Classroom behavior detection method based on dimension diffusion perception adaptive fusion

Abstract

The invention discloses a classroom behavior detection method based on dimension diffusion perception adaptive fusion, and the method comprises the steps: replacing a part of C2f layer of a Backbone network in a YOLOv8 model with parallelization plaque perception attention PPA, constructing a dimension perception fusion diffusion pyramid network DAIDFPN through employing a dimension perception selective fusion module DASI, and replacing a Neck feature fusion network in the YOLOv8 model. Therefore, an improved YOLOv8 model is obtained, and classroom behavior detection is carried out on the classroom image by using the improved YOLOv8 model. Compared with other methods, the classroom behavior detection method based on dimension diffusion perception adaptive fusion does not need to depend on a large number of data annotation data and computing resources, has the advantages of being small in computing overhead, high in real-time performance and high in detection precision, is low in cost in smart classroom deployment and has good practical application value.

Description

Technical Field

[0001] The present invention belongs to the technical field of target detection, and specifically relates to a classroom behavior detection method based on dimensional diffusion perception adaptive fusion. Background Art

[0002] In traditional teaching activities, classroom teaching has always been the main form of current school education and teaching, and it is also the focus of educational research. In the past, the observation and recording of students' classroom learning behavior for classroom teaching mainly relied on teachers' subjective judgment. This method is not only time-consuming and labor-intensive, but also easily affected by teachers' personal preferences and emotions, making it difficult to achieve objective, comprehensive and timely classroom behavior analysis. With the development of artificial intelligence technology, it has become a popular trend to empower traditional education forms with computer vision deep learning technology. Compared with traditional classroom behavior recognition and analysis methods, it has the characteristics of high automation, real-time and high efficiency, which can provide accurate and objective data support for the subsequent analysis of students' classroom teaching behavior and the construction of a teaching quality evaluation system, thereby promoting the unification of subjective and objective evaluation, qualitative analysis and quantitative analysis, and making education and teaching feedback more comprehensive and accurate.

[0003] Existing algorithms for detecting student classroom behavior are mainly divided into three categories: the first is based on video action recognition, which directly extracts visual features from classroom videos and can fully consider the dynamic characteristics of the time and space dimensions, but it also requires a relatively more complex network structure and higher computational complexity, and requires a large amount of sample data labeling work in the early stage. The relevant theoretical research is not yet mature, so it is not suitable for current practical applications; the second is key point recognition based on posture estimation algorithm, which recognizes behavior by extracting key points and motion features of the human skeleton, which is easy to ignore some important details and is not suitable for crowded and severely occluded classroom scenes. At the same time, the computing resources required for skeleton extraction are large, which is difficult to meet the monitoring field with high real-time requirements; the third is recognition based on target detection algorithm. Due to its good performance in speed, computational overhead and accuracy, it is currently considered to be one of the most suitable algorithms for related applications in the field of student classroom behavior detection.

[0004] However, it is a complex problem to identify students' classroom behaviors in natural classroom scenes. Due to the crowded crowd in real classroom scenes, students' behaviors are unclear and severely occluded. At the same time, the scales of different positions in the classroom are not uniform, and the ratio of pixels between students in the front row and students in the back row is large. Therefore, this undoubtedly poses a higher challenge to the target detection algorithm. Researching the problems of small target detection, occlusion and scale inconsistency in classroom scenes, so as to improve the recall rate and accuracy of students' classroom behavior detection has become an urgent problem to be solved in the field of educational informatization. Summary of the invention

[0005] In view of the shortcomings of the existing technology, the present invention provides a classroom behavior detection method based on dimensional diffusion perception adaptive fusion, which utilizes a multi-branch extraction strategy and a dimensional adaptive feature fusion network to improve the model's ability to detect student behavior in classroom scenarios, and effectively solves the technical problems of unclear and severe occlusion of student behavior in the existing technology. At the same time, it has the characteristics of low computational overhead, strong real-time performance, and high detection accuracy, which is helpful to further promote the development of smart education.

[0006] The present invention can be achieved through the following technical solutions:

[0007] A classroom behavior detection method based on dimension diffusion-aware adaptive fusion is proposed. Parallelized patch-aware attention (PPA) is used to replace part of the C2f layer of the Backbone network in the YOLOv8 model. The dimension-aware selective fusion module (DASI) is used to construct a dimension-aware fusion diffusion pyramid network (DAIDFPN) to replace the Neck feature fusion network in the YOLOv8 model, thereby obtaining an improved YOLOv8 model. The improved YOLOv8 model is used to perform classroom behavior detection on classroom images.

