Online education platform user authentication method and system

Abstract

The present invention discloses a user authentication method and system for an online education platform, which solves the problem that the existing method adopts frame-by-frame analysis of facial images and can only simply process each frame independently and is sensitive to interference information. The method comprises: normalizing target image information to obtain normalized target image information; combining a multi-layer spatial displacement perception mechanism and a cascaded posture regression model to construct a multi-target recognition model, the multi-target recognition model performs noise reduction processing on the target image information and then segments and extracts feature points, and calculates the confidence of the target image information and the standard facial image; the present invention can overcome the distraction of the model caused by background information by normalizing the target image information, and can quickly remove background interference factors; the multi-target recognition model approximates the real face shape and feature point position by cluster analysis, and corrects, reduces noise and optimizes the results of the previous level, thereby improving the positioning accuracy of the feature points.

Description

Technical Field

[0001] The present invention belongs to the field of face recognition technology, and specifically relates to a user authentication method and system for an online education platform. Background Art

[0002] Facial recognition is a technology that identifies or verifies personal identity by analyzing and comparing facial features. It integrates computer image processing technology with biostatistics principles, and uses computer image processing technology to extract character feature points from videos. At the same time, with the popularization of personal computers, the application of facial recognition has penetrated into many corners of people's daily lives. This is also reflected in education, training and examination activities. The application of facial recognition technology on online education platforms, especially in sign-in authentication, has significant advantages.

[0003] Chinese patent CN112232206B discloses a face recognition method and face recognition platform based on big data and artificial intelligence. First, the background of the target image to be recognized is segmented to obtain a target background image and a target face image; second, the target background image is matched with each frame of the reference image in the first target database to obtain a target matching result; then, based on the target matching result, a target recognition rule is determined from multiple pre-formed recognition rules; finally, the target face image is recognized based on the target recognition rule. However, the existing method uses frame-by-frame analysis of face images and can only simply process each frame independently. It is sensitive to interference information and cannot fully utilize the correlation and continuity between previous and next frames, resulting in limited accuracy of the recognition results. To address the above problems, we propose a user authentication method and system for an online education platform. Summary of the Invention

[0004] The purpose of the present invention is to address the shortcomings of the existing technology and provide a user authentication method and system for an online education platform, which solves the problem that the existing method uses frame-by-frame analysis of facial images and can only simply process each frame independently, is sensitive to interference information, and cannot fully utilize the correlation and continuity between previous and next frames, resulting in limited accuracy of recognition results.

[0005] The present invention is implemented as follows: an online education platform user authentication method, the online education platform user authentication method comprising:

[0006] Real-time collection of target image information in the authentication scene, normalization of the target image information, and obtaining normalized target image information;

[0007] Based on the image acquisition device, standard facial images are pre-collected, an image database is constructed using the standard facial images, the image database is traversed, standard facial images are extracted from the image database, and the standard facial images are processed based on different angles, brightness, and noise factors to create a complex image set. A multi-target recognition model is constructed by combining a multi-layer spatial displacement perception mechanism and a cascaded posture regression model. The multi-target recognition model is iteratively trained using the complex image set to output a converged multi-target recognition model.

[0008] Taking the normalized target image information as input, the multi-target recognition model is executed. The multi-target recognition model performs noise reduction processing on the target image information, then segments and extracts feature points. The feature points are fused based on the layer space displacement perception mechanism, and a feature fusion set is output. The confidence level of the target image information and the standard face image is calculated, and it is determined whether the confidence level of the target image information and the standard face image meets the confidence threshold.

[0009] If the confidence level of the target image information and the standard face image meets the confidence threshold, it is determined that the target image information in the scene is successfully authenticated, and the target image information in the scene is entered to complete the education platform user authentication and identification.

[0010] Preferably, the method for normalizing target image information specifically includes:

[0011] Obtain target image information, process the target image information using a two-dimensional empirical mode decomposition algorithm, and generate a low-discrepancy point set based on cardinality flip sampling;

[0012] Among them, the base flip sampling converts the target image information into integer The exponential expansion of is:

[0013] (1)

[0014] is the input target image information, Cardinality expansion information representing target image information;

[0015] The cardinality expansion information is sampled l times to generate a Hammersley sequence;

[0016] The Hammersley sequence is represented as:

[0017] (2)

[0018] in, Indicates the number of sampling times, Indicates the projection direction vector number;

[0019] Calculate the average value of the mean envelope in all projection directions to obtain the low-discrepancy point set;

[0020] (3)

[0021] Obtain a low-difference point set, and normalize the grayscale of the low-difference point set by combining histogram equalization and gamma correction to obtain a grayscale normalized set;

[0022] Load the grayscale normalization set and combine it with the PP-HumanSeg algorithm to remove the background pigment in the grayscale normalization set to achieve normalization processing of the target image information.

[0023] Preferably, the method for processing a standard face image based on different angles, brightness, and noise factors specifically includes:

[0024] Obtain a standard face image, randomly rotate the standard face image at a set rotation angle (-30°-30°), and mirror-flip the randomly rotated standard face image to generate a random angle set;

[0025] Load a random angle set, adjust the brightness of the random angle set within a preset brightness range (10-150%) based on the multi-scale Retinex algorithm, and output the brightness-adjusted random angle set;

[0026] Among them, the random angle set output is expressed as:

[0027] (4)

[0028] in, is the input representation of a random angle set, represents the illumination component of the random angle set data, Represents the pixel value gain factor;

[0029] Get a random angle set after brightness adjustment, and simultaneously add preset proportions of salt and pepper noise and Gaussian noise to the random angle set to create a complex image set.

[0030] Preferably, the method of iteratively training a multi-target recognition model using a complex image set specifically includes:

[0031] The cascaded pose regression model is used as the initial model of the multi-target recognition model. The initial model consists of a shape estimation layer, a feature extraction layer, and a cascade regressor. The shape estimation layer is connected to the feature extraction layer, which is then connected to the cascade regressor. The feature extraction layer includes three groups of convolutional layers, one group of pooling layers, one group of upsampling layers, and one group of feature fusion layers.

[0032] A confusion matrix analysis module is introduced in the shape estimation layer, and the pooling layer in the feature extraction layer is frozen, and a layer-space displacement perception mechanism is introduced to replace the pooling layer;

[0033] Load a complex image set and perform hybrid enhancement on it based on Mosaic and Mixup hybrid enhancement techniques to obtain 40 groups of facial posture images with an interval of 2. Divide the hybrid enhanced complex image set into a training set and a test set.

[0034] Obtain the training set and set the initial model parameters. The initial model has the following parameters: the number of feature points, the number of clustering iterations is 20, the number of random forest decision trees is 250, the feature point radius is 2, and the pixel feature is 400-500.

[0035] Taking the training set as input, the initial model is trained based on random forest local encoding and K-means cluster analysis until the initial model converges, and at least one set of initial models in the multi-view recognition mode is obtained;

[0036] A test set is obtained, and the test set is used as input. At least one set of initial models is used to perform posture evaluation and confidence evaluation on the test set to obtain a posture deflection angle and an image confidence.

[0037] Preset the measurement standard threshold, combine the posture deflection angle and image confidence to calculate the model recognition accuracy of the initial model, select the initial model whose model recognition accuracy meets the measurement standard threshold range as the final multi-target recognition model, and output the converged multi-target recognition model;

[0038] The model recognition accuracy is calculated by the following formula:

[0039] (5)

[0040] in, represents the model recognition accuracy, is the number of feature points, represents the attitude deflection angle, Represents the image confidence.

[0041] Preferably, the method of fusing feature points based on the layer space displacement perception mechanism and outputting a feature fusion set specifically includes:

[0042] Taking the normalized target image information as input, the multi-target recognition model is executed. The shape estimation layer makes a preliminary estimate and judgment of the target image information based on the confusion matrix algorithm to determine whether the target image information is a living image;

[0043] If the target image information is a living body image, the target image information is pushed to the feature extraction layer;

[0044] The feature extraction layer obtains the target image information, introduces an equivalent weighted filter to perform low-pass filtering and noise reduction on the target image information, filters out the interference components of the target image information, and outputs the noise reduction information set;

[0045] Obtain the denoising information set, segment and extract feature points after denoising the target image information in the feature extraction layer, use the domain classifier to extract the bounding box and facial feature information of the denoising information set, and perform ROI pooling with the feature map of the corresponding scale of the standard face image;

[0046] Based on the layer space displacement perception mechanism, feature points are fused and processed to output a feature fusion set;

[0047] Calculate the confidence of target image information and standard face image;

[0048] Determine whether the confidence level of the target image information and the standard face image meets the confidence threshold;

[0049] The confidence of the target image information and the standard face image is calculated using the following formula:

[0050] (9)

[0051] in, Respectively represent the target image information and the fusion value of the face standard image feature, is the label confidence coefficient of the target image information;

[0052] The noise reduction information set is expressed as:

[0053] (6)

[0054] (7)

[0055] (8)

[0056] in, is the output representation of the denoised information set, represents the equivalent weighted filter, Indicates the target image information input information, Indicates the number of times of filtering and removing impurities, is the amplitude response of the filter, represents the filter signal-to-noise ratio, represents the overlapping correlation coefficient, is the overlap proportionality constant.

[0057] On the other hand, the present invention also provides an online education platform user authentication system, which specifically includes:

[0058] The data normalization module is used to collect target image information in the authentication scene in real time, normalize the target image information, and obtain normalized target image information;

[0059] The recognition modeling module pre-collects standard facial images based on image acquisition equipment, builds an image database based on these standard facial images, traverses the image database, extracts standard facial images from the image database, processes the standard facial images based on different angles, brightness, and noise factors, creates a complex image set, and builds a multi-target recognition model by combining a multi-layer spatial displacement perception mechanism and a cascaded posture regression model. The multi-target recognition model is iteratively trained using the complex image set to output a converged multi-target recognition model.

[0060] The face recognition module is used to execute the multi-target recognition model. The multi-target recognition model performs noise reduction on the target image information, then segments and extracts feature points. It then fuses the feature points based on the layer-space displacement perception mechanism, outputs a fusion feature set, and calculates the confidence level between the target image information and the standard face image.

[0061] The authentication judgment module is used to judge whether the confidence of the target image information and the standard face image meets the confidence threshold. If the confidence of the target image information and the standard face image meets the confidence threshold, it is judged that the authentication of the target image information in the scene is successful, and the target image information in the scene is entered to complete the education platform user authentication and identification.

[0062] Preferably, the data normalization module specifically includes:

[0063] A modal decomposition unit is used to obtain target image information, process the target image information using a two-dimensional empirical mode decomposition algorithm, and generate a low-discrepancy point set based on cardinality flip sampling;

[0064] A grayscale processing unit is used to obtain a low-difference point set, and normalize the grayscale of the low-difference point set by combining histogram equalization and gamma correction to obtain a grayscale normalized set;

[0065] The background selection unit is used to load the grayscale normalization set and, combined with the PP-HumanSeg algorithm, remove the background pigment in the grayscale normalization set to achieve normalization processing of the target image information.

[0066] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0067] In an embodiment of the present invention, the problem of background information distracting the model and causing misidentification or missed identification can be overcome by normalizing the target image information. By normalizing the target image information, background interference factors can be quickly removed, so that the multi-target recognition model can be more focused on feature extraction of the face area, thereby improving the accuracy of recognition. At the same time, a multi-target recognition model is constructed by combining a multi-layer spatial displacement perception mechanism and a cascaded posture regression model. The multi-target recognition model approximates the real face shape and feature point position through cluster analysis, and corrects, reduces noise and optimizes the results of the previous level, thereby improving the positioning accuracy of the feature points. It overcomes the problem that the existing method uses frame-by-frame analysis of face images and can only simply process each frame independently, is sensitive to interference information, and cannot fully utilize the correlation and continuity between the previous and next frames, resulting in the accuracy of the recognition results being limited.

[0068] In an embodiment of the present invention, two-dimensional empirical mode decomposition is performed on the target image information to generate a low-difference point set, and then facial features that are not affected by illumination are extracted. These features remain relatively stable under different illumination conditions, which can reduce the interference of illumination changes on face recognition and improve the accuracy and reliability of recognition. Then, the grayscale normalization processing of the low-difference point set is combined with histogram equalization and gamma correction, which can overcome the situation where some grayscale levels in the low-difference point set are relatively concentrated. Histogram equalization can effectively widen the gap between these grayscale levels, thereby enhancing the overall contrast of the image, which makes the details in the face image clearer. Finally, background pigment removal in the grayscale normalization set can highlight the main features of the face and reduce interference factors, thereby improving the robustness and efficiency of the multi-target recognition model.

[0069] In an embodiment of the present invention, a complex image set is created based on different angles, brightness, and noise factors. This training image set, which incorporates a variety of angles, brightness, and noise factors, provides the model with comprehensive and rich learning material, helping it to more accurately capture the key features and detailed information of a face. This significantly improves the model's accuracy in recognizing faces, while overcoming the issues of image quality, angle, and brightness that may arise when multi-target recognition models are run on different hardware devices. It also meets the needs of face recognition in strong outdoor light, dim indoor light, and during motion.

[0070] In an embodiment of the present invention, a multi-target recognition model is constructed by combining a multi-layer spatial displacement perception mechanism and a cascaded posture regression model, and a confusion matrix analysis module is introduced into the multi-target recognition model, so that changes in the face under different postures can be effectively processed. For example, when the face is in different postures such as sideways, looking up, and lowering the head, the multi-target recognition model can gradually correct the influence of the posture on the facial features, so that the model can more accurately extract the key features of the face, further improve the accuracy of recognition, and thus make the entire multi-target recognition model more robust to complex environmental factors. Whether it is a face image from the front or side, or a face image with expression changes or partial occlusion, the model can perform recognition well, which improves the application range and reliability of face recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 It is a schematic diagram of the implementation flow of the online education platform user authentication method provided by the present invention.

[0072] Figure 2 The figure shows a schematic diagram of the implementation process of the method for normalizing target image information.

[0073] Figure 3 The figure shows a schematic diagram of the implementation process of the method for processing standard face images based on different angles, brightness and noise factors.

[0074] Figure 4 The figure shows the implementation process of the iterative training method for multi-target recognition model using complex image sets.

[0075] Figure 5 The figure shows a schematic diagram of the implementation process of the method for fusing feature points based on the layer space displacement perception mechanism and outputting a feature fusion set.

[0076] Figure 6 It is a structural diagram of the online education platform user authentication system provided by the present invention. DETAILED DESCRIPTION

[0077] Unless otherwise defined, all technical and scientific terms used herein have the same meanings as commonly understood by those skilled in the art to which this application belongs. The terms used in the specification of the application are only for the purpose of describing specific embodiments and are not intended to limit this application. The terms "including" and "having" and any variations thereof in the specification and claims of this application and the above-mentioned drawings are intended to cover non-exclusive inclusions. The terms "first", "second", etc. in the specification and claims of this application or the above-mentioned drawings are used to distinguish different objects, not to describe a specific order.

[0078] Existing methods that analyze facial images frame by frame can only simply process each frame independently, are sensitive to interference information, and cannot fully utilize the correlation and continuity between previous and subsequent frames, resulting in limited accuracy of recognition results. To address the above problems, we propose a user authentication method and system for online education platforms. In short, when executing the user authentication method for online education platforms, the target image information in the authentication scene is first collected in real time, the target image information is normalized to obtain normalized target image information, and an image database is constructed using standard facial images. The standard facial images are processed based on different angles, brightness, and noise factors to create a complex image set. A multi-layer spatial displacement perception mechanism and a cascaded posture regression model are combined to construct a multi-target recognition model. The multi-target recognition model is used to reduce noise in the target image information and then segment and extract feature points. The feature points are fused based on the layer spatial displacement perception mechanism to output a fusion feature set. The confidence of the target image information and the standard facial image is calculated, and it is determined whether the confidence of the target image information and the standard facial image meets the confidence threshold. If the confidence of the target image information and the standard facial image meets the confidence threshold, the target image information in the scene is deemed to be successfully authenticated and the target image information in the scene is recorded. In an embodiment of the present invention, the problem of background information distracting the model and causing misidentification or missed identification can be overcome by normalizing the target image information. By normalizing the target image information, background interference factors can be quickly removed, so that the multi-target recognition model can be more focused on feature extraction of the face area, thereby improving the accuracy of recognition. At the same time, a multi-target recognition model is constructed by combining a multi-layer spatial displacement perception mechanism and a cascaded posture regression model. The multi-target recognition model approximates the real face shape and feature point position through cluster analysis, and corrects, reduces noise and optimizes the results of the previous level, thereby improving the positioning accuracy of the feature points. It overcomes the problem that the existing method uses frame-by-frame analysis of face images and can only simply process each frame independently, is sensitive to interference information, and cannot fully utilize the correlation and continuity between the previous and next frames, resulting in the accuracy of the recognition results being limited.

[0079] The embodiment of the present invention provides a method for authenticating users of an online education platform. Figure 1 The following is a schematic diagram of the implementation process of the online education platform user authentication method, which specifically includes:

[0080] Step S10: collecting target image information in the authentication scene in real time, normalizing the target image information, and obtaining normalized target image information;

[0081] It should be noted that the online education platform in the embodiment of the present invention includes but is not limited to a traffic platform, a vertical content platform, and a tool platform. The teaching mode of online education can be a recorded class, a live class, or a hybrid teaching platform. The target image information in the acquisition and authentication scene can be collected through a communicative mobile terminal, a mobile phone, a computer, a web page, or an attendance machine. For example, by using artificial intelligence image recognition technology, personnel authentication can be performed through a fingerprint machine and a face recognition machine. The real-time acquisition and authentication scene on the web page can be used to track images captured by the camera in real time. At the same time, the data object being tracked can be a color or a human image. If the data object being tracked is a face image, the specific tracking detection target can also be set to be the eyes, nose, and mouth in the face image. When the target is detected to move or appear, the face recognition authentication instruction will be triggered.

[0082] Step S20: pre-collecting standard facial images based on an image acquisition device, constructing an image database using the standard facial images, traversing the image database, extracting standard facial images from the image database, processing the standard facial images based on different angles, brightness, and noise factors, creating a complex image set, constructing a multi-target recognition model by combining a multi-layer spatial displacement perception mechanism and a cascaded posture regression model, iteratively training the multi-target recognition model using the complex image set, and outputting a converged multi-target recognition model;

[0083] Step S30: Using the normalized target image information as input, executing a multi-target recognition model, which performs noise reduction processing on the target image information, then segments and extracts feature points, fuses the feature points based on a layer-space displacement perception mechanism, outputs a feature fusion set, and calculates the confidence level between the target image information and the standard face image.

[0084] Step S40, determining whether the confidence level of the target image information and the standard face image meets the confidence threshold;

[0085] It's important to note that the confidence threshold can be set between 0.98 and 1. This threshold comprehensively determines the success or failure of facial authentication and is a key parameter for determining whether two facial images match. After the facial images undergo a series of processing steps, including feature extraction and comparison, a similarity score or probability value is generated, which is the confidence level. If the confidence level is higher than the set threshold, the two faces are considered a match, meaning the recognition is successful; otherwise, it is considered a mismatch.

[0086] Step S50: If the confidence level of the target image information and the standard face image meets the confidence threshold, it is determined that the authentication of the target image information in the scene is successful;

[0087] If the confidence level of the target image information and the standard face image does not meet the confidence threshold, the education platform user authentication and recognition failure instruction is triggered, and the target image information is uploaded to the image database.

[0088] Step S60: input the target image information in the scene to complete the education platform user authentication and identification.

[0089] In an embodiment of the present invention, the problem of background information distracting the model and causing misidentification or missed identification can be overcome by normalizing the target image information. By normalizing the target image information, background interference factors can be quickly removed, so that the multi-target recognition model can be more focused on feature extraction of the face area, thereby improving the accuracy of recognition. At the same time, a multi-target recognition model is constructed by combining a multi-layer spatial displacement perception mechanism and a cascaded posture regression model. The multi-target recognition model approximates the real face shape and feature point position through cluster analysis, and corrects, reduces noise and optimizes the results of the previous level, thereby improving the positioning accuracy of the feature points. It overcomes the problem that the existing method uses frame-by-frame analysis of face images and can only simply process each frame independently, is sensitive to interference information, and cannot fully utilize the correlation and continuity between the previous and next frames, resulting in the accuracy of the recognition results being limited.

[0090] The embodiment of the present invention provides a method for normalizing target image information. Figure 2 The following is a schematic diagram of a process for implementing a method for normalizing target image information. The method for normalizing target image information specifically includes:

[0091] Step S101: acquiring target image information, processing the target image information using a two-dimensional empirical mode decomposition algorithm, and generating a low-discrepancy point set based on cardinality flip sampling;

[0092] Among them, the base flip sampling converts the target image information into integer The exponential expansion of is:

[0093] (1)

[0094] is the input target image information, Cardinality expansion information representing target image information;

[0095] The cardinality expansion information is sampled l times to generate a Hammersley sequence;

[0096] The Hammersley sequence is represented as:

[0097] (2)

[0098] in, Indicates the number of sampling times, Indicates the projection direction vector number;

[0099] Calculate the average value of the mean envelope in all projection directions to obtain the low-discrepancy point set;

[0100] (3)

[0101] It should be noted that in formula (3), are the maximum and minimum values ​​of the Hammersley sequence, respectively.

[0102] Step S102: obtaining a low-difference point set, and performing grayscale normalization processing on the low-difference point set by combining histogram equalization and gamma correction to obtain a grayscale normalized set;

[0103] Step S103 , loading the grayscale normalization set, combining the PP-HumanSeg algorithm, removing the background pigment in the grayscale normalization set, and realizing normalization processing of the target image information.

[0104] In this embodiment, two-dimensional empirical mode decomposition is performed on the target image information through steps S101-S103 to generate a low-difference point set, and then facial features that are not affected by illumination are extracted. These features remain relatively stable under different illumination conditions, which can reduce the interference of illumination changes on face recognition and improve the accuracy and reliability of recognition. Then, the grayscale normalization of the low-difference point set is performed in combination with histogram equalization and gamma correction, which can overcome the situation where some grayscale levels in the low-difference point set are relatively concentrated. Histogram equalization can effectively widen the gap between these grayscale levels, thereby enhancing the overall contrast of the image, which makes the details in the face image clearer. Finally, background pigment removal in the grayscale normalization set can highlight the main features of the face and reduce interference factors, thereby improving the robustness and efficiency of the multi-target recognition model.

