Intelligent education management system and method based on AI and cloud computing

Abstract

The invention provides an intelligent education management system and method based on AI and cloud computing, and relates to the field of education management.The method comprises the steps that firstly, classroom monitoring images are obtained on the end side, student behavior detection such as hand raising, standing and walking is conducted on the classroom monitoring images, and when the student behaviors are detected, the student behaviors are uploaded to a cloud computing platform; otherwise, receiving a next frame of classroom monitoring image, and carrying out student behavior region-of-interest detection and multi-scale image extraction processing on the classroom monitoring image on a cloud computing platform by adopting an image analysis technology based on AI and cloud computing to obtain a student behavior local image coding feature set; and a classroom behavior recognition result is intelligently obtained based on the student behavior feature aggregation representation of the student behavior local image coding feature set. Therefore, the hand raising behavior of the student can be identified more accurately, the mode and the dynamic change trend of the classroom participation degree can be comprehensively known, and the intelligent degree of teaching management can be effectively improved.

Description

Technical Field

[0001] The present application relates to the field of education management, and more specifically, to an intelligent education management system and method based on AI and cloud computing. Background Art

[0002] Hand-raising behavior in class is an important indicator of student participation in classroom activities, which can reflect students' participation and initiative. By identifying hand-raising behavior in class, teachers can better understand students' participation, adjust teaching strategies in time, improve teaching quality, and provide necessary help and support for individual students.

[0003] In traditional classrooms, the recognition of hand-raising behavior relies on manual recognition by teachers, which is easily affected by factors such as teacher distraction and limited vision, and there are cases of missed judgments and misjudgments. Especially in large classes, it is difficult to fully and accurately capture the hand-raising behavior of each student. In addition, manual recognition cannot systematically record the data of hand-raising behavior, making it difficult to grasp the overall classroom participation mode and change trend of students, and unable to provide strong data support for teaching improvement. Summary of the invention

[0004] In order to solve the above technical problems, this application is proposed. The embodiments of this application provide a smart education management system and method based on AI and cloud computing.

[0005] According to one aspect of the present application, a smart education management system based on AI and cloud computing is provided, which includes: a data acquisition and transmission module, used to obtain classroom monitoring images, and perform student behavior judgment on the classroom monitoring images, if there is student behavior in the classroom monitoring images, the classroom monitoring images are uploaded to the cloud computing platform; an image receiving module, used to receive the next frame of the classroom monitoring image if there is no student behavior in the classroom monitoring image; a classroom behavior recognition module, used to perform student behavior analysis on the classroom monitoring images on the cloud computing platform to obtain classroom behavior recognition results, wherein the classroom behavior recognition module includes: a behavior aggregation unit, used to perform semantic aggregation of behavior interest regions on the classroom monitoring images to obtain semantic-level aggregation coding features of student behavior images; an identification unit, used to obtain the classroom behavior recognition results based on the semantic-level aggregation coding features of the student behavior images.

[0006] According to another aspect of the present application, a smart education management method based on AI and cloud computing is provided, which includes: obtaining a classroom monitoring image, and judging student behavior on the classroom monitoring image, if there is student behavior in the classroom monitoring image, uploading the classroom monitoring image to a cloud computing platform; if there is no student behavior in the classroom monitoring image, receiving the next frame of the classroom monitoring image; on the cloud computing platform, performing student behavior analysis on the classroom monitoring image to obtain a classroom behavior recognition result, including: performing semantic aggregation of behavior interest regions on the classroom monitoring image to obtain semantic-level aggregation coding features of the student behavior image; obtaining the classroom behavior recognition result based on the semantic-level aggregation coding features of the student behavior image.

[0007] Compared with the prior art, the smart education management system and method based on AI and cloud computing provided by this application first obtains classroom monitoring images on the end side and detects student behavior. When student behavior is detected, it is uploaded to the cloud computing platform. Otherwise, the next frame of classroom monitoring images is received. The image analysis technology based on AI and cloud computing is used on the cloud computing platform to detect the student behavior area of ​​interest and multi-scale image extraction processing to obtain the local image coding feature set of student behavior, and the classroom behavior recognition result is intelligently obtained based on the student behavior feature aggregation representation of the local image coding feature set of student behavior. In this way, the student's hand-raising behavior can be more accurately identified, and the pattern and dynamic change trend of classroom participation can be fully understood, which can effectively improve the intelligence of teaching management. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0009] In the attached picture: Figure 1 This is a system block diagram of an intelligent education management system based on AI and cloud computing according to an embodiment of the present application.

[0010] Figure 2 It is a block diagram of the data acquisition and transmission module in the smart education management system based on AI and cloud computing according to an embodiment of the present application.

[0011] Figure 3 This is a block diagram of a classroom behavior recognition module in an AI and cloud computing-based smart education management system according to an embodiment of the present application.

[0012] Figure 4It is a block diagram of a behavior aggregation unit in a smart education management system based on AI and cloud computing according to an embodiment of the present application.

[0013] Figure 5 This is a flowchart of a smart education management method based on AI and cloud computing according to an embodiment of the present application. DETAILED DESCRIPTION

[0014] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0015] The hand-raising behavior of students in class is one of the important indicators of their participation in classroom activities. It can reflect the degree of participation and initiative of students. By identifying the hand-raising behavior in class, teachers can have a more comprehensive understanding of students' participation, so as to flexibly adjust teaching strategies, optimize classroom teaching effects, and provide targeted help and support to students in need.

[0016] In traditional classrooms, hand-raising behavior usually relies on manual observation by teachers to identify it. However, this method is easily affected by factors such as teacher distraction and limited vision, which may lead to omissions or misjudgments. Especially in large-class teaching environments, it is difficult for teachers to fully and accurately capture each student's hand-raising behavior. In addition, manual recognition methods cannot systematically record data on hand-raising behavior, making it difficult to analyze students' classroom participation patterns and their changing trends as a whole, and therefore cannot provide reliable data support for teaching improvements.

[0017] Based on this, the technical concept of the present application is to obtain classroom monitoring images collected by the camera on the end side, and when student behavior is detected, the image is transmitted to the cloud computing platform. When there is no student behavior, the next frame of classroom monitoring image is received. In the cloud computing platform, image analysis technology based on AI and cloud computing is used to detect student behavior interest areas and multi-scale image extraction processing on classroom monitoring images to obtain a set of local image coding features of student behavior, and based on the student behavior feature aggregation representation of the local image coding feature set of student behavior, the classroom behavior recognition result is intelligently obtained. In this way, by real-time analysis of classroom monitoring image data, not only can the accuracy of student hand-raising behavior recognition be improved, but also the overall pattern of classroom participation and its dynamic change trend can be systematically grasped, thereby improving the intelligence of teaching management and providing reliable data support for the optimization of teaching strategies.