[0008] Among them, the dimension-aware fusion diffusion pyramid network DAIDFPN adopts a two-layer structure. The first layer uses the first dimension-aware selective fusion module DASI to perform preliminary multi-dimensional aggregation on the different-scale feature maps output by the Backbone backbone network; the second layer uses the second dimension-aware selective fusion module DASI to perform fine multi-dimensional aggregation on the multi-dimensional feature maps output by the first layer and the different-scale feature maps output by the Backbone backbone network, so as to obtain small-size, medium-size, and large-size feature maps for detection processing in the Head detection network in the YOLOv8 model.

[0009] Further, the Backbone network outputs a low-level feature map feat1, a middle-level feature map feat2, and a high-level feature map feat3;

[0010] For the first layer, the first dimension-aware selective fusion module DASI performs dimension-aware adaptive fusion on the low-level feature map feat1, the mid-level feature map feat2 and the high-level feature map feat3 to obtain the feature map feat4, and then upsamples and downsamples the feature map feat4 respectively to obtain the feature map feat5 and the feature map feat6 respectively, and then concatenates the feature map fe5 with the low-level feature map fe1, and then performs feature enhancement fusion through a C2f module to obtain the feature map feat7, and at the same time, concatenates the feature map fe6 with the high-level feature map fe3, and then performs feature enhancement fusion through a C2f module to obtain the feature map feat8.

[0011] Further, for the second layer, the second dimension-aware selective fusion module DASI performs dimension-aware adaptive fusion on the feature map feat4, the feature map feat7 and the feature map feat8 to obtain the feature map feat9, that is, the medium-size feature map feat9, and then upsamples the feature map feat9, and then splices it with the low-level feature map feat1 and the feature map feat5 and the feature map fe7, and then performs feature enhancement fusion through a C2f module to obtain the large-size feature map feat10. At the same time, the feature map feat9 is downsampled, and then spliced ​​with the high-level feature map feat3 and the feature map feat6 and the feature map fe8, and then performs feature enhancement fusion through a C2f module to obtain the small-size feature map fe11.

[0012] Furthermore, the parallelized patch-aware attention (PPA) is used to replace the C2f layers corresponding to the 2nd, 4th, and 6th layers of the Backbone network in the YOLOv8 model.

[0013] Furthermore, the layers of the Backbone network are set to be a convolutional layer, a convolutional layer, a parallelized patch-aware attention PPA, a convolutional layer, a parallelized patch-aware attention PPA, a convolutional layer, a parallelized patch-aware attention PPA, a convolutional layer, a C2f layer and a SPPF layer in sequence, and they start counting the layers from zero, where the low-level feature map feat1, the middle-level feature map feat2 and the high-level feature map feat3 come from the outputs of the 4th, 6th and 9th layers respectively.

[0014] Furthermore, the categories of the classroom behaviors are set to raising hands, reading, writing, using mobile phones, lowering heads, lying on tables, looking up, and standing.

[0015] Compared with the prior art, the present invention has the following beneficial effects:

[0016] 1. The present invention is based on an improved YOLOv8 detection model. By introducing parallelized patch-aware attention PPA into the Backbone network and adopting a multi-branch extraction strategy, the model's multi-scale feature selection capability is improved, so that the model can extract as much effective multi-scale feature information as possible. At the same time, a redesigned dimension-aware fusion diffusion pyramid network DAIDFPN is used for feature fusion, thereby realizing the model's preliminary multi-dimensional aggregation and fine multi-dimensional aggregation of high-dimensional features and low-dimensional features, so that each scale feature has rich context while reducing the number of model parameters, effectively solving the problems of dense small targets, occlusion and multi-scale spanning in natural classroom scenes, and improving the accuracy of student behavior detection and classification in classroom scenes.

[0017] 2. The dimension-aware fusion diffusion pyramid network DAIDFPN proposed in the present invention adopts the idea of ​​adaptive fusion-diffusion of high- and low-dimensional features, making full use of shallow and deep features. Compared with the traditional top-down and bottom-up PAN-FPN structure, it can reduce the loss of target feature information caused by up and down sampling in the feature fusion process, so that it has rich contextual information on features of different scales. At the same time, the number of model parameters is reduced, and the target detection accuracy is further improved.

[0018] 3. Compared with other methods, the classroom behavior detection method based on dimensional diffusion perception adaptive fusion proposed in the present invention does not need to rely on a large amount of data annotation data and computing resources. It has the characteristics of low computing overhead, strong real-time performance, and high detection accuracy. It has low cost in smart classroom deployment and has good practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a flow chart of the classroom behavior detection method based on dimensional diffusion perception adaptive fusion of the present invention;

[0020] Figure 2 It is the overall structure diagram of the improved YOLOv8 detection model of the classroom behavior detection method based on dimensional diffusion perception adaptive fusion of the present invention;

[0021] Figure 3 It is a structural diagram of the parallelized patch perception attention module PPA of the classroom behavior detection method based on dimensional diffusion perception adaptive fusion of the present invention;

[0022] Figure 4 It is a structural diagram of the DAIDFPN dimensional perception fusion diffusion pyramid network of the classroom behavior detection method based on dimensional diffusion perception adaptive fusion of the present invention;

[0023] Figure 5It is a simplified comparison diagram of the DAIDFPN structure of the classroom behavior detection method based on dimensional diffusion perception adaptive fusion of the present invention and the existing PAN-FPN structure;

[0024] Figure 6 It is a structural diagram of the DASI module of the classroom behavior detection method based on dimensional diffusion perception adaptive fusion of the present invention;

[0025] Figure 7 This is a diagram showing the effect of using the classroom behavior detection method based on dimensional diffusion perception adaptive fusion of the present invention for detection. DETAILED DESCRIPTION

[0026] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the following examples and drawings specifically illustrate a classroom behavior detection method based on dimensional diffusion perception adaptive fusion of the present invention. It should be noted that the description of these implementation methods is used to help understand the present invention, but does not constitute a limitation of the present invention.

[0027] like Figure 1 As shown, the present invention provides a classroom behavior detection method based on dimensional diffusion perception adaptive fusion, which replaces part of the C2f layer of the Backbone network in the YOLOv8 model with parallelized patch perception attention PPA, and constructs a dimensional perception fusion diffusion pyramid network DAIDFPN with a dimensional perception selective fusion module DASI to replace the Neck feature fusion network in the YOLOv8 model, thereby obtaining an improved YOLOv8 model, and then uses the improved YOLOv8 model to perform classroom behavior detection on classroom images. The details are as follows:

[0028] Step 1, obtaining a classroom surveillance video containing students' classroom behaviors, sampling useful frames in the video, and obtaining a classroom image containing students' classroom behaviors;

[0029] First, we obtained videos ranging from 10 to 20 minutes from more than 10 classroom scenes, and then randomly extracted corresponding frames from different videos as the images of the final data set (the images were not repeated). Finally, we obtained 1,140 classroom images.

[0030] Step 2: Preprocess the classroom images obtained in step 1, use the LabelImg annotation tool to annotate the students’ classroom behaviors on the preprocessed images, and then divide the sample dataset into a training dataset and a validation set according to a data ratio of 8:2 to construct a student classroom behavior dataset.

[0031] Common classroom behaviors are divided into 8 categories: raising hands, reading, writing, using mobile phones, looking down, lying on the table, looking up, and standing. In the end, we obtained 1,140 annotated classroom images, containing a total of 20,132 annotation tags.

[0032] Step 3: Build an improved YOLOv8 detection model

[0033] Step 3.1: Improve the Backbone network

[0034] The parallelized patch-aware attention module PPA is introduced into the 2nd, 4th and 6th layers of the Backbone backbone network to replace the original C2f layer. Combined with the multi-branch extraction strategy in the PPA module, the features of the students' classroom behavior images are extracted to obtain the corresponding feature maps. In this way, the layers of the entire Backbone backbone network are set as follows: a convolutional layer, a convolutional layer, a parallelized patch-aware attention PPA, a convolutional layer, a parallelized patch-aware attention PPA, a convolutional layer, a parallelized patch-aware attention PPA, a convolutional layer, a C2f layer and a SPPF layer. They start counting the number of layers from zero. The low-level feature map feat1, the mid-level feature map feat2 and the high-level feature map feat3 described below come from the outputs of the 4th, 6th and 9th layers respectively.

[0035] The specific process is as follows:

[0036] Step 3.1.1, input a classroom image of size 640×640, and perform multi-branch feature extraction through two Conv modules, one PPA module, one Conv module and one PPA module in sequence to obtain a low-level feature map feat1 of size 80×80;

[0037] Step 3.1.2, input the feature map feat1 obtained in step 3.1 into a Conv module for downsampling, and perform multi-branch feature extraction through the PPA module to obtain a mid-level feature map feat2 with a size of 40×40;

[0038] Step 3.1.3, input the feature map feat2 obtained in step 3.1 into a Conv module and a C2f module, and finally input it into the spatial pyramid pooling SPPF module to obtain a high-level feature map feat3 of size 20×20;

[0039] Among them, the overall structure of the PPA module is as follows Figure 3 As shown in the figure, the PPA module uses multiple parallel branches to process the input feature map. Each branch is responsible for capturing features of different scales. This branching strategy enables the module to focus on local details and global context at the same time, thereby better identifying and locating small objects.

[0040] Specifically, the PPA module divides the input feature map into three independent branches after a point-by-point convolution: local, global, and serial convolution branches. The local branch and the global branch are respectively implemented by controlling the size parameter p of the block patch. Here, we set the parameter p of the local branch and the global branch to 2 and 4, respectively, and the serial convolution branch is composed of three 3×3 convolutional layers.

[0041] The local branch and the global branch realize local and global feature extraction and interaction by aggregating and shifting non-overlapping control block patches and calculating the attention matrix between blocks. Specifically, the input feature map is first divided into a set of spatially continuous control blocks. Among them, H', W', and C' represent the height, width, and number of channels of the input feature map, respectively, and then the channel average is performed to produce p×p, Input to FFN to apply activation function to obtain the probability distribution of linear calculation features in spatial dimension, adjust their weights accordingly, and use feature selection to select task-related features from token and channel in the weighted result, and then calculate the corresponding feature map through these three branches respectively Feature Map Feature Map Then add these three results to get the final fused feature map Finally, after feature fusion is performed on multi-branch feature extraction, the attention mechanism is used to perform adaptive feature enhancement and output.

[0042] Step 3.2: Improve the Neck feature fusion network

[0043] The dimension-aware fusion diffusion pyramid network DAIDFPN is designed to replace the original PAN-FPN structure in the Neck part of the YOLOv8 model. The feature map extracted in step 3.1 is aggregated twice in multiple dimensions. The DAIDFPN network structure is as follows: Figure 4 As shown, Figure 5The structural difference between the DAIDFPN network and the PAN-FPN network is demonstrated. The dimension-aware fusion diffusion pyramid network DAIDFPN adopts a two-layer structure. The first layer uses the first-dimensional-aware selective fusion module DASI to perform preliminary multi-dimensional aggregation on the different-scale feature maps output by the Backbone network; the second layer uses the second-dimensional-aware selective fusion module DASI to perform fine multi-dimensional aggregation on the multi-dimensional feature maps output by the first layer and the different-scale feature maps output by the Backbone network, thereby obtaining small-, medium-, and large-scale target detection feature maps for detection processing in the Head detection network in the YOLOv8 model. In this way, after two link aggregations, the expressive power of the features can be effectively enhanced and the detection effect can be improved. Although simply aggregating features directly can reduce the amount of calculation, it will lose some recognition capabilities for small targets and complex backgrounds. Through two aggregations, the information loss caused by a single aggregation can be alleviated.

[0044] The specific implementation process of the DAIDFPN network is as follows:

[0045] Step 3.2.1, the feature maps feat1, feat2, and feat3 described in step 3.1 respectively contain low-level, middle-level, and high-level feature information of the input student classroom behavior image, and the feature maps of different levels are input into the first DASI module for dimension-aware adaptive feature fusion, thereby obtaining a feature map feat4 that fuses semantic information of different levels;

[0046] Step 3.2.2, upsample and downsample the feature map feat4 obtained in step 3.2.1 respectively, and obtain feature maps feat5 and feat6 respectively;

[0047] Step 3.2.3, concatenate the upsampled feature map feat5 described in step 3.2.2 and the feature map feat1 described in step 3.1 along the channel dimension using the Concat module, and then enhance and fuse the features through a C2f module to obtain the feature map feat7;