[0105] The embodiment of the present invention provides a method for processing a standard face image based on different angles, brightness, and noise factors. Figure 3 The following is a schematic diagram of the implementation process of a method for processing standard face images based on different angles, brightness, and noise factors. The method for processing standard face images based on different angles, brightness, and noise factors specifically includes:

[0106] Step S201: obtaining a standard face image, randomly rotating the standard face image at a set rotation angle (-30° to 30°), and mirror-flipping the randomly rotated standard face image to generate a random angle set;

[0107] In this embodiment, the rotation angle (-30°-30°) is set to meet the recognition requirements of target image information obtained at different angles, while also helping to improve the generalization ability and robustness of the model.

[0108] At the same time, the standard face image is pre-collected based on the image acquisition device, which includes but is not limited to ordinary cameras, professional face acquisition cameras, 3D cameras, smart attendance machines, self-service terminals, and handheld 3D scanners. The image database constructed based on the standard face image can be a SQL Server, MySQL database, etc., and the database server is built using Oracle Database Link technology and database connection method; finally, the file server is used to provide file storage services for users.

[0109] Step S202 , loading a random angle set, adjusting the brightness of the random angle set within a preset brightness range (10-150%) based on a multi-scale Retinex algorithm, and outputting the brightness-adjusted random angle set;

[0110] Among them, the random angle set output is expressed as:

[0111] (4)

[0112] in, is the input representation of a random angle set, represents the illumination component of the random angle set data, represents a pixel value gain factor. In this embodiment, the pixel value gain factor may be 0.1-0.5;

[0113] Step S203 , obtaining a random angle set after brightness adjustment, and synchronously adding salt and pepper noise and Gaussian noise of preset proportions to the random angle set to create a complex image set.

[0114] In this embodiment, by synchronously adding preset proportions of salt and pepper noise and Gaussian noise to the random angle set, the robustness of the model to noise interference can be enhanced. At the same time, in order to further improve the performance of the model, the color distortion, geometric transformation and generative adversarial network enhancement of the image set can also be performed when creating a complex image set.

[0115] In an embodiment of the present invention, a complex image set is created based on different angles, brightness, and noise factors. This training image set, which incorporates a variety of angles, brightness, and noise factors, provides the model with comprehensive and rich learning material, helping it to more accurately capture the key features and detailed information of a face. This significantly improves the model's accuracy in recognizing faces, while overcoming the issues of image quality, angle, and brightness that may arise when multi-target recognition models are run on different hardware devices. It also meets the needs of face recognition in strong outdoor light, dim indoor light, and during motion.

[0116] The embodiment of the present invention provides a method for iteratively training a multi-target recognition model using a complex image set. Figure 4The following is a schematic diagram of the implementation process of the iterative training method for a multi-target recognition model using a complex image set. The iterative training method for a multi-target recognition model using a complex image set specifically includes:

[0117] Step S301: Using a cascaded pose regression model as the initial model of a multi-target recognition model. The initial model consists of a shape estimation layer, a feature extraction layer, and a cascade regressor. The shape estimation layer is connected to the feature extraction layer, which is then connected to the cascade regressor. The feature extraction layer includes three groups of convolutional layers, one group of pooling layers, one group of upsampling layers, and one group of feature fusion layers.

[0118] It should be noted that the convolution kernel size of the convolution layer in the feature extraction layer is 3*3, the number of channels of the convolution layer is 128, the number of channels of the pooling layer is 256, and the number of channels of the upsampler is 512, the parameter = 5, and the number of channels of the feature fusion layer is 512.

[0119] Step S302: introduce a confusion matrix analysis module into the shape estimation layer, freeze the pooling layer in the feature extraction layer, and introduce a layer space displacement perception mechanism to replace the pooling layer;

[0120] Step S303: Load the complex image set, perform hybrid enhancement on the complex image set based on Mosaic and Mixup hybrid enhancement technology, obtain 40 groups of facial posture images with an interval of 2, and divide the hybrid enhanced complex image set into a training set and a test set;

[0121] In this embodiment, the ratio of the training set to the test set may be 4:1 or 3:1, and the training set and the test set may be randomly allocated using a K-fold cross-validation method.

[0122] Step S304: Obtain a training set and set initial model parameters, where the initial model has a set number of feature points, a set number of clustering iterations of 20, a set number of random forest decision trees of 250, a feature point radius of 2, and a pixel feature of 400-500.

[0123] Step S305: using the training set as input, training the initial model based on random forest local encoding and K-means cluster analysis until the initial model converges, thereby obtaining at least one set of initial models in a multi-view recognition mode;

[0124] Step S306: obtaining a test set, taking the test set as input, and performing posture evaluation and confidence evaluation on the test set using at least one set of initial models to obtain a posture deflection angle and image confidence;

[0125] Step S307: Preset a standard threshold and calculate the model recognition accuracy of the initial model by combining the posture deflection angle and the image confidence;

[0126] In this embodiment, the measurement standard threshold can be 0.05-0.055. If the model recognition accuracy of the initial model calculated by the posture deflection angle and the image confidence meets the measurement standard threshold, the corresponding initial model is set as the multi-target recognition model. If there are multiple initial models, the initial model with the highest model recognition accuracy is selected as the multi-target recognition model.

[0127] Step S308 , selecting the initial model whose model recognition accuracy meets the measurement standard threshold range as the final multi-target recognition model, and outputting the converged multi-target recognition model.

[0128] In this embodiment, the model recognition accuracy is calculated using the following formula:

[0129] (5)

[0130] in, represents the model recognition accuracy, is the number of feature points. In this embodiment, the feature points can be 5-10, including but not limited to the mouth, eyes, nose, ears, and forehead feature points in the face. represents the attitude deflection angle, Represents the image confidence.

[0131] In an embodiment of the present invention, a multi-target recognition model is constructed by combining a multi-layer spatial displacement perception mechanism and a cascaded posture regression model, and a confusion matrix analysis module is introduced into the multi-target recognition model, so that changes in the face under different postures can be effectively processed. For example, when the face is in different postures such as sideways, looking up, and lowering the head, the multi-target recognition model can gradually correct the influence of the posture on the facial features, so that the model can more accurately extract the key features of the face, further improve the accuracy of recognition, and thus make the entire multi-target recognition model more robust to complex environmental factors. Whether it is a face image from the front or side, or a face image with expression changes or partial occlusion, the model can perform recognition well, which improves the application range and reliability of face recognition.

[0132] The embodiment of the present invention provides a method for fusing feature points based on a layer space displacement perception mechanism and outputting a feature fusion set. Figure 5 The figure shows a schematic diagram of the implementation process of a method for fusing feature points based on a layer-space displacement perception mechanism and outputting a feature fusion set. The method for fusing feature points based on a layer-space displacement perception mechanism and outputting a feature fusion set specifically includes:

[0133] Step S401: using the normalized target image information as input, executing the multi-target recognition model, and the shape estimation layer preliminarily estimates and determines the target image information based on the confusion matrix algorithm;

[0134] Step S402, determining whether the target image information is a living body image;

[0135] It's important to note that the confusion matrix algorithm uses SIFT (Scale-Invariant Feature Transform) and HOG (Histogram of Oriented Gradients) methods to initially estimate and determine target image information. These features can include color, texture, shape, and edges. A confusion matrix is ​​then constructed to predict whether each image is alive or not. Based on the true class labels and the predicted results, the rows of the confusion matrix represent the actual class (live / not alive), while the columns represent the predicted class (live / not alive).

[0136] Step S403: If the target image information is a living body image, the target image information is pushed to the feature extraction layer;

[0137] Step S404: If the target image information is not a living body image, the face recognition authentication task is terminated.

[0138] In step S405, the feature extraction layer obtains target image information, introduces an equivalent weighted filter to perform low-pass filtering and noise reduction processing on the target image information, filters out interference components of the target image information, and outputs a noise reduction information set.

[0139] It should be noted that the noise reduction information set is expressed as:

[0140] (6)

[0141] (7)

[0142] (8)

[0143] in, is the output representation of the denoised information set, represents the equivalent weighted filter, Indicates the target image information input information, Indicates the number of times of filtering and removing impurities. In this embodiment, the number of times of filtering and removing impurities can be 2-15. is the amplitude response of the filter, represents the filter signal-to-noise ratio, represents the overlapping correlation coefficient, is the overlap ratio constant. In this embodiment, the overlap ratio constant may be 0.1-0.9.

[0144] Step S406: Obtain a denoising information set. The feature extraction layer performs denoising on the target image information and then segments and extracts feature points. A domain classifier is used to extract the bounding box and facial feature information of the denoising information set, and ROI pooling is performed with a feature map of a scale corresponding to the standard face image.

[0145] Step S407: performing fusion processing on the feature points based on the layer space displacement perception mechanism and outputting a feature fusion set;

[0146] Step S408, calculating the confidence level of the target image information and the standard face image;

[0147] Step S409, determining whether the confidence level of the target image information and the standard face image meets the confidence threshold;

[0148] The confidence of the target image information and the standard face image is calculated using the following formula:

[0149] (9)

[0150] in, Respectively represent the target image information and the fusion value of the face standard image feature, is the label confidence coefficient of the target image information. In this embodiment, the label confidence coefficient can be determined by domain experts based on their experience in label confidence assessment. Experts can assign a confidence value based on factors such as image quality and feature visibility. The label confidence coefficient is in the range of 0.1-1.

[0151] The embodiment of the present invention also provides an online education platform user authentication system, Figure 6 The following is a schematic diagram showing the structure of an online education platform user authentication system, which specifically includes:

[0152] The data normalization module 100 is used to collect target image information in the authentication scene in real time, normalize the target image information, and obtain normalized target image information;

[0153] The recognition modeling module 200 pre-collects standard facial images based on an image acquisition device, constructs an image database using the standard facial images, traverses the image database, extracts standard facial images from the image database, processes the standard facial images based on different angles, brightness, and noise factors, creates a complex image set, constructs a multi-target recognition model by combining a multi-layer spatial displacement perception mechanism and a cascaded posture regression model, iteratively trains the multi-target recognition model using the complex image set, and outputs a converged multi-target recognition model;

[0154] The face recognition module 300 is used to execute a multi-target recognition model, which performs noise reduction processing on the target image information, then segments and extracts feature points, fuses the feature points based on a layer-space displacement perception mechanism, outputs a fusion feature set, and calculates the confidence level between the target image information and the standard face image;

[0155] The authentication judgment module 400 is used to judge whether the confidence of the target image information and the standard face image meets the confidence threshold. If the confidence of the target image information and the standard face image meets the confidence threshold, it is judged that the authentication of the target image information in the scene is successful, and the target image information in the scene is entered to complete the education platform user authentication and identification.

[0156] In an embodiment of the present invention, the user authentication system of the online education platform is composed of a data normalization module 100, a recognition modeling module 200, a face recognition module 300, and an authentication judgment module 400. The data normalization module 100, the recognition modeling module 200, the face recognition module 300, and the authentication judgment module 400 can be connected and data interaction can be achieved by static linking, dynamic linking at loading time, and dynamic linking at runtime.

[0157] In this embodiment, the data normalization module 100 specifically includes:

[0158] The modal decomposition unit 110 is used to obtain target image information, process the target image information using a two-dimensional empirical mode decomposition algorithm, and generate a low-discrepancy point set based on cardinality reversal sampling;

[0159] Grayscale processing unit 120, used to obtain a low-difference point set, and perform grayscale normalization processing on the low-difference point set in combination with histogram equalization and gamma correction to obtain a grayscale normalized set;

[0160] The background selection unit 130 is used to load the grayscale normalization set and remove the background pigment in the grayscale normalization set in combination with the PP-HumanSeg algorithm to achieve normalization processing of the target image information.

[0161] It should be noted that the online education platform user authentication system provided by the embodiment of the present invention corresponds to the above-mentioned online education platform user authentication method. The explanations, examples, beneficial effects, etc. of the relevant contents can refer to the corresponding contents in the online education platform user authentication system and will not be repeated here.

[0162] On the other hand, an embodiment of the present invention further provides a computer-readable storage medium storing computer program instructions, which can be executed by a processor. When the computer program instructions are executed, the method of any of the above embodiments is implemented.

[0163] The computer-readable storage medium (e.g., memory) herein may be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. By way of example and not limitation, non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM), which may act as external cache memory. By way of example and not limitation, RAM may be available in a variety of forms, such as synchronous RAM (DRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), and direct Rambus RAM (DRRAM). The memory devices of the disclosed aspects are intended to include, but are not limited to, these and other suitable types of memory.

[0164] At the same time, an embodiment of the present invention further provides a computer device, which includes a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the method of any one of the above embodiments is implemented.

[0165] As a non-volatile computer-readable storage medium, the memory can be used to store non-volatile software programs, non-volatile computer executable programs and modules, such as the program instructions / modules corresponding to the online education platform user authentication method in the embodiment of the present application. The memory may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function; the data storage area may store data created by the use of the online education platform user authentication method, etc. In addition, the memory may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory may optionally include a memory remotely located relative to the processor 410, and these remote memories may be connected to the local module via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof.

[0166] In summary, the present invention provides a user authentication method and system for an online education platform. In an embodiment of the present invention, the problem of background information distracting the model's attention, resulting in misidentification or missed identification, can be overcome by normalizing the target image information. By normalizing the target image information, background interference factors can be quickly removed, so that the multi-target recognition model can be more focused on feature extraction of the face area, thereby improving recognition accuracy. At the same time, a multi-target recognition model is constructed by combining a multi-layer spatial displacement perception mechanism and a cascaded posture regression model. The multi-target recognition model approximates the real face shape and feature point position through clustering analysis, and corrects, de-noises and optimizes the results of the previous level, thereby improving the positioning accuracy of the feature points. It overcomes the problem that the existing method uses frame-by-frame analysis of face images and can only simply process each frame independently, is sensitive to interference information, and cannot fully utilize the correlation and continuity between the previous and subsequent frames, resulting in the accuracy of the recognition results being limited.

[0167] It should be noted that for the aforementioned embodiments, for simplicity of description, they are all expressed as a series of action combinations. However, those skilled in the art should be aware that the present invention is not limited by the order of the actions described, because according to the present invention, certain steps may be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present invention.

[0168] In the several embodiments provided in this application, it should be understood that the disclosed devices can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the above-mentioned units is only a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or communication connection shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be in the form of telecommunications or other forms.

[0169] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the scope of protection of the invention. Obviously, the embodiments described are only some embodiments of the present invention, rather than all embodiments. Based on these embodiments, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in this field can still combine, add, delete or make other adjustments to the features in the various embodiments of the present invention according to the circumstances without conflict, without making creative work, so as to obtain different other technical solutions that do not deviate from the concept of the present invention in essence, and these technical solutions also fall within the scope of protection of the present invention.

Claims

1. Online education platform user authentication method, characterized in that: The online education platform user authentication method includes: Real-time collection of target image information in the authentication scene, normalization of the target image information, and obtaining normalized target image information; Based on the image acquisition device, standard facial images are pre-collected, an image database is constructed using the standard facial images, the image database is traversed, standard facial images are extracted from the image database, and the standard facial images are processed based on different angles, brightness, and noise factors to create a complex image set. A multi-target recognition model is constructed by combining a multi-layer spatial displacement perception mechanism and a cascaded posture regression model. The multi-target recognition model is iteratively trained using the complex image set to output a converged multi-target recognition model. Taking the normalized target image information as input, the multi-target recognition model is executed. The multi-target recognition model performs noise reduction processing on the target image information, then segments and extracts feature points. The feature points are fused based on the layer space displacement perception mechanism, and a feature fusion set is output. The confidence level of the target image information and the standard face image is calculated, and it is determined whether the confidence level of the target image information and the standard face image meets the confidence threshold. If the confidence level of the target image information and the standard face image meets the confidence threshold, the target image information in the scene is judged to be successfully authenticated, and the target image information in the scene is entered to complete the education platform user authentication and identification; The method for normalizing target image information specifically includes: Obtain target image information, process the target image information using a two-dimensional empirical mode decomposition algorithm, and generate a low-discrepancy point set based on cardinality flip sampling; Among them, the base flip sampling converts the target image information into integer The exponential expansion of is: （1） is the input target image information, Cardinality expansion information representing target image information; The cardinality expansion information is sampled l times to generate a Hammersley sequence; The Hammersley sequence is represented as: （2） in, Indicates the number of sampling times, Indicates the projection direction vector number; Calculate the average value of the mean envelope in all projection directions to obtain the low-discrepancy point set; （3） Obtain a low-difference point set, and normalize the grayscale of the low-difference point set by combining histogram equalization and gamma correction to obtain a grayscale normalized set; Load the grayscale normalization set and combine it with the PP-HumanSeg algorithm to remove the background pigment in the grayscale normalization set to achieve normalization processing of the target image information.

2. The online education platform user authentication method according to claim 1, wherein: The method for processing a standard face image based on different angles, brightness, and noise factors specifically includes: Obtain a standard face image, randomly rotate the standard face image at a set rotation angle (-30°-30°), and mirror-flip the randomly rotated standard face image to generate a random angle set; Load a random angle set, adjust the brightness of the random angle set within a preset brightness range (10-150%) based on the multi-scale Retinex algorithm, and output the brightness-adjusted random angle set; Among them, the random angle set output is expressed as: （4） in, is the input representation of a random angle set, represents the illumination component of the random angle set data, Represents the pixel value gain factor; Get a random angle set after brightness adjustment, and simultaneously add preset proportions of salt and pepper noise and Gaussian noise to the random angle set to create a complex image set.

3. The online education platform user authentication method according to claim 2, characterized in that: The method for iteratively training a multi-target recognition model using a complex image set specifically includes: The cascaded pose regression model is used as the initial model of the multi-target recognition model. The initial model consists of a shape estimation layer, a feature extraction layer, and a cascade regressor. The shape estimation layer is connected to the feature extraction layer, which is then connected to the cascade regressor. The feature extraction layer includes three groups of convolutional layers, one group of pooling layers, one group of upsampling layers, and one group of feature fusion layers. A confusion matrix analysis module is introduced in the shape estimation layer, and the pooling layer in the feature extraction layer is frozen, and a layer-space displacement perception mechanism is introduced to replace the pooling layer; Load a complex image set and perform hybrid enhancement on it based on Mosaic and Mixup hybrid enhancement techniques to obtain 40 groups of facial posture images with an interval of 2. Divide the hybrid enhanced complex image set into a training set and a test set. Obtain the training set and set the initial model parameters. The initial model has the following parameters: the number of feature points, the number of clustering iterations is 20, the number of random forest decision trees is 250, the feature point radius is 2, and the pixel feature is 400-500. Taking the training set as input, the initial model is trained based on random forest local encoding and K-means cluster analysis until the initial model converges, and at least one set of initial models in the multi-view recognition mode is obtained; A test set is obtained, and the test set is used as input. At least one set of initial models is used to perform posture evaluation and confidence evaluation on the test set to obtain a posture deflection angle and an image confidence. The measurement standard threshold is preset, and the model recognition accuracy of the initial model is calculated by combining the posture deflection angle and image confidence. The initial model whose model recognition accuracy meets the measurement standard threshold interval is selected as the final multi-target recognition model, and the converged multi-target recognition model is output.

4. The online education platform user authentication method according to claim 3, wherein: The model recognition accuracy is calculated by the following formula: （5） in, represents the model recognition accuracy, is the number of feature points, represents the attitude deflection angle, Represents the image confidence.

5. The online education platform user authentication method according to claim 4, characterized in that: The method of fusing feature points based on the layer space displacement perception mechanism and outputting a feature fusion set specifically includes: Taking the normalized target image information as input, the multi-target recognition model is executed. The shape estimation layer makes a preliminary estimate and judgment of the target image information based on the confusion matrix algorithm to determine whether the target image information is a living image; If the target image information is a living body image, the target image information is pushed to the feature extraction layer; The feature extraction layer obtains the target image information, introduces an equivalent weighted filter to perform low-pass filtering and noise reduction on the target image information, filters out the interference components of the target image information, and outputs the noise reduction information set.

6. The online education platform user authentication method according to claim 5, characterized in that: The method of fusing feature points based on the layer space displacement perception mechanism and outputting a feature fusion set specifically includes: Obtain the denoising information set, segment and extract feature points after denoising the target image information in the feature extraction layer, use the domain classifier to extract the bounding box and facial feature information of the denoising information set, and perform ROI pooling with the feature map of the corresponding scale of the standard face image; Based on the layer space displacement perception mechanism, feature points are fused and processed to output a feature fusion set; Calculate the confidence of target image information and standard face image; Determine whether the confidence level of the target image information and the standard face image meets the confidence threshold; The confidence of the target image information and the standard face image is calculated using the following formula: （9） in, Respectively represent the target image information and the fusion value of the face standard image feature, is the label confidence coefficient of the target image information.

7. The online education platform user authentication method according to claim 5, characterized in that: The denoising information set is expressed as: （6） （7） （8） in, is the output representation of the denoised information set, represents the equivalent weighted filter, Indicates the target image information input information, Indicates the number of times of filtering and removing impurities, is the amplitude response of the filter, represents the filter signal-to-noise ratio, represents the overlapping correlation coefficient, is the overlap proportionality constant.

8. An online education platform user authentication system, configured to implement the online education platform user authentication method according to any one of claims 1 to 7, characterized in that: The online education platform user authentication system specifically includes: The data normalization module is used to collect target image information in the authentication scene in real time, normalize the target image information, and obtain normalized target image information; The recognition modeling module pre-collects standard facial images based on image acquisition equipment, builds an image database based on these standard facial images, traverses the image database, extracts standard facial images from the image database, processes the standard facial images based on different angles, brightness, and noise factors, creates a complex image set, and builds a multi-target recognition model by combining a multi-layer spatial displacement perception mechanism and a cascaded posture regression model. The multi-target recognition model is iteratively trained using the complex image set to output a converged multi-target recognition model. The face recognition module is used to execute the multi-target recognition model. The multi-target recognition model performs noise reduction on the target image information, then segments and extracts feature points. It then fuses the feature points based on the layer-space displacement perception mechanism, outputs a fusion feature set, and calculates the confidence level between the target image information and the standard face image. The authentication judgment module is used to judge whether the confidence of the target image information and the standard face image meets the confidence threshold. If the confidence of the target image information and the standard face image meets the confidence threshold, it is judged that the authentication of the target image information in the scene is successful, and the target image information in the scene is entered to complete the education platform user authentication and identification.

9. The online education platform user authentication system according to claim 8, characterized in that: The data normalization module specifically includes: A modal decomposition unit is used to obtain target image information, process the target image information using a two-dimensional empirical mode decomposition algorithm, and generate a low-discrepancy point set based on cardinality flip sampling; A grayscale processing unit is used to obtain a low-difference point set, and normalize the grayscale of the low-difference point set by combining histogram equalization and gamma correction to obtain a grayscale normalized set; The background selection unit is used to load the grayscale normalization set and, combined with the PP-HumanSeg algorithm, remove the background pigment in the grayscale normalization set to achieve normalization processing of the target image information.

Images