Development and Application of a Deep Learning-Based Image Processing System for Classroom Behavior Analysis

With the continuous advancement of educational informatization, how to effectively leverage advanced technologies to enhance classroom teaching quality has become an important research topic [1-4]. In recent years, deep learning-based image processing technologies have achieved significant accomplishments across various fields, and their application in education has gradually gained attention [5-9]. By automating the analysis of classroom videos, student behaviors can be objectively recorded and assessed, enabling teachers to better understand teaching effectiveness and adjust teaching strategies in a timely manner. Consequently, the development of a deep learning-based image processing system for classroom behavior analysis holds considerable practical significance.

Classroom behavior analysis not only reflects student engagement and learning status but also provides data support for personalized teaching, thereby enabling precision in teaching [10-13]. Through the quantitative analysis of student behaviors during class, issues in the teaching process can be identified, which can subsequently inform the optimization of teaching design and improve the overall quality of education. Furthermore, classroom behavior analysis systems can be employed in teaching research, providing empirical data that contribute to the advancement of educational theory [14-16]. Therefore, research into deep learning-based image processing systems for classroom behavior analysis is not only of theoretical significance but also holds wide practical value.

Although some studies have attempted to apply deep learning to classroom behavior analysis, there remain several limitations [17-22]. For example, traditional behavior analysis methods largely rely on manual feature extraction, which fails to fully exploit the information contained in image data, leading to insufficient accuracy and robustness in the analysis results [23-25]. Within the theoretical framework of deep learning, Convolutional Neural Networks (CNNs) are the foundational architecture for visual feature extraction. Their hierarchical feature learning mechanism provides a crucial basis for the analysis of image sequences [26]. In the domain of temporal modeling, temporal models such as Recurrent Neural Networks (RNNs) and their variants, notably Long Short-Term Memory (LSTM) [27], have been widely employed to capture temporal dependencies via the transmission of hidden states. However, these models are often constrained by issues such as vanishing gradients and limited computational efficiency. Additionally, existing CNNs often overlook the temporal correlations when processing sequential data, which negatively affects the recognition performance of classroom behaviors. In recent years, artificial intelligence technologies-particularly image processing techniques driven by deep learning-have demonstrated growing potential within the field of education and have garnered increasing scholarly attention. CNNs have been employed to identify students’ postures in classroom environments, marking preliminary efforts toward the automation of behavioral analysis. Additional studies have focused respectively on video-based assessments of student attentiveness and analyses of classroom engagement, thereby underscoring the value of deep learning in interpreting classroom dynamics. Moreover, sequential models have been applied to capture patterns of teacher-student interaction. These investigations have collectively indicated that deep learning-based automated analysis of classroom behavior from image and video data offers promising pathways for enhancing the objectivity of teaching assessments, optimizing instructional strategies, and promoting personalized learning. Nevertheless, current methodologies still face significant challenges in fully extracting complex temporal information embedded in classroom behavior image sequences, effectively perceiving behavior variations across different time scales, and constructing robust recognition models capable of maintaining high generalizability. Therefore, there is an urgent need to develop a more effective deep learning model that can better capture and analyze classroom behaviors.

In response to the aforementioned issues, a deep learning-based image processing system for classroom behavior analysis was proposed in this study. The primary research content includes three key components: a) the development of a temporal 2D convolution model for classroom behavior analysis to fully exploit the temporal information embedded in image data; b) the design of a method to expand the receptive field of the temporal 2D convolution, enhancing the model's ability to perceive behaviors at different time scales; c) the construction of a classroom behavior recognition network to improve the accuracy and robustness of behavior recognition. Through these efforts, this research aims to provide an efficient and accurate solution for classroom behavior analysis and contribute to the advancement of educational informatization.

When processing classroom behavior video images, the temporal perception range of the feature maps generated by the temporal 2D convolution must be considered. By stacking a sufficient number of temporal 2D convolutional layers, a broader temporal perception range and richer semantic features can, in theory, be obtained. However, the depth of the neural network cannot be increased indefinitely, as both computational cost and training difficulty rise significantly with the increase in layers. To effectively explore and optimize this process, a feedforward neural network was discussed in this study using the nonlinear unit as the unit. This approach aids in a clearer analysis and understanding of the role and contribution of the temporal 2D convolution at specific layers. Let the neural network D(a;r) consist of M stacked nonlinear units. Assuming that the computation of the neural network is denoted as G and the activation function is represented by h, the net input c and output x of the m-th layer of the neural network can be expressed by the following equation:

$\begin{aligned} & c^{(m)}=G\left(x^{(m-1)}\right) \\ & x^{(m)}=h\left(x^{(m)}\right)\end{aligned}$         (9)

To harness the advantages of deep networks, it is necessary to address issues such as overfitting, gradient vanishing, and gradient explosion effectively. The classroom behavior analysis model utilized in this study expands the perception range through multiple layers of the temporal 2D convolution, capturing the temporal dynamics within the video data to identify and analyze students' behaviors. To prevent overfitting, several regularization techniques were employed. These include L2 regularization, dropout, and data augmentation methods. L2 regularization involves adding a weight penalty term to the loss function to suppress excessively large network parameters, thereby reducing the risk of overfitting. Dropout works by randomly deactivating a portion of neurons during training, enhancing the model's robustness and preventing dependence on specific nodes. Additionally, data augmentation generates more diverse training samples through operations such as rotation, translation, and scaling, further enhancing the model's generalization capability. The combined use of these methods helps the deep neural network to better learn features during training, rather than memorizing the training data.

To address the issues of gradient vanishing and explosion, several advanced techniques and optimization strategies were introduced in this study. For example, batch normalization was employed to standardize the input of each batch of data, ensuring that the input distribution of each layer remains stable. This technique accelerates training convergence and reduces the risk of gradient vanishing. Additionally, appropriate weight initialization methods were adopted to ensure that the initial weights are within a suitable range, thereby preventing the gradual disappearance or explosion of gradients during both forward and backward propagation. Furthermore, the selection of suitable activation functions, such as ReLU and its variants, can effectively mitigate the gradient vanishing problem and improve training efficiency. Specifically, the parameter update at the M-th layer requires the computation of the gradient of the loss Z with respect to the layer. This gradient depends on the error term of the layer, denoted as σ=∂γ/∂c^(m). According to the chain rule, σ^(m) is related to the error term σ^(m+1) of the subsequent layer:

$\sigma^{(m)}=\frac{\partial c^{(m+1)}}{\partial c^{(m)}} \cdot \sigma^{(m+1)}$         (10)

Let ε^(m)∂c^(m+1)/∂c^(m), then it leads to the following expression:

$\sigma^{(m)}=\varepsilon^{(m)} \sigma^{(m+1)}$         (11)

Assuming that the output and input dimensions of the neural network are consistent, the function G that needs to be fitted can be divided into two parts as follows:

$c^{\left(m_2\right)}=G\left(x^{(m-1)}\right)=x^{(m-1)}+D\left(x^{(m-1)}\right)$         (12)

For computational convenience, if no activation functions are applied, the following expression can be obtained:

$\begin{aligned} x^{\left(m_2\right)} & =x^{\left(m_2-1\right)}+D\left(x^{\left(m_2-1\right)}\right) \\ & =x^{\left(m_2-2\right)}+D\left(x^{\left(m_2-2\right)}\right)+D\left(x^{\left(m_2-1\right)}\right)\end{aligned}$          (13)

By recursively expanding this expression, the final result can be derived as:

$x^{\left(m_2\right)}=x^{\left(m_1\right)}+\sum_{u=m_1}^{m_2-1} D\left(x^{(u)}\right)$         (14)

The gradient of the final loss γ with respect to a lower-layer output can be expanded as:

$\begin{aligned} & \frac{\partial \gamma}{\partial x^{\left(m_1\right)}}=\frac{\partial \gamma}{\partial x^{\left(m_2\right)}} \frac{\partial x^{\left(m_2\right)}}{\partial x^{\left(m_1\right)}} \\ & =\frac{\partial \gamma}{\partial x^{\left(m_2\right)}}\left(1+\frac{\partial x^{\left(m_2\right)}}{\partial x^{\left(m_1\right)}} \sum_{u=m_1}^{m_2-1} D\left(x^{(u)}\right)\right)\end{aligned}$          (15)

The choice of optimization algorithm and hyperparameter tuning are also crucial for the training of deep neural networks. The Adam optimizer was employed in this study, which adjusts the update step size of each layer's parameters through an adaptive learning rate. The incorporation of momentum terms helps to reduce oscillations, further stabilizing the training process.

As previously described, the standard T2D-Conv module captures sequential frame information through 1D convolutional kernels. However, classroom behaviors exhibit pronounced multi-scale temporal characteristics: certain actions occur over brief durations, whereas others span significantly longer periods. To enhance the model’s ability to perceive long-duration behavioral patterns-while avoiding the computational burden and optimization challenges associated with excessively deep convolutional stacks-a dilated convolution mechanism was introduced into the temporal 1D convolutional component, forming the D-T2D-Conv module.

The primary advantage of dilated convolution lies in its ability to exponentially expand the receptive field. By simply increasing the dilation rate, D-T2D-Conv significantly extends the model’s temporal receptive field without increasing the number of convolutional parameters or the network depth. This enhancement markedly improves the model’s capability to capture long-range dependencies, thereby enabling the effective recognition of classroom behaviors that span dozens or even hundreds of frames. In contrast, achieving an equivalent receptive field with standard convolution would necessitate deep layer stacking, substantially increasing model complexity and training difficulty.

In the system design, multiple groups of D-T2D-Conv modules with different dilation rates were applied across various network hierarchies, forming a multi-scale temporal feature extraction pyramid. Lower layers employed smaller dilation rates to capture fine-grained, short-term actions, while upper layers utilized larger dilation rates to model coarse-grained, long-duration behavioral patterns. This architectural strategy constitutes a core technique for enhancing the model’s perceptual capacity across diverse temporal scales in classroom behavior recognition.

As shown in Table 1, the classroom behavior recognition network architecture describes a multi-level deep learning model specifically designed for the analysis and recognition of behaviors in classroom environments. Each input image to the network contains multiple temporal frames. A series of different convolutional layers, combined with the temporal offset dilation structure, was used to progressively extract features from low-level to high-level. In the initial stages, multiple standard convolutional layers were used for spatial feature extraction, with the output channel count increasing, while convolution operations with a stride of 1 were employed to maintain spatial resolution while ensuring feature dimension consistency. As the network deepened, the temporal offset dilation convolution structure was introduced to enhance the ability to perceive temporal information. Particularly, through the use of different dilation rates and repetition counts, the network was able to model and extract temporal features across various time scales, effectively capturing the dynamic changes in classroom behaviors.

Table 1. Classroom behavior recognition network architecture

Input	Network Type	Number of Output Channels	Repetition Count of the Layer	Stride
11*215*215*3	Standard 3×3 convolution	31	1	2
11*125*125*31	Standard 3×3 convolution	15	1	1
11*125*125*31	Standard 3×3 convolution	23	1	2
11*55*55*23	Temporal offset dilation structure	23	3	1
11*55*55*23	Standard 1×1 convolution	31	1	2
11*27*27*31	Temporal offset dilation structure	31	2	1
11*27*27*31	Standard 1×1 convolution	62	1	2
11*12*12*62	Temporal offset dilation structure	62	3	1
11*12*12*62	Standard 1×1 convolution	95	1	1
11*12*12*94	Temporal offset dilation structure	95	3	1
11*12*12*94	Standard 1×1 convolution	158	1	2
11*7*7*158	Temporal offset dilation structure	158	1	1
11*7*7*335	Standard 1×1 convolution	1125	1	1
11*7*7*1128	Average pooling	-	1	-
11*1*1*1128	Fully connected	4	1	-

In the experimental section, the first task was the prediction and detection of classroom behavior boundaries, with the goal of distinguishing between intervals containing classroom behaviors and those without. Figure 4 illustrates the conceptual approach for boundary prediction, which effectively segments long videos into multiple segments. Each segment was marked with the action intervals of both students and teachers, allowing the model to focus on analyzing only those parts of the video that contain classroom behavior. This segmentation improves data processing efficiency and model accuracy. This step not only reduces interference from irrelevant data but also provides more precise input data for the subsequent behavior recognition network. The segments selected through this method serve as the input for the classroom behavior recognition network, enabling further detailed action recognition.

4a.png

4b.png

Figure 4. Prediction of classroom behavior boundaries to identify the student-teacher action intervals

Table 2. Results of classroom behavior recognition testing

Dataset	Correct Action Accuracy	Incorrect Action Accuracy	False Positive Rate	False Negative Rate
Training dataset	95%	91%	0%	0%
Testing dataset	87%	85%	22%	12%

Based on the classroom behavior recognition test results presented in Table 2, it can be observed that the system's performance differs between the training and testing datasets. On the training dataset, the model achieved high accuracy, with a correct action accuracy of 95%, an incorrect action accuracy of 91%, and both false positive and false negative rates of 0%, indicating that the model was able to effectively identify classroom behaviors during training without any misclassifications or omissions. In contrast, the results for the testing dataset were slightly lower, with a correct action accuracy of 87%, an incorrect action accuracy of 85%, a false positive rate of 22%, and a false negative rate of 12%. This suggests that, while the model's behavior recognition ability remains highly accurate in practical applications, the increased diversity and complexity of the testing data led to a higher occurrence of false positives and false negatives.

Table 3. Performance comparison of different models

Network Model	Backbone Network	Number of Sample Frames	Weight	mAP
Classroom behavior recognition network	Temporal expansion network	11	8.8 M	44.6
Two-stream networks	VGG	8	46.2 M	38.9
SlowFast networks	ResNet	31	34.5 M	42.3
Temporal segment networks	ResNet	8	32.6 M	37.8
Inflated 3D ConvNet	Inception-v1	8	11.2 M	18.6

As shown in Table 3, the proposed classroom behavior recognition network demonstrates excellent performance in terms of mean Average Precision (mAP), achieving a value of 44.6, which outperforms several other mainstream network models. The model employs a temporal expansion network as the backbone and was trained with a sample size of 11 frames, with a weight parameter of 8.8 M. In comparison, the two-stream networks, which use the Visual Geometry Group (VGG) backbone, have a higher weight (46.2 M) but achieve an mAP of 38.9, which is slightly lower than the proposed network. SlowFast networks, based on the ResNet backbone and with 31 sample frames, achieve an mAP of 42.3, indicating good performance in spatiotemporal feature extraction, although it does not surpass the proposed network. Additionally, temporal segment networks and inflated 3D ConvNet, which use the ResNet and Inception-v1 backbones, respectively, show weaker performance with mAP values of 37.8 and 18.6. The comparison results clearly show that the classroom behavior recognition network proposed in this study performs superiorly across several aspects, particularly in accurately identifying classroom behaviors. While other models, such as two-stream networks and SlowFast networks, also perform well in processing temporal and spatial features, their more complex network structures and higher parameter weights limit their computational efficiency and generalization ability. The temporal expansion network designed in this study, with fewer parameters (8.8 M) and a reasonable frame count (11 frames), achieves a balance between accuracy, computational efficiency, and robustness.

Table 4. Comparison of processing speeds of different models

Network Model	FLOPs	Weight	Processing Time
Inflated 3D ConvNet	315 G	34.6 M	156.2 ms
Two-Stream Networks	63 G	46.2 M	31.2 ms
SlowFast Networks	32 G	28.5 M	24.3 ms
Classroom behavior recognition network	16 G	8.8 M	12.4 ms

As shown in Table 4, the classroom behavior recognition network proposed in this study demonstrates a significant advantage in processing speed. The model has a floating-point operations (FLOPs) count of 16 G, with a weight of 8.8 M and a processing time of only 12.4 ms, indicating high computational efficiency. In contrast, although the inflated 3D ConvNet and two-stream networks perform well in terms of recognition accuracy, their processing times are 156.2 ms and 31.2 ms, respectively. Furthermore, the former has a FLOPs count of 315 G, and the latter 63 G, suggesting that these models incur higher computational costs. Even the SlowFast Networks, with a FLOPs count of 32 G, has a processing time of 24.3 ms, which still does not match the performance of the proposed network.

To provide a more intuitive and in-depth understanding of system performance beyond quantitative metrics, a detailed qualitative analysis was also conducted. A comprehensive review of numerous video samples was undertaken to summarize representative cases of successful recognition and to examine challenging scenarios that led to misclassification, thereby revealing both the strengths and current limitations of the system. In the successfully recognized cases, the system demonstrated robust capability in capturing a variety of typical classroom behaviors. For instance, in the recognition of the “raising hand to ask a question” behavior, the model was able to accurately track the full sequence of motion-from the initial lift of the arm, through the maintained raised-hand posture, to the eventual lowering of the hand-while consistently producing high-confidence outputs. This outcome substantiates the effectiveness of the T2D-Conv model in modeling the dynamic evolution of continuous actions. In the detection of the “taking notes” behavior, the recognition results remained stable even when students momentarily looked up at the blackboard, thereby interrupting the writing motion. This stability was attributed to the moderate window-overlap strategy and the model’s temporal contextual memory, which allowed temporal segments to be associated across time, effectively preventing misclassification due to short-term interruptions. For the “attentive listening” state, high-confidence results were continuously produced when students maintained a standard seated posture facing the lectern. Even in the presence of minor background disturbances, the AFG mechanism was presumed to have effectively suppressed noise activation in non-critical regions, maintaining focus on the target student. In addition, for long-duration “inattentive” behaviors, the use of high-level D-T2D-Conv modules with large dilation rates enabled the successful capture of these prolonged deviation patterns spanning dozens of frames.

5.png

Figure 5. Testing results on the actual classroom video dataset

Based on the data presented in Figure 5, which shows the variation in average recall with respect to the average number of predicted results per video, differences in performance at varying overlap rates can be observed. When the overlap rate is 0.45, the average recall increases gradually from 0.52 to 0.91, with an Area Under the Curve (AUC) of 67.26. As the number of predicted results increases, the average recall steadily rises, indicating that the model performs well in capturing actual behavioral segments at lower overlap rates. When the overlap rate is 0.55, the initial average recall is 0.48, which eventually reaches 0.87, with an AUC of 54.59. As the overlap rate increases to 0.65, the initial recall decreases to 0.44, and it eventually reaches 0.83, with an AUC of 50.61. Further increasing the overlap rate to 0.75 results in an initial recall of 0.34, which rises to 0.73, with an AUC of only 42.33. At an overlap rate of 0.85, the initial recall remains 0.34, but the final recall reaches only 0.73, with an AUC of 24.61. In comparison, the AR-AN method shows a gradual increase in average recall from 0.24 to 0.47, with an AUC of 43.15, overall underperforming. From the experimental results, it can be observed that the proposed deep learning-based classroom behavior analysis system demonstrates a superior recall at lower overlap rates. The average recall steadily improves as the number of predicted results increases, suggesting that the model can accurately capture time segments that overlap with actual behavior segments. However, as the overlap rate increases, the recall performance of the system significantly decreases, and the AUC value also drops sharply, reflecting the model's difficulty in maintaining high accuracy at higher overlap rates. Particularly at an overlap rate of 0.85, despite the increase in the number of predicted results, the final recall shows only limited improvement, with the AUC dropping to 24.61, indicating a significant decline in the model's recognition performance in this scenario.

To assess the system's sensitivity to behavior boundary delineation, performance was evaluated under different sliding window overlap ratios during input video segmentation for prediction generation. The experimental results demonstrated that as the overlap ratio increased from 0% to 50%, the overall recognition accuracy initially improved and then plateaued. Notably, optimal or near-optimal performance was typically achieved when the overlap ratio ranged between 25% and 33%. The performance variation under different overlap settings can be attributed to the following factors: a) Continuity of behaviors and boundary ambiguity: Classroom behaviors generally evolve continuously, with indistinct onset and offset boundaries. When the window overlap is too low, a severe discontinuity is introduced between adjacent segments. As a result, