smooth normalization and activation function The enhancement filtering process involves two steps. Here, It is the energy value of the image. It is the median energy value in the 3×3 template convolution image region. It is the image after template convolution. It is the frequency of the image after convolution. and The threshold is given empirically. After the first two steps, the image's feature matrix is ​​passed to the next sampling stage. After three samplings, the dimensions (length, width, and height) of the original input image are increased by 8 times, resulting in improved clarity.

[0083] Due to various factors, including image acquisition systems and different physical phenomena such as the imperfect uniformity of illumination, the edge intensity of acquired images varies. For example, images with strong light require toning, while backlit images need to be intensified to increase tone. Furthermore, in real-world scenarios, the presence of varying noise around the target, combined with the characteristics of the scene, makes subsequent interpretation extremely difficult.

[0084] In step 106, the step of performing image denoising on the second sample image set to obtain the third sample image set includes:

[0085] First, define the threshold function as follows:

[0086] (3)

[0087] in, It is the grayscale value of the image. It is within the 3x3 template The grayscale value of each pixel, γ is the adjustment coefficient;

[0088] By using the transformation threshold defined above, the image can be... x Noise reduction is implemented. This is based on a defined threshold. ε All gray values ​​less than or equal to ε are taken from the original image value; their values ​​are still... ， Exceeding the threshold ε The values ​​are classified as noise and replaced with new values. This method means that thresholding removes small-amplitude noise or unwanted signals. After template transformation, the desired image is obtained. The noise reduction method is given below:

[0089] (4)

[0090] in, , , For the change threshold, The grayscale value of the image after threshold denoising. The corresponding image is the new image after noise reduction.

[0091] Using the denoising method described above, after denoising the noisy image, the mean square error of the noisy image is reduced, the signal-to-noise ratio is improved, and the denoised image is clearer and smoother.

[0092] In step 108, as Figure 3 As shown, the segmentation of each image in the third sample image set to obtain several sub-images includes:

[0093] S402, calculate the derivative of each image in the third sample image set, and find the points that are collinear with the points with the largest derivative values;

[0094] S404, Connect the collinear points to obtain the segmentation curve of the image;

[0095] S406, the segmentation curve is used to segment each image in the third sample image set to obtain several sub-images.

[0096] Image semantic extraction is performed in two steps: first, the images in the third sample image set are segmented; then, semantic association is performed on the segmented images. The first step, image segmentation, is implemented as follows:

[0097] If we take the derivative of the image obtained above, the curve formed by connecting the points that are collinear with the point of maximum derivative value is the segmentation curve for the semantics of the image. By finding the collinear points, we can construct the equation of the straight line. By finding the key points of each line, the corresponding line can be determined. , Values, yielding several points. Connect all the points. The resulting curve is a segmented curve.

[0098] The method for determining specific dividing lines is: in the image In the middle, put each pair , Treat it as a constant and solve the equation. Determine several pairs , Value. Connect these points. The resulting curve is a segmented curve. This segmentation method has strong anti-interference capabilities.

[0099] However, directly solving the equation of the line cannot yield the dividing line in all cases. For example, if the line tends to be perpendicular, then... ,equation There is no solution. Therefore, a more general solution is to use parametric equations: According to this equation, points on a straight line in the image are mapped to... A sine curve in space. Following the method described above for solving the equation of a straight line, solve the parametric equation to obtain several parametric coordinate points. Connect these coordinate points to achieve image segmentation. The basic procedure for creating segmentation lines is as follows:

[0100] 1) Construct a P, Q Two-dimensional accumulation array in space ;

[0101] 2) From Take from the specified area According to the equation exist Take all possible values ​​from the middle Value calculation yields possible value;

[0102] 3) Calculate at the corresponding position ;

[0103] 4) Repeat steps 2) and 3) until you will be able to... All points within the specified area have been retrieved. At this point, The maximum value in the array 、 That is, the equation In 、 value;

[0104] 5) According to Draw The straight line in the middle.

[0105] In step 108, the step of establishing semantic associations between the several sub-images based on association rules to obtain several semantic feature vectors includes:

[0106] To establish semantic associations, the segmented images described above are connected according to the association rules defined below, resulting in mutually independent features with different semantics. These features are then stored in the feature database as a benchmark expert database.