[0048] Step 3.2.4, concatenate the feature map feat6 after downsampling convolution in step 3.2.2 and the feature map feat3 in step 3.1 along the channel dimension using the Concat module, and then enhance and fuse the features through a C2f module to obtain the feature map feat8;

[0049] Step 3.2.5, inputting the high-level, mid-level, and low-level features of the feature maps feat7, feat4, and feat8 described in step 3.2.1 and steps 3.2.2-3.2.3 into the second DASI module again for adaptive feature fusion to obtain the fused feature map feat9;

[0050] Step 3.2.6, upsample the feature map feat9 obtained in step 3.2.5, and then input it into the Concat module along the channel dimension together with the feature map feat5 obtained in step 3.2.2, the feature map feat7 obtained in step 3.2.3, and the feature map feat1 obtained in step 3.1, and then enhance and fuse the features through a C2f module to obtain the feature map feat10;

[0051] Step 3.2.7, perform downsampling convolution on the feature map feat9 obtained in step 3.2.5, and then input it into the Concat module along the channel dimension with the feature map feat6 obtained in step 3.2.2, the feature map feat8 obtained in step 3.2.4, and the feature map feat3 obtained in step 3.1, and then enhance and fuse the features through a C2f module to obtain the feature map feat11;

[0052] In step 3.2.8, the feature maps feat10, feat9 and feat11 obtained in steps 3.2.5-3.2.6 are respectively input into the Head detection network as the feature maps for small-size, medium-size and large-size target detection.

[0053] Among them, the DASI module structure diagram is as follows Figure 6 As shown, the specific implementation process is as follows:

[0054] First, through operations such as convolution and interpolation, high-dimensional features are transformed and low-dimensional features With the current layer features After preliminary alignment, they are divided into four equal segments in the channel dimension, resulting in

[0055] where h i , l i ,u i Represents the i-th partition feature of high-dimensional, low-dimensional and current layer features. H, W, and C usually represent the three dimensions of the image.

[0056] H: Height, that is, the number of rows of the image;

[0057] W: Width, that is, the number of columns of the image;

[0058] C: Channels, that is, the number of color channels in the image, for example, an RGB image has 3 channels.

[0059] The partitions are calculated according to the following formula:

[0060] α=sigmoid(u i ),u i'=αl i +(1-α)h i (1)

[0061]

[0062] in, Indicates u i The value obtained after the sigmoid activation function; u i ' represents the result of selective aggregation on each partition. After merging the channel dimension, the feature map is obtained. Then after convolution Conv(), batch normalization And the corrected linear unit δ(), finally the output feature map is obtained

[0063] In step 3.3, the conventional method is used to build a Head detection network to locate the position of the students based on the feature map obtained in step 3.2, and to classify the student behavior; it is combined with the Backbone network obtained in step 3.1 and the Neck feature fusion network obtained in step 3.2 to jointly construct an improved YOLOv8 model.

[0064] Step 4, use the student classroom behavior dataset described in step 2 to train the improved YOLOv8 detection model, evaluate the training results, and save the optimal training weight file;

[0065] The training parameters of the model are set as follows: 300 rounds of iterative training, Batchsize is set to 16, the input image size is 640×640, the stochastic gradient descent (SGD) algorithm is used to optimize the model weight information, the learning rate is set to 0.01, the momentum parameter is 0.937, and the optimizer weight decay is 0.0005. After the training is completed, the model with the best effect is selected and saved.

[0066] Step 5, load the optimal training weight file described in step 4, input the classroom monitoring video or image into the trained model, and obtain the student classroom behavior detection results.

[0067] In order to further evaluate the effectiveness of the model improvement, we compared it with several of the most popular methods and used the same experimental environment for comparative experiments. The comparative experimental data set uses our self-built college student behavior detection data set. The evaluation indicators are model parameter quantity (Parameters), precision (P), recall rate (R), average precision (mAP50) under a confidence threshold (IOU) of 0.5, and average precision under different confidence thresholds (IOU) (mAP50-95). The experimental results of the models under different algorithms are shown in Table 1. The experimental results show that the improved model we proposed has the best performance under the same scale of parameters. Compared with the original YOLOv8 algorithm, the P value, R value, mAP50, and mAP50-95 increased by 2.5%, 1.2%, 1.9%, and 2.3%, respectively, and the number of parameters decreased by 2.3%.

[0068] Table 1 Performance comparison of different models

[0069]

[0070] The trained improved YOLOv8 model is used to detect students’ classroom behaviors. The model input can be classroom surveillance videos or students’ classroom behavior images, and the model outputs classroom behavior detection results. Figure 7 A set of image detection renderings obtained using this method are shown. It can be seen from the renderings that the proposed invention method can accurately detect students' classroom behavior under multi-target occlusion and crowded scenes, proving the effectiveness and practical application value of this method.

[0071] The above description is only a specific implementation of the present application, so that those skilled in the art can understand or implement the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest range consistent with the principles and novel features applied for herein.

Claims

1. A classroom behavior detection method based on dimensional diffusion perception adaptive fusion, characterized by: The parallelized patch-aware attention PPA is used to replace part of the C2f layer of the Backbone network in the YOLOv8 model. The dimension-aware selective fusion module DASI is used to construct the dimension-aware fusion diffusion pyramid network DAIDFPN to replace the Neck feature fusion network in the YOLOv8 model, thereby obtaining an improved YOLOv8 model. The improved YOLOv8 model is used to detect classroom behavior in classroom images. Among them, the dimension-aware fusion diffusion pyramid network DAIDFPN adopts a two-layer structure. The first layer uses the first dimension-aware selective fusion module DASI to perform preliminary multi-dimensional aggregation on the different-scale feature maps output by the Backbone backbone network; the second layer uses the second dimension-aware selective fusion module DASI to perform fine multi-dimensional aggregation on the multi-dimensional feature maps output by the first layer and the different-scale feature maps output by the Backbone backbone network, so as to obtain small-size, medium-size, and large-size feature maps for detection processing in the Head detection network in the YOLOv8 model.

2. The classroom behavior detection method based on dimensional diffusion perception adaptive fusion according to claim 1 is characterized by: The Backbone network outputs a low-level feature map feat1, a mid-level feature map feat2, and a high-level feature map feat3; For the first layer, the first dimension-aware selective fusion module DASI performs dimension-aware adaptive fusion on the low-level feature map feat1, the mid-level feature map feat2 and the high-level feature map feat3 to obtain the feature map feat4, and then upsamples and downsamples the feature map feat4 respectively to obtain the feature map feat5 and the feature map feat6 respectively, and then concatenates the feature map fe5 with the low-level feature map fe1, and then performs feature enhancement fusion through a C2f module to obtain the feature map feat7, and at the same time, concatenates the feature map fe6 with the high-level feature map fe3, and then performs feature enhancement fusion through a C2f module to obtain the feature map feat8.

3. The classroom behavior detection method based on dimensional diffusion perception adaptive fusion according to claim 2 is characterized by: For the second layer, the second dimension-aware selective fusion module DASI performs dimension-aware adaptive fusion on the feature map feat4, the feature map feat7 and the feature map feat8 to obtain the feature map feat9, that is, the medium-sized feature map feat9, and then upsamples the feature map feat9, and then splices it with the low-level feature map feat1 and the feature map feat5 and the feature map fe7, and then performs feature enhancement fusion through a C2f module to obtain the large-sized feature map feat10. At the same time, the feature map feat9 is downsampled, and then spliced ​​with the high-level feature map feat3 and the feature map feat6 and the feature map fe8, and then performs feature enhancement fusion through a C2f module to obtain the small-sized feature map fe11.

4. The classroom behavior detection method based on dimensional diffusion perception adaptive fusion according to claim 3 is characterized by: The parallelized patch-aware attention (PPA) is used to replace the C2f layers corresponding to the 2nd, 4th, and 6th layers of the Backbone network in the YOLOv8 model.

5. The classroom behavior detection method based on dimensional diffusion perception adaptive fusion according to claim 4 is characterized by: The layers of the Backbone network are set to be a convolutional layer, a convolutional layer, a parallelized patch-aware attention PPA, a convolutional layer, a parallelized patch-aware attention PPA, a convolutional layer, a parallelized patch-aware attention PPA, a convolutional layer, a C2f layer and a SPPF layer in sequence, and they start counting the layers from zero, where the low-level feature map feat1, the middle-level feature map feat2 and the high-level feature map feat3 come from the outputs of the 4th, 6th and 9th layers respectively.

6. The classroom behavior detection method based on dimensional diffusion perception adaptive fusion according to claim 1 is characterized by: The categories of classroom behaviors are set as raising hands, reading, writing, using mobile phones, looking down, lying on the table, looking up, and standing.

Images