[0018] Figure 1 : is a system block diagram of a smart education management system based on AI and cloud computing according to an embodiment of the present application. Figure 1As shown, in the smart education management system 100 based on AI and cloud computing, it includes: a data acquisition and transmission module 110, which is used to obtain classroom monitoring images and judge student behavior on the classroom monitoring images. If there is student behavior in the classroom monitoring image, the classroom monitoring image is uploaded to the cloud computing platform; an image receiving module 120, which is used to receive the next frame of the classroom monitoring image if there is no student behavior in the classroom monitoring image; a classroom behavior recognition module 130, which is used to analyze the student behavior of the classroom monitoring image on the cloud computing platform to obtain a classroom behavior recognition result.

[0019] In the embodiment of the present application, the data acquisition and transmission module 110 is used to obtain classroom monitoring images and to judge student behavior in the classroom monitoring images. If student behavior exists in the classroom monitoring images, the classroom monitoring images are uploaded to the cloud computing platform. Specifically, Figure 2 FIG. 1 is a block diagram of a data acquisition and transmission module in a smart education management system based on AI and cloud computing according to an embodiment of the present application. Figure 2 As shown, the data acquisition and transmission module 110 includes: a classroom monitoring image collection unit 111, which is used to obtain classroom monitoring images collected by a camera on the terminal side; a classroom monitoring image detection and transmission unit 112, which is used to detect whether there is student behavior in the classroom monitoring image on the terminal side, and if there is student behavior in the classroom monitoring image, upload the classroom monitoring image to the cloud computing platform.

[0020] In the embodiment of the present application, the classroom monitoring image collection unit 111 is used to obtain classroom monitoring images collected by the camera on the terminal side. It should be understood that the classroom is a dynamically changing scene, and student behavior may occur at any time. By directly acquiring monitoring images on the terminal side, student behavior information can be collected at the fastest speed. For example, the behavior of a student raising his hand occurs instantly. If it is not captured in time on the terminal side, the delay between the occurrence of the behavior and the data collection will cause the information to be acquired in time, affecting the subsequent real-time analysis and teaching intervention. The classroom monitoring images collected on the terminal side contain students' facial expression information, body movement information, classroom environment information, and interaction information between students, such as eye contact between students, and actions of discussing with each other. This information is an important basis for determining the label of the hand-raising behavior. They are interrelated and complementary to each other, and together provide support for the identification of the hand-raising behavior. Specifically, student body movement information is the core clue for judging the hand-raising behavior. When the image shows that the student's arm is raised upward, and the palm is open or clenched, etc., which meets the characteristics of the hand-raising action, this is a key sign for identifying the hand-raising behavior. Facial expressions and eye direction can also assist in judgment. If a student raises his hand with a positive expression and looks at the teacher, it can be confirmed that his intention to raise his hand is to actively participate in classroom interaction. Although classroom environment information cannot directly indicate hand-raising behavior, it will affect the judgment of hand-raising actions. For example, if the light in the image is dim, it may make the hand-raising action difficult to clearly identify, and it is necessary to combine other information for judgment. Information on the interaction scene with other students can help eliminate misjudgment. If there is no abnormal reaction from the surrounding students, and a student raises his arm, it is more likely to be a hand-raising behavior; if the surrounding students are discussing, the student's arm raising may be an action to participate in the discussion rather than raising his hand. Through the comprehensive analysis and processing of this information in the classroom monitoring image, the type label of the student's hand-raising behavior can be accurately obtained, and the effective identification of students' hand-raising behavior in class can be achieved.

[0021] In the embodiment of the present application, the classroom monitoring image detection and transmission unit 112 is used to detect whether there is student behavior in the classroom monitoring image on the terminal side, and if there is student behavior in the classroom monitoring image, the classroom monitoring image is uploaded to the cloud computing platform. It should be understood that classroom monitoring continues, and a large amount of image data will be generated. If all are uploaded, it will not only occupy too much network bandwidth, but also may cause transmission delay or even interruption due to excessive data volume. Detecting whether there is student behavior on the terminal side and only uploading images containing student behavior can greatly reduce unnecessary data transmission. For example, when students are in a quiet listening state and have no obvious behavioral movements, the corresponding images do not need to be uploaded. Only images when behaviors such as raising hands and standing up appear will be transmitted, which effectively reduces the network transmission pressure and ensures the efficiency of data transmission. Moreover, after preliminary screening on the terminal side, the cloud computing platform only needs to process images containing student behavior, and can concentrate resources for subsequent complex processing, which can speed up the processing speed of the entire system and meet the real-time requirements of classroom teaching. In particular, the cloud computing platform has powerful computing resources and high-performance computing equipment, which can quickly process complex image data. For images containing student behavior, the cloud computing platform can efficiently perform data analysis tasks. Compared with end-side devices, cloud computing platforms can complete the processing of large amounts of data in a short period of time, ensuring the timeliness and accuracy of behavior recognition. In addition, cloud computing platforms provide convenient data storage and management functions. Images containing student behaviors uploaded to the platform can be stored securely, facilitating subsequent data query, analysis, and backtracking. At the same time, cloud computing platforms also support data classification, labeling, and integration, which helps to establish a complete student classroom behavior database and provide a data basis for long-term teaching research and teaching quality evaluation.

[0022] The following is a detailed description of a specific implementation process of "detecting whether there is student behavior in the classroom monitoring image on the terminal side, and uploading the classroom monitoring image to the cloud computing platform if there is student behavior in the classroom monitoring image": The collected classroom monitoring image must first be preprocessed before entering the student behavior detection link. This process mainly includes image resolution adjustment and noise reduction processing. The image resolution is adjusted to make the image meet the input specifications of the subsequent student behavior detection model to ensure that the model can effectively identify the features in the image; noise reduction processing is to remove the noise introduced by the image during the acquisition and transmission process. These noises may interfere with the accuracy of behavior detection. Through professional noise reduction algorithms, the image quality can be improved and the behavioral features in the image can be made clearer and more discernible.

[0023] The pre-processed images are input into the student behavior detection model pre-deployed on the end-side device. The model is trained based on a deep learning algorithm and has powerful pattern recognition capabilities. It can identify a variety of student behaviors, such as raising hands, standing, walking, and other common classroom behaviors. After receiving the image, the model will conduct an in-depth analysis of each area in the image, and by comparing the preset behavioral feature patterns, determine whether there is a part in the image that meets the definition of student behavior. In this process, the model will perform complex calculations and processing on the pixel information in the image to extract key feature information for behavior judgment.

[0024] After the model outputs the detection results, the end-side device will make a judgment based on the preset rules. If the model detects that the features of a certain area in the image are highly matched with the preset student behavior features, and the confidence of this matching result reaches the preset threshold (for example, 0.8), it is determined that there is student behavior in the image; conversely, if the model does not detect the qualified behavior features, or the confidence of the detection result is lower than the threshold, the end-side device will consider that there is no student behavior in the image. This confidence-based judgment method can balance the accuracy and false positive rate of detection to a certain extent, ensuring that only behavioral detection results with higher credibility will be recognized.

[0025] When the end-side device determines that there is student behavior in the image, the image upload process will be started immediately. First, the image will be converted into a format suitable for network transmission, such as the commonly used JPEG or PNG format. This format can effectively compress the file size while ensuring image quality, reducing the bandwidth and time required for transmission. Subsequently, the end-side device establishes a reliable communication link with the cloud computing platform through a network connection (such as a stable Wi-Fi or high-speed Ethernet). During the transmission process, the end-side device will accurately send the image data to the storage location or processing interface specified by the cloud computing platform in accordance with the interface protocol and data transmission method specified by the cloud computing platform, ensuring that the data can be smoothly received and processed by the cloud computing platform.

[0026] For those images that are judged to not contain student behavior, the end-side device will not perform meaningless transmission and storage, but will directly discard them or only store them briefly before discarding them. This processing method can greatly save network bandwidth resources and storage space of end-side devices, avoid processing a large amount of invalid data, improve the overall operating efficiency of the system, and enable the end-side device to focus more on processing and uploading images with student behavior.

[0027] In an embodiment of the present application, the image receiving module 120 is used to receive the next frame of the classroom monitoring image if there is no student behavior in the classroom monitoring image. It should be understood that when there is no student behavior in the monitoring image, subsequent complex processing (such as uploading to a cloud computing platform, performing behavior analysis, etc.) is meaningless and a waste of resources. By directly skipping images without student behavior and concentrating limited resources on images with student behavior, optimal resource allocation can be achieved. And the operating efficiency of the system is affected by the amount of data processing. When an image without student behavior is detected, the next frame is directly received, which can shorten the processing cycle of a single image and speed up the overall data processing speed. In a real-time monitoring scenario, this allows the system to capture student behavior faster, improve the real-time nature of behavior detection, and enable teachers to obtain student behavior information in a timely manner. For example, after a student raises his hand, the teacher can more quickly discover and respond, thereby improving the classroom interaction effect.

[0028] In the embodiment of the present application, the classroom behavior recognition module 130 is used to perform student behavior analysis on the classroom monitoring image on the cloud computing platform to obtain classroom behavior recognition results. Specifically, Figure 3 FIG. 1 is a block diagram of a classroom behavior recognition module in a smart education management system based on AI and cloud computing according to an embodiment of the present application. Figure 3 As shown, the classroom behavior recognition module 130 includes: a behavior aggregation unit 131, which is used to perform semantic aggregation of behavior interest regions on the classroom monitoring image to obtain semantic-level aggregation coding features of student behavior images; and an identification unit 132, which is used to obtain the classroom behavior recognition result based on the semantic-level aggregation coding features of the student behavior images.

[0029] In the embodiment of the present application, the behavior aggregation unit 131 is used to perform semantic aggregation of the behavior interest region of the classroom monitoring image to obtain semantic-level aggregation coding features of the student behavior image. Specifically, Figure 4 FIG. 1 is a block diagram of a behavior aggregation unit in a smart education management system based on AI and cloud computing according to an embodiment of the present application. Figure 4 As shown, the behavior aggregation unit 131 includes: a behavior region of interest detection subunit 1311, which is used to input the classroom monitoring image into a student behavior region of interest detector based on the Yolo model to obtain a student behavior ROI area image; a behavior local image coding feature extraction subunit 1312, which is used to perform student behavior multi-scale image feature extraction on the student behavior ROI area image to obtain a set of student behavior local image coding features; a behavior image semantic level aggregation coding feature generation subunit 1313, which is used to perform student behavior feature aggregation coding with a concentrated information core on the set of student behavior local image coding features to obtain the student behavior image semantic level aggregation coding features.

[0030] In an embodiment of the present application, the behavior region of interest detection subunit 1311 is used to input the classroom monitoring image into a student behavior region of interest detector based on the Yolo model to obtain a student behavior ROI area image. It should be understood that the classroom monitoring image contains a large amount of background information, such as desks, chairs, walls, etc. in the classroom. If the subsequent behavior analysis is performed directly on the entire image, it will involve processing a large amount of data unrelated to the student behavior, increasing the computing cost and processing time. By detecting the region of interest (ROI), it is possible to focus on the part of the image containing the student behavior, greatly reducing the amount of data processing and improving the processing efficiency. That is, when identifying the hand-raising behavior, only the area where the student is located is focused on to avoid unnecessary calculations on other irrelevant parts of the classroom. Specifically, in the present application, it is necessary to input the classroom monitoring image into a student behavior region of interest detector based on the Yolo model to obtain a student behavior ROI area image. Those of ordinary skill in the art should understand that the Yolo model simplifies the target detection task into a regression problem, directly predicting the category and location information of the target from the input image. In the detection of student behavior ROI, this means that it can skip the complex candidate region generation, classification and positioning in traditional methods, directly process the classroom monitoring image, and output the ROI region containing the student behavior. This data processing method reduces the complexity and computational complexity of the intermediate links, making the detection process more efficient and direct. In addition, the Yolo model has a unique multi-scale feature fusion capability, which can integrate feature maps of different scales to detect targets of different sizes. In the classroom monitoring scene, due to the different distances between students and the camera, the sizes of students in the image vary greatly. The Yolo model fully utilizes the advantages of multi-scale feature fusion by extracting features from convolutional layers at different levels. Among them, the high-level convolutional layer has strong semantic information, which is more conducive to detecting large targets; while the bottom-level convolutional layer retains more image detail information, which is more helpful for detecting small targets. By cleverly integrating these different-scale features, the Yolo model can adapt well to student behavior targets of different sizes in classroom images. Whether it is a student who is closer to the camera and appears larger in the image, or a student who is farther away and appears smaller in the image, the ROI area where their behavior is located can be accurately detected, effectively improving the accuracy of detection.

[0031] The following is a detailed description of a specific implementation process of "inputting the classroom monitoring image into the student behavior region of interest detector based on the Yolo model to obtain the student behavior ROI region image": First, make full preparations for using the Yolo model. In terms of model selection, it is necessary to comprehensively consider factors such as the characteristics of the classroom monitoring scene, the required detection accuracy, and computing resources, and select a suitable model from a variety of Yolo model versions, such as Yolov5, Yolov7, or Yolov8. These models are pre-trained on different data sets and have different performances. After determining the model, download the corresponding model code and pre-trained weight files from the official code library or open source platform. Next, build a development environment that adapts to the model operation. This requires the installation of the Python programming language environment, as well as the installation of a deep learning framework such as PyTorch, and a series of related dependent libraries to ensure that the model can run smoothly. After the environment is set up, since the pre-trained model is trained based on a general data set, it is difficult to fully fit the classroom monitoring scenario, so the model needs to be fine-tuned. This requires collecting a large number of classroom surveillance images, carefully annotating student behaviors such as raising hands, standing up to speak, walking around, etc., and generating annotation files (txt files) that conform to the YOLO format. The annotated data set is reasonably divided into training sets, validation sets, and test sets. The training set is used to fine-tune the model and continuously adjust the model parameters to improve its ability to detect classroom student behaviors.

[0032] After the preparations are ready, start preprocessing the classroom monitoring images. Use image processing libraries such as OpenCV to read classroom monitoring images, ensuring that various common image formats, such as JPEG, PNG, etc., can be accurately read. Since different versions of the Yolo model have specific requirements for the input image size, usually square, such as 640x640 pixels, the read image needs to be resized. During the adjustment process, in order to avoid image deformation affecting the detection results, the aspect ratio of the image must be maintained, and methods such as bilinear interpolation are often used to scale the image. After resizing, the image needs to be normalized to scale the image pixel values ​​from the range of [0,255] to the range of [0,1]. This can not only speed up the training and reasoning of the model, but also improve the stability of the model.

[0033] After image preprocessing, the target detection stage based on the Yolo model is entered. The fine-tuned Yolo model is loaded into memory and switched to inference mode. In inference mode, the model focuses on forward propagation and no longer performs gradient calculations and parameter updates, thereby efficiently outputting detection results. The preprocessed classroom monitoring image is input into the loaded Yolo model for forward propagation. The model will carry out complex processing on the image. Through a series of convolution, pooling and fully connected layer operations, image features are extracted and target detection is performed to predict the target category, bounding box position and confidence that may exist in the image. However, the detection results output by the model may contain multiple overlapping bounding boxes, which may correspond to the same target. In order to remove redundant bounding boxes and retain only the bounding boxes with the highest confidence, the non-maximum suppression (NMS) algorithm needs to be used. The algorithm screens based on the confidence and overlap of the bounding box (such as the intersection over union ratio IoU), and finally outputs a unique bounding box for each target.

[0034] After target detection, the student behavior ROI area needs to be screened and extracted. According to the target category output by the model, the categories related to student behavior are screened out. The bounding box information of each student behavior target is extracted from the screened detection results, including the coordinates of the upper left corner of the bounding box as well as the width and height. These bounding boxes accurately define the region of interest (ROI) of the student behavior. Finally, based on the determined bounding box information, the original classroom monitoring image is cropped using the image processing library to extract the image of each student behavior ROI area. The cropped image only contains the student behavior target, removing the background and other irrelevant information, which can facilitate subsequent feature extraction and behavior analysis.

[0035] The last step is the result output and storage. The extracted student behavior ROI area images are saved to the specified folder for subsequent use. The images can be saved in common image formats such as JPEG, PNG, etc. with the help of the image processing library. At the same time, the relevant metadata of each ROI area image is recorded, including the corresponding original image file name, target category, bounding box coordinates, etc. These metadata can provide key basis for data tracing.

[0036] In an embodiment of the present application, the behavior local image coding feature extraction subunit 1312 is used to extract student behavior multi-scale image features from the student behavior ROI area image to obtain a set of student behavior local image coding features. Specifically, in an embodiment of the present application, the behavior local image coding feature extraction subunit is used to: use a pyramid behavior feature capturer based on the SimAM module to extract student behavior multi-scale image features from the student behavior ROI area image to obtain a set of student behavior local image coding feature vectors as the set of student behavior local image coding features. Accordingly, considering that in a classroom scene, the scale of presentation of students' behaviors in images is significantly different. Due to the different distances between students and the camera, and the size differences of different behavioral actions themselves, the scale of student behavior targets in the ROI area image varies greatly. For example, when a student sitting in the front row of the classroom raises his hand, his arm occupies a larger area in the image, while when a student sitting in the back row raises his hand, his arm occupies a smaller proportion in the image. By extracting multi-scale image features of student behavior from the ROI regional image of student behavior, feature capture can be performed for targets of different scales to ensure that both large-scale hand-raising actions and small-scale subtle behavioral changes can be effectively identified. In a specific embodiment of the present application, it is necessary to use a pyramid behavior feature capturer based on the SimAM module to extract multi-scale image features of student behavior from the ROI regional image of the student behavior to obtain a set of local image coding feature vectors of student behavior as a set of local image coding features of the student behavior. Specifically, the SimAM module can assign importance weights to each pixel based on the principle of local self-similarity of the image. In the pyramid behavior feature capturer, this feature enables the capturer to accurately focus on the key areas and key features of student behavior when extracting multi-scale image features. When extracting hand-raising behavior features, it can automatically highlight areas such as arms and palms that are closely related to hand-raising behavior, and suppress irrelevant information in the background. This ability to accurately focus on key features can effectively improve the quality and pertinence of feature extraction, making the extracted features more representative, thereby improving the accuracy of subsequent behavior recognition. The pyramid behavior feature capturer uses a pyramid structure to obtain image features of different scales. It can organically combine the overall semantic information in large-scale features with the detailed information in small-scale features to form a more comprehensive and discriminative student behavior local image encoding feature vector. It is particularly important to note that the SimAM module is a parameter-free attention mechanism. In the pyramid behavior feature capturer based on the SimAM module, no additional network parameters are added. This means that while improving the multi-scale image feature extraction effect, the computational complexity and training time of the model will not be significantly increased.That is, it can improve the performance of the model without increasing the hardware burden, ensuring that the system can quickly and accurately extract multi-scale image features of student behavior and meet the strict efficiency requirements of real-time classroom behavior detection.

[0037] The following is a detailed description of a specific implementation process of "using a pyramid behavior feature capturer based on the SimAM module to extract multi-scale image features of student behavior from the student behavior ROI area image to obtain a set of local image encoding feature vectors of student behavior": First, the obtained student behavior ROI area map needs to be preprocessed. The first step of preprocessing is to adjust the image size and uniformly adjust the image to a specific specification, such as 224×224 pixels. The purpose of this is to ensure the consistency of subsequent processing so that the model can better process the image. Then normalization is performed to map the image pixel values ​​to a specific range, such as [0,1] or [-1,1]. In addition, image enhancement operations such as random flipping, rotation, brightness adjustment, etc. can also be performed. These operations can increase the diversity of the data, so that the model can have a better recognition effect when facing student behaviors under different postures and lighting conditions, and improve the generalization ability of the model.

[0038] Next, we need to build a pyramid behavior feature capturer. It processes the input image through different levels of convolutional layers and pooling layers. The lower-level networks mainly extract detailed features of the image, such as edges, textures, etc., while the higher-level networks extract more abstract semantic features. The pyramid behavior feature capturer is usually built using the architecture of a convolutional neural network, which includes multiple convolutional layers, pooling layers, and activation function layers. The convolutional layer is responsible for extracting local features of the image. It slides on the image through different convolution kernels to extract features from different areas of the image. The pooling layer is used to reduce the dimension of the feature map and reduce the amount of calculation while retaining important feature information. The activation function layer, such as ReLU, introduces nonlinear factors to enhance the expressiveness of the model, allowing the model to learn more complex features.

[0039] The SimAM module is a key part of the entire feature capturer. It is a biologically inspired attention mechanism module that simulates the selective response of neurons in the human visual nervous system to different stimuli. Traditional attention mechanisms generally allocate attention based on channels or spatial dimensions, while the SimAM module starts from the perspective of neurons. It believes that the response of neurons depends not only on their own inputs, but also on the states of surrounding neurons. The importance of neurons is measured by calculating the energy function of neurons, which takes into account the input value of the neuron itself, the average input value of surrounding neurons, and the difference between them. In the specific implementation, the feature map from a certain layer of the pyramid behavior feature capturer is first received as input. Then the energy function is calculated for each neuron in the input feature map, and the optimal weight of each neuron is solved through an optimization problem. According to the calculation result of the energy function, the attention weight of each neuron is generated. The larger the weight value, the more important the neuron is to the current task. Finally, the generated attention weight is applied to the input feature map, and each neuron in the feature map is weighted, so as to enhance important features and suppress irrelevant features. In addition, the SimAM module is inserted after each convolution layer of the pyramid behavior feature capturer. In this way, in the process of extracting features at each layer, the SimAM module can adjust the attention of the feature map of that layer, so that the model can pay more attention to the key features related to student behavior.

[0040] When the input student behavior ROI area image is processed by different layers of the pyramid behavior feature capturer, feature extraction is performed layer by layer. The convolution layer of each layer extracts local features of the image, and the SimAM module enhances attention to these features. As the number of network layers increases, the size of the feature map gradually decreases, the number of channels gradually increases, and the extracted features become more and more abstract. In order to obtain a more comprehensive and richer multi-scale feature representation, it is necessary to fuse the feature maps output by different layers of the pyramid behavior feature capturer. Different scale feature information can be integrated together by splicing, addition, etc., so that both the detailed features of the image and the semantic features can be used to improve the recognition ability of student behavior.

[0041] Next, a global pooling operation is performed on the fused multi-scale feature maps, such as global average pooling or global maximum pooling. Global pooling converts each feature map into a feature vector of fixed length for easy subsequent processing and analysis. Combining feature vectors of different scales together forms a set of feature vectors encoding local images of student behavior.

[0042] Finally, the generated feature vector set needs to be post-processed. The feature vector set is normalized, such as using L2 normalization, so that all feature vectors have the same scale and distribution. This can improve the stability and accuracy of the subsequent model. After post-processing, the processed set of student behavior local image encoding feature vectors is output as the final result for data processing and analysis in subsequent units.

[0043] In the embodiment of the present application, the behavior image semantic level aggregate coding feature generation subunit 1313 is used to perform student behavior feature aggregate coding of the set of student behavior local image coding features with a centralized information core to obtain the student behavior image semantic level aggregate coding features. Specifically, in the embodiment of the present application, the behavior image semantic level aggregate coding feature generation subunit includes: a behavior feature coarse-grained aggregation first-level subunit, used to input the set of student behavior local image coding feature vectors into the behavior information core coarse-grained aggregation network to obtain the student behavior image semantic coarse-grained aggregation coding feature vector; a behavior feature fine-grained aggregation compensation first-level subunit, used to perform behavior information fine-grained compensation aggregation on the set of student behavior local image coding feature vectors to obtain the student behavior image semantic fine-grained compensation aggregation coding feature vector; a behavior semantic feature aggregation residual first-level subunit, used to input the student behavior image semantic coarse-grained aggregation coding feature vector and the student behavior image semantic fine-grained compensation aggregation coding feature vector into the residual unit to obtain the student behavior image semantic level aggregate coding feature vector as the student behavior image semantic level aggregate coding feature.

[0044] It should be understood that after performing multi-scale image feature extraction on the student behavior ROI area image, the obtained set of student behavior local image coding feature vectors contains rich information at different scales, but these features are relatively scattered, lacking an overall understanding and in-depth mining of student behavior. That is, each student behavior local image coding feature vector may only reflect the characteristics of a certain part or specific scale of the student behavior. If effective aggregation coding is not performed, it is difficult to grasp the essential characteristics of the student behavior from a global perspective. Based on this, the present application needs to perform student behavior feature aggregation coding of the set of student behavior local image coding features with a concentrated information core to obtain the semantic-level aggregate coding features of the student behavior image. That is, through the feature aggregation coding operation of the concentrated information core, the global feature summary is constructed using coarse-grained aggregation, which can integrate these scattered local features to form a more macro and comprehensive understanding of student behavior, so as to make up for the lack of a global perspective in the previous feature extraction process. At the same time, this data processing can also perform high-fidelity modeling of the local detailed characteristics of each behavior, carefully enhancing the significance of the personalized characteristics of each behavior, making it possible to deeply analyze the uniqueness of each student's behavior, such as the special behavioral habits of individual students or subtle movements expressed by emotional changes, thereby meeting the dual needs of global and local information in complex classroom behavior analysis.

[0045] Specifically, the set of student behavior local image encoding feature vectors is first input into the behavior information core coarse-grained aggregation network to obtain the student behavior image semantic coarse-grained aggregation encoding feature vector. The above process can be expressed as: ;in, represents the set of local image encoding feature vectors of student behavior, The first, second, and third vectors in the set of local image encoding feature vectors of student behavior are and The local image encoding feature vector of student behavior, Indicates taking The maximum value of Indicates taking The minimum value of express The corresponding extreme value mean, represents the normalization function, express The corresponding student behavior local image encoding feature weights, express The number of vectors in Represents the semantic coarse-grained aggregated encoding feature vector of student behavior images.

[0046] It should be understood that the set of local image coding feature vectors of student behavior is the result of multi-scale image feature extraction, which reflects the local or specific scale characteristics of student behavior, and the information is relatively scattered. The behavior information kernel coarse-grained aggregation network uses the information kernel as the core to compress the global features and integrate these scattered local features. Specifically, when analyzing students' hand-raising behavior in class, the network can comprehensively consider local features such as arm movements and hand postures at different scales, and model the associations between each local feature in an implicit high-dimensional space through the nonlinear mapping of the information kernel, thereby forming a student behavior image semantic coarse-grained aggregation coding feature vector that can represent the macro characteristics of students' hand-raising behavior.

[0047] Specifically, in an embodiment of the present application, the first-level subunit of fine-grained convergence compensation of behavioral features includes: a second-level subunit for calculating behavioral semantic core convergence compensation factors, which is used to calculate the behavioral semantic core convergence compensation factors of each student behavior local image coding feature vector in the set of student behavior local image coding feature vectors relative to the student behavior image semantic coarse-grained convergence coding feature vector to obtain a set of behavioral semantic core convergence compensation factors; a second-level subunit for calculating behavioral semantic core convergence compensation weight factors, which is used to perform compensation explicit modeling based on a gating function on the set of behavioral semantic core convergence compensation factors to obtain a set of behavioral semantic core convergence compensation weight factors; a second-level subunit for generating fine-grained compensation convergence coding features, which is used to input the set of behavioral semantic core convergence compensation weight factors, the set of student behavior local image coding feature vectors, and the student behavior image semantic coarse-grained convergence coding feature vector into a behavioral fine-grained dynamic compensation convergence network to obtain the student behavior image semantic fine-grained compensation convergence coding feature vector.

[0048] More specifically, in an embodiment of the present application, the behavioral semantic core convergence compensation factor calculation secondary subunit is used to: modulate and encode the student behavior local image coding feature vector to obtain the student behavior local image modulation coding feature vector; modulate and encode the student behavior image semantic coarse-grained convergence coding feature vector to obtain the student behavior image semantic coarse-grained convergence modulation coding feature vector; calculate the student behavior local difference modulation coding feature vector between the student behavior local image modulation coding feature vector and the student behavior image semantic coarse-grained convergence modulation coding feature vector; multiply the student behavior local difference modulation coding feature vector with the student behavior weight matrix and then add it by position with the student behavior bias value to obtain the student behavior weighted local difference modulation coding feature vector, and multiply the student behavior weighted local difference modulation coding feature vector with the transformation score vector to obtain the behavioral semantic core convergence compensation factor. The above process can be expressed by the formula: ;in, The first in the set of local image encoding feature vectors representing student behavior The local image encoding feature vector of student behavior, Represents the semantic coarse-grained aggregation encoding feature vector of student behavior images, represents point convolutional coding, represents the first weight matrix, represents the activation function, express The corresponding student behavior local image modulation coding feature vector, represents the second weight matrix, Represents the semantic coarse-grained aggregation modulation coding feature vector of student behavior images, It means to subtract by position point. Indicates taking the absolute value, express and The local difference between the student behaviors modulates the encoding feature vector, represents matrix multiplication, express The corresponding transformed score vector, express The corresponding student behavior weight matrix, express The corresponding student behavior bias value, express The corresponding behavioral semantic kernel convergence compensation factor.

[0049] It should be understood that in the process of obtaining the semantic coarse-grained converged coding feature vector of the student behavior image, although the global convergence method enhances the expressive power of the overall modeling, due to the compressive aggregation of the global features, it is inevitable that the personalized information of some important behaviors will be diluted or even completely lost. That is, the local image coding feature vector of the student behavior contains rich details and personalized characteristics of the student behavior, such as the unique movement habits of individual students when raising their hands. However, when generating the semantic coarse-grained converged coding feature vector, these personalized information may be weakened or ignored due to global aggregation. In order to ensure and utilize these important personalized behavior information, in this application, it is necessary to calculate the behavior semantic core convergence compensation factor of each student behavior local image coding feature vector in the set of student behavior local image coding feature vectors relative to the student behavior image semantic coarse-grained converged coding feature vector to obtain a set of behavior semantic core convergence compensation factors. The behavior semantic core convergence compensation factor is calculated by measuring the deviation between each student behavior local image coding feature vector and the student behavior image semantic coarse-grained converged coding feature vector. This process is equivalent to formulating exclusive correction rules for each local feature vector. This allows the personalized information of the behavior represented by each local feature vector to be retained. For example, when analyzing students' classroom behavior, even if some students' behaviors are not prominent in the global characteristics, their unique behaviors can be retained through the behavioral semantic kernel aggregation compensation factors and will not be completely erased due to global aggregation, which can provide a basis for subsequent in-depth analysis of students' individual behaviors.

[0050] More specifically, in an embodiment of the present application, the student behavior bias value is related to the modulation coding feature vector of the student behavior local image and the modulation coding feature vector of the semantic coarse-grained convergence of the student behavior image, wherein, in response to the second norm of the modulation coding feature vector of the student behavior local image being less than the second norm of the modulation coding feature vector of the semantic coarse-grained convergence of the student behavior image, the logarithmic function value of the logarithm with a base of 2 and the value added by the constant 1 to the ratio of the second norm of the modulation coding feature vector of the student behavior local image and the second norm of the modulation coding feature vector of the semantic coarse-grained convergence of the student behavior image is calculated to obtain the student behavior bias value; in response to the second norm of the modulation coding feature vector of the student behavior local image being greater than or equal to the second norm of the modulation coding feature vector of the semantic coarse-grained convergence of the student behavior image, the ratio of the second norm of the modulation coding feature vector of the student behavior local image to the second norm of the modulation coding feature vector of the semantic coarse-grained convergence of the student behavior image is used as the student behavior bias value. The above expression can be expressed by the formula: ;in, represents the two-norm of a vector, express The corresponding student behavior local image modulation coding feature vector, Represents the semantic coarse-grained aggregation modulation coding feature vector of student behavior images, represents the logarithmic function value with base 2, express The corresponding student behavior bias value.

[0051] Here, for the deviation compensation between the local image coding feature vector of student behavior and the coarse-grained convergence coding feature vector of the semantics of student behavior image, the performance deviation of the kernel convergence strategy as a situational strategy can be measured by quantifying the regret metric based on the information kernel compression hypothesis in the kernel convergence decision process, that is, the game counterfactual regret value. Specifically, the counterfactual regret value is provided with a normalized description of the decision point loss based on the strategic action, that is, the vector norm representation of the local image coding feature vector of student behavior and the coarse-grained convergence coding feature vector of the semantics of student behavior image through the norm representation of the vector. Then, for the possible differences in the game scenarios of vector distribution actions, the compensation rules of node personalized information are corrected by the information distribution degree of the regret value and the relative distribution amplitude of the regret value, so that the personalized information of the local image coding feature of student behavior is considered as the untaken action in the decision, and the biased compensation is performed based on the potential benefits of the information kernel convergence hypothesis.

[0052] Then, the set of behavioral semantic core convergence compensation factors is subjected to compensation explicit modeling based on the gating function to obtain a set of behavioral semantic core convergence compensation weight factors. The above process can be expressed by the formula: ;in, express The corresponding behavioral semantic kernel convergence compensation factor, is the preset threshold, represents the gating function, express The corresponding behavioral semantic kernel converges the compensation weight factor.

[0053] It should be understood that although the set of local image encoding feature vectors of student behavior retains the personalized information of student behavior, not all the information contained therein is of equal value for behavior analysis. Compensation explicit modeling based on gating functions dynamically selects information according to nonlinear constraints, and can screen out information that plays a key role in student behavior analysis. Specifically, when analyzing students' hand-raising behavior in class, the process can select the parts closely related to the hand-raising action from many compensation factors, such as the compensation factors corresponding to key features such as the angle of arm raising, and enhance their effects to make these local significant features more prominent. In this way, in subsequent analysis, the model can focus more on important behavioral features, avoid being interfered by irrelevant or redundant information, and improve the accuracy of judging students' hand-raising behavior.

[0054] Next, the set of behavioral semantic core convergence compensation weight factors, the set of student behavior local image encoding feature vectors and the student behavior image semantic coarse-grained convergence encoding feature vector are input into the behavior fine-grained dynamic compensation convergence network to obtain the student behavior image semantic fine-grained compensation convergence encoding feature vector. The above process can be expressed as: ;in, express The number of vectors in express The corresponding behavioral semantic kernel convergence compensation weight factor, The first in the set of local image encoding feature vectors representing student behavior The local image encoding feature vector of student behavior, Represents the semantic coarse-grained aggregation encoding feature vector of student behavior images, Representing the semantic fine-grained compensated aggregation encoding feature vector of student behavior images.

[0055] It should be understood that although the semantic coarse-grained convergence encoding feature vector of student behavior image provides overall behavior trend information, it cannot accurately reflect the subtle differences and personalized characteristics of student behavior. The set of student behavior local image encoding feature vectors retains the personalized local information of student behavior, and the set of behavioral semantic core convergence compensation weight factors can filter and regulate the effect of this information. Combining them with the semantic coarse-grained convergence encoding feature vector of student behavior image and inputting them into the behavior fine-grained dynamic compensation convergence network can meet the needs of fine-grained analysis of student behavior and help the model to have a deeper understanding of each student's behavior performance, such as identifying the special behavior patterns of individual students or subtle movements caused by emotional changes. In short, the behavior fine-grained dynamic compensation convergence network dynamically adjusts the characteristics of the current student behavior by using the behavioral semantic core convergence compensation weight factors to achieve a high-fidelity description of the details of the student behavior. This dynamic adjustment mechanism is similar to the attention mechanism, which can achieve the coordinated expression of features at the global and local levels of behavior, and can provide strong support for the comprehensive and in-depth analysis of student classroom behavior.

[0056] Finally, the coarse-grained aggregated coding feature vector of the student behavior image semantics and the fine-grained compensated aggregated coding feature vector of the student behavior image semantics are input into the residual unit to obtain the semantic-level aggregated coding feature vector of the student behavior image. The above process can be expressed as: ;in, Represents the semantic coarse-grained aggregation encoding feature vector of student behavior images, Represents the semantic fine-grained compensation aggregation encoding feature vector of student behavior images, and represents the weighted hyperparameter, Represents the semantic-level aggregated encoding feature vector of student behavior images.

[0057] It should be understood that the semantic coarse-grained converged encoding feature vector of student behavior images contains the global features of student behavior, while the semantic fine-grained compensated converged encoding feature vector of student behavior images focuses on capturing local significant details. Through the fusion mechanism of residual units, fine-grained features can inherit the global vision of coarse-grained features while retaining their own local details, thus forming a refined feature expression that combines global and local features - the semantic-level aggregated encoding feature vector of student behavior images. It is particularly important to note that in the process of deep learning model training, as the number of network layers increases, the gradient vanishing problem often occurs, causing the information to gradually decay during the transmission process, affecting the learning effect of the model. The introduction of residual units effectively solves this problem. Its unique residual connection structure allows features to be directly transmitted across layers, ensuring the stable propagation of student behavior features in the network. At the same time, this can enable the model to learn accurate student behavior feature representations more efficiently, thereby improving the saliency and robustness of the encoding results. When faced with behavioral data from different classroom environments and different student groups, the optimized model can more keenly capture key behavioral characteristics, and has stronger resistance to noise such as light changes and background interference, and can stably output accurate student behavior-related characteristics, enhancing the applicability and reliability of the model in complex real-world scenarios.

[0058] In the embodiment of the present application, the identification unit 132 is used to obtain the classroom behavior recognition result based on the semantic level aggregate coding feature of the student behavior image. Specifically, in the embodiment of the present application, the identification unit is used to: input the semantic level aggregate coding feature vector of the student behavior image into the classifier-based classroom behavior identifier to obtain the classroom behavior recognition result, and the classroom behavior recognition result is the type label of the classroom behavior. It should be understood that although the semantic level aggregate coding feature vector of the student behavior image contains rich student behavior information, this information is only a numerical expression of the behavior characteristics, and there is no clear behavior category indication itself. The classifier-based classroom behavior identifier has the ability to classify and judge this feature vector. Specifically, after the classifier is trained with a large amount of sample data with clear behavior labels, it learns the correspondence between different behavior features and specific behavior types. By inputting the semantic level aggregate coding feature vector of the student behavior image, the classifier can "translate" these feature vectors into specific behavior type labels based on the learned knowledge, thereby realizing the classification of students' classroom behavior. Teachers can understand students' participation and learning status in real time based on the classroom behavior recognition results. If the recognition result shows that many students raise their hands, the teacher can know that the students have questions about the current teaching content or actively participate in the discussion, so as to adjust the teaching rhythm in time, answer students' questions, or guide students to have a more in-depth discussion to improve the teaching effect. On the contrary, if it is found that some students are in a state of non-focused behavior for a long time, such as frequent distraction, making small movements, etc., the teacher can take corresponding measures, such as changing the teaching method, increasing interactive links, etc., to attract students' attention and enhance students' learning enthusiasm. In particular, in a specific embodiment of the present application, the semantic-level aggregated encoding feature vector of the student behavior image is input into a classifier-based classroom behavior identifier to obtain the classroom behavior recognition result, and the classroom behavior recognition result is a type label of classroom behavior, including: using the fully connected layer of the classifier to fully connect the semantic-level aggregated encoding feature vector of the student behavior image to obtain the semantic-level aggregated fully connected encoding classification feature vector of the student behavior image; the semantic-level aggregated fully connected encoding classification feature vector of the student behavior image is input into the Softmax classification function of the classifier to obtain the classroom behavior recognition result.

[0059] In summary, the AI ​​and cloud computing-based smart education management system 100 based on the embodiment of the present application is explained, which first obtains classroom monitoring images on the end side and detects student behavior. When student behavior is detected, it is uploaded to the cloud computing platform, otherwise the next frame of classroom monitoring images is received, and the image analysis technology based on AI and cloud computing is used on the cloud computing platform to detect the student behavior area of ​​interest and multi-scale image extraction processing of the classroom monitoring image to obtain the student behavior local image coding feature set, and the student behavior feature aggregation representation based on the student behavior local image coding feature set is used to intelligently obtain the classroom behavior recognition result. In this way, the student's hand-raising behavior can be more accurately identified, and the pattern and dynamic change trend of classroom participation can be fully understood, which can effectively improve the intelligence of teaching management.

[0060] Figure 5 Flowchart of a smart education management method based on AI and cloud computing according to an embodiment of the present application. Figure 5 As shown, in the smart education management method based on AI and cloud computing, it includes: S110, obtaining a classroom monitoring image, and judging the student behavior of the classroom monitoring image. If there is student behavior in the classroom monitoring image, uploading the classroom monitoring image to the cloud computing platform; S120, if there is no student behavior in the classroom monitoring image, receiving the next frame of the classroom monitoring image; S130, on the cloud computing platform, analyzing the student behavior of the classroom monitoring image to obtain a classroom behavior recognition result.

[0061] Here, those skilled in the art can understand that the specific operations of each step in the above-mentioned smart education management method based on AI and cloud computing have been referred to above. Figures 1 to 4 The description of the smart education management system based on AI and cloud computing has been introduced in detail, and therefore, its repeated description will be omitted.

[0062] In summary, the smart education management method based on AI and cloud computing based on the embodiment of the present application is explained, which first obtains classroom monitoring images on the end side and detects student behavior. When student behavior is detected, it is uploaded to the cloud computing platform, otherwise the next frame of classroom monitoring images is received, and the image analysis technology based on AI and cloud computing is used on the cloud computing platform to detect the student behavior area of ​​interest and multi-scale image extraction processing of the classroom monitoring image to obtain the local image coding feature set of the student behavior, and the classroom behavior recognition result is intelligently obtained based on the student behavior feature aggregation representation of the local image coding feature set of the student behavior. In this way, the student's hand-raising behavior can be more accurately identified, and the pattern and dynamic change trend of classroom participation can be fully understood, which can effectively improve the intelligence of teaching management.

Claims

1. A smart education management system based on AI and cloud computing, characterized in that: include: A data acquisition and transmission module is used to obtain classroom monitoring images and to judge student behavior in the classroom monitoring images. If student behavior exists in the classroom monitoring images, the classroom monitoring images are uploaded to the cloud computing platform. An image receiving module, used for receiving the next frame of the classroom monitoring image if there is no student behavior in the classroom monitoring image; A classroom behavior recognition module is used to perform student behavior analysis on the classroom monitoring image on the cloud computing platform to obtain classroom behavior recognition results, wherein the classroom behavior recognition module includes: a behavior aggregation unit, used to perform semantic aggregation of behavior interest regions on the classroom monitoring image to obtain semantic-level aggregation coding features of student behavior images; and an identification unit, used to obtain the classroom behavior recognition results based on the semantic-level aggregation coding features of the student behavior images.

2. The intelligent education management system based on AI and cloud computing according to claim 1 is characterized in that: The data acquisition and transmission module includes: a classroom monitoring image collection unit, which is used to obtain classroom monitoring images collected by a camera on the terminal side; a classroom monitoring image detection and transmission unit, which is used to detect whether there is student behavior in the classroom monitoring image on the terminal side, and if there is student behavior in the classroom monitoring image, upload the classroom monitoring image to the cloud computing platform.

3. The intelligent education management system based on AI and cloud computing according to claim 2 is characterized in that: The behavior aggregation unit includes: a behavior region of interest detection subunit, which is used to input the classroom monitoring image into a student behavior region of interest detector based on the Yolo model to obtain a student behavior ROI area image; a behavior local image coding feature extraction subunit, which is used to perform student behavior multi-scale image feature extraction on the student behavior ROI area image to obtain a set of student behavior local image coding features; a behavior image semantic level aggregation coding feature generation subunit, which is used to perform student behavior feature aggregation coding with a concentrated information core on the set of student behavior local image coding features to obtain the student behavior image semantic level aggregation coding features.

4. The intelligent education management system based on AI and cloud computing according to claim 3 is characterized in that: The behavior local image coding feature extraction subunit is used to: use a pyramid behavior feature capturer based on the SimAM module to perform student behavior multi-scale image feature extraction on the student behavior ROI area image to obtain a set of student behavior local image coding feature vectors as the set of student behavior local image coding features.

5. The intelligent education management system based on AI and cloud computing according to claim 4 is characterized in that: The behavior image semantic-level aggregated coding feature generation subunit includes: a behavior feature coarse-grained aggregation first-level subunit, which is used to input the set of student behavior local image coding feature vectors into the behavior information core coarse-grained aggregation network to obtain a student behavior image semantic coarse-grained aggregated coding feature vector; a behavior feature fine-grained aggregation compensation first-level subunit, which is used to perform behavior information fine-grained compensation aggregation on the set of student behavior local image coding feature vectors to obtain a student behavior image semantic fine-grained compensation aggregation coding feature vector; a behavior semantic feature aggregation residual first-level subunit, which is used to input the student behavior image semantic coarse-grained aggregation coding feature vector and the student behavior image semantic fine-grained compensation aggregation coding feature vector into a residual unit to obtain a student behavior image semantic-level aggregated coding feature vector as the student behavior image semantic-level aggregated coding feature.

6. The intelligent education management system based on AI and cloud computing according to claim 5 is characterized in that: The first-level subunit of fine-grained convergence compensation of behavioral features includes: a second-level subunit for calculating behavioral semantic core convergence compensation factors, which is used to calculate the behavioral semantic core convergence compensation factors of each student behavior local image coding feature vector in the set of student behavior local image coding feature vectors relative to the student behavior image semantic coarse-grained convergence coding feature vector to obtain a set of behavioral semantic core convergence compensation factors; a second-level subunit for calculating behavioral semantic core convergence compensation weight factors, which is used to perform compensation explicit modeling based on a gating function on the set of behavioral semantic core convergence compensation factors to obtain a set of behavioral semantic core convergence compensation weight factors; and a second-level subunit for generating fine-grained compensation convergence coding features, which is used to input the set of behavioral semantic core convergence compensation weight factors, the set of student behavior local image coding feature vectors and the student behavior image semantic coarse-grained convergence coding feature vector into a behavioral fine-grained dynamic compensation convergence network to obtain the student behavior image semantic fine-grained compensation convergence coding feature vector.

7. The intelligent education management system based on AI and cloud computing according to claim 6 is characterized in that: The behavior semantic core convergence compensation factor calculation secondary subunit is used to: modulate and encode the student behavior local image coding feature vector to obtain the student behavior local image modulation coding feature vector; modulate and encode the student behavior image semantic coarse-grained convergence coding feature vector to obtain the student behavior image semantic coarse-grained convergence modulation coding feature vector; calculate the student behavior local difference modulation coding feature vector between the student behavior local image modulation coding feature vector and the student behavior image semantic coarse-grained convergence modulation coding feature vector; multiply the student behavior local difference modulation coding feature vector with the student behavior weight matrix and then add them with the student behavior bias value by position to obtain the student behavior weighted local difference modulation coding feature vector, and multiply the student behavior weighted local difference modulation coding feature vector with the transformation score vector to obtain the behavior semantic core convergence compensation factor.

8. The intelligent education management system based on AI and cloud computing according to claim 7 is characterized in that: The student behavior bias value is related to the student behavior local image modulation coding feature vector and the student behavior image semantic coarse-grained convergence modulation coding feature vector, wherein, in response to the second norm of the student behavior local image modulation coding feature vector being less than the second norm of the student behavior image semantic coarse-grained convergence modulation coding feature vector, the logarithmic function value of the logarithm with a base of 2 and the value added by a constant 1 to the ratio of the second norm of the student behavior local image modulation coding feature vector and the second norm of the student behavior image semantic coarse-grained convergence modulation coding feature vector is calculated to obtain the student behavior bias value; in response to the second norm of the student behavior local image modulation coding feature vector being greater than or equal to the second norm of the student behavior image semantic coarse-grained convergence modulation coding feature vector, the ratio of the second norm of the student behavior local image modulation coding feature vector to the second norm of the student behavior image semantic coarse-grained convergence modulation coding feature vector is used as the student behavior bias value.

9. The intelligent education management system based on AI and cloud computing according to claim 8, characterized in that: The recognition unit is used to: input the semantic-level aggregated coding feature vector of the student behavior image into a classifier-based classroom behavior recognizer to obtain the classroom behavior recognition result, and the classroom behavior recognition result is a type label of the classroom behavior.

10. A smart education management method based on AI and cloud computing, characterized in that: include: Acquire classroom monitoring images, and judge student behavior on the classroom monitoring images. If student behavior exists in the classroom monitoring images, upload the classroom monitoring images to the cloud computing platform; If there is no student behavior in the classroom monitoring image, receiving the next frame of the classroom monitoring image; On the cloud computing platform, the classroom monitoring image is subjected to student behavior analysis to obtain classroom behavior recognition results, including: semantic aggregation of behavior interest regions on the classroom monitoring image to obtain semantic-level aggregation coding features of student behavior images; Based on the semantic-level aggregate coding features of the student behavior image, the classroom behavior recognition result is obtained.

Images