[0107] Semantic association is defined by the following rules: Let Is the image region containing the target? The minimum region and the maximum possible region set. and Experience or expertise (hereinafter collectively referred to as experts) Give a goal The minimum and maximum membership values ​​that appear in the selected region are called... For experts Give The approximate accuracy. Here, .but A comprehensive assessment by experts X approximate accuracy Through comprehensive and The calculation is as follows:

[0108] (5)

[0109] Additionally, if the image Divided into Sub-images Image region Sub-images Under the same conditions for the same target Assignment, then experts For each sub-image Includes target The minimum and maximum membership values ​​assigned and Experts Give The approximate accuracy is:

[0110] (6)

[0111] Then we get, Each expert assigned In sub-image and sub-image The overall approximate accuracy of the regions is as follows:

[0112] (7)

[0113] (8)

[0114] Given an expected threshold ,if Then define a sub-image and sub-image If the semantic content of a region is related, then two related semantic sub-images are merged into one block. Then, the same method is used to detect the relatedness of this block with other sub-images until no more semantically related sub-images remain. This block is then stored in a predefined expert library. Next, the remaining pairwise related sub-images are detected sequentially until all related sub-images have been traversed. The newly generated, independent semantically related blocks are then stored in a predefined expert library as a baseline expert library.

[0115] In step 110, comparing the feature vector of the image to be detected with the semantic feature vector of a known category that has already been trained includes:

[0116] In the feature matching network structure, the feature vector of the image to be detected with unknown semantic content is compared with the semantic feature vector of the known categories that has been trained. A match is made if and only if its feature vector matches the first... i When the similarity between the feature vectors of the class is maximized, the target to be detected is determined to belong to the class according to the maximum membership principle. i Class. The specific matching algorithm is given below: Assume there are a total of n There are 3 semantic classes, and each semantic feature vector is composed of 1 semantic class. k It consists of several feature parameters, such as semantics like anger, negativity, and positivity. And it is assumed that the first... Class semantics in the first In the direction of each characteristic parameter Each value , Indicates that the first... i Class semantics in the first j The first parameter direction m Membership degree value, Let it be its mean. The detected semantics are in the first position. Membership values ​​in each parameter direction, Let it be its mean. and They represent and The membership function, based on experience, can be defined as follows: and ,have:

[0117] and (9)

[0118] here, u It corresponds to semantic factors, and They represent and The longest extension width of semantic factors relative to the peak point.

[0119] To determine the type of the detected semantics, it is necessary to calculate and Similarity between .because yes and The supremum of the intersection, that is, the boundary between the two distribution curves at... and The height of the intersection between them, therefore by achievable .

[0120] Therefore:

[0121] (10)

[0122] Therefore, we can obtain the detected semantics and the known semantics. Class semantics in The first parameter direction The similarity of matching each value is:

[0123] (11)

[0124] In step 110, determining the category of the image to be detected based on the maximum membership principle according to similarity includes:

[0125] Due to the Class semantics in the first In the direction of each characteristic parameter Therefore, the th value is... Class semantic common 1 eigenvector A total of semantic classes There are eigenvectors. Therefore, from formula (11), we have:

[0126] (12)

[0127] Therefore, the detected vector and the first... The semantic similarity vector is .

[0128] The output of the governance decision-making network adopts the vector norm method, that is, if , making Then, according to the principle of maximum membership, the semantic term to be detected is determined to belong to the first... Classes are used to detect semantics. Here, It is a vector norm. It is a set of indicators.

[0129] It should be understood that although the steps in the flowchart above are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps.

[0130] In one embodiment, such as Figure 4 As shown, this embodiment of the invention provides a data security governance system based on machine vision, the system comprising:

[0131] Image acquisition module 502 is used to acquire a first sample image set, which is obtained by scanning screen content data by a scanner at different scanning angles and / or different scanning fields;

[0132] The image augmentation module 504 is used to augment the first sample image set to obtain a second sample image set.

[0133] Image denoising module 506 is used to denoise the second sample image set to obtain a third sample image set;

[0134] The feature extraction module 508 is used to segment each image in the third sample image set to obtain several sub-images, and establish semantic associations between the several sub-images based on association rules to obtain several semantic feature vectors.

[0135] The governance decision module 510 is used to compare the feature vector of the image to be detected with the semantic feature vector of the known categories that have been trained, and to determine the category of the image to be detected according to the principle of maximum membership based on the similarity.

[0136] In one embodiment, the present invention provides a computer device including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps in any of the above embodiments of the data security governance method based on machine vision.

[0137] In one embodiment, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in any of the above embodiments of the data security governance method based on machine vision.

[0138] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0139] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0140] The above embodiments merely illustrate several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A data security governance method based on machine vision, characterized in that, The method includes: A first sample image set is obtained by scanning screen content data with a scanner at different scanning angles and / or different scanning fields. The first sample image set is augmented to obtain the second sample image set; Denoising the second sample image set yields the third sample image set; Each image in the third sample image set is segmented to obtain several sub-images. Semantic associations are established between the several sub-images based on association rules to obtain several semantic feature vectors. The feature vector of the image to be detected is compared with the semantic feature vector of the known categories that have been trained, and the category of the image to be detected is determined according to the principle of maximum membership based on the similarity. The segmentation of each image in the third sample image set yields several sub-images, including: For each image in the third sample image set, find the derivatives and the points that are collinear with the points that have the largest derivative values; Connecting the collinear points yields the segmentation curves of the image; The segmentation curve is used to segment each image in the third sample image set to obtain several sub-images; The step of establishing semantic associations between the several sub-images based on association rules to obtain several semantic feature vectors includes: Get Each expert assigned a target In sub-image and sub-image The combined approximation accuracy is as follows: in, Experts For each sub-image Includes target The minimum membership value assigned; Experts For each sub-image Includes target The highest membership value assigned; if , To represent the desired threshold, we define the sub-image. and sub-image The semantic content is related, linking two related sub-images. and sub-image Merge into a single image patch; The image patch is then compared with other sub-images according to the above association rules until there are no more semantically related sub-images. Then, proceed to check the remaining pairs of related sub-images in turn, until all sub-images have been traversed.

2. The method according to claim 1, characterized in that, The step of augmenting the first sample image set to obtain the second sample image set includes: The first sample image set is augmented using a trained adversarial network (GAN) to obtain a second sample image set; the GAN includes a discriminator D and a generator G. The discriminator D and the generator G implement the following game discrimination functions: in, This represents the generated noisy image. Let D be the probability that the discriminator classifies it as a real image. To generate the probability that the image discriminator D classifies it as a real image, Representing an image Distribution of real images The mean of the following, Indicates the generated image In random noise Distribution The mean of the values ​​below.

3. The method according to claim 1, characterized in that, The step of performing image denoising on the second sample image set to obtain the third sample image set includes: The threshold function is defined as follows: in, It is the grayscale value of the image. It is within the 3x3 template The grayscale value of each pixel, γ is the adjustment coefficient; Image denoising using denoising methods: in, , , For the change threshold, The grayscale value of the image after threshold denoising. The corresponding image is the new image after noise reduction.

4. The method according to claim 1, characterized in that, The step of comparing the feature vector of the image to be detected with the semantic feature vector of a known category that has already been trained includes: Calculate the feature vector of the image to be detected and the known first feature vector. Class semantic feature vectors in The first parameter direction The similarity of matching membership values ​​is: in, and They represent and The longest extension width of semantic factors relative to the peak point. u Indicates corresponding to semantic factors; express membership function, Indicates that the first... Class semantics in the first j The first feature parameter direction m One possible value; express Membership function; The feature vector of the image to be detected is represented at the th... Membership values ​​in each parameter direction; Represents the membership value The mean, Indicates that the first... Class semantic feature vectors in The first parameter direction The mean of the membership values.

5. The method according to claim 1, characterized in that, The method of determining the category of the image to be detected based on the maximum membership principle according to similarity includes: The vector norm method is used to determine the category of the image to be detected based on the maximum membership principle according to the similarity, so as to detect the semantics of the image to be detected.

6. A data security governance system based on machine vision, characterized in that, The system includes: An image acquisition module is used to acquire a first sample image set, which is obtained by scanning screen content data by a scanner at different scanning angles and / or different scanning fields. An image augmentation module is used to augment the first sample image set to obtain a second sample image set. An image denoising module is used to denoise the second sample image set to obtain a third sample image set; The feature extraction module is used to segment each image in the third sample image set to obtain several sub-images, and establish semantic associations between the several sub-images based on association rules to obtain several semantic feature vectors. The governance decision module is used to compare the feature vector of the image to be detected with the semantic feature vector of the known categories that have been trained, and to determine the category of the image to be detected according to the principle of maximum membership based on the similarity. The segmentation of each image in the third sample image set yields several sub-images, including: For each image in the third sample image set, find the derivatives and the points that are collinear with the points that have the largest derivative values; Connecting the collinear points yields the segmentation curves of the image; The segmentation curve is used to segment each image in the third sample image set to obtain several sub-images; The step of establishing semantic associations between the several sub-images based on association rules to obtain several semantic feature vectors includes: Get Each expert assigned a target In sub-image and sub-image The combined approximation accuracy is as follows: in, Experts For each sub-image Includes target The minimum membership value assigned; Experts For each sub-image Includes target The highest membership value assigned; if , To represent the desired threshold, we define the sub-image. and sub-image The semantic content is related, linking two related sub-images. and sub-image Merge into a single image patch; The image patch is then compared with other sub-images according to the above association rules until there are no more semantically related sub-images. Then, proceed to check the remaining pairs of related sub-images in turn, until all sub-images have been traversed.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

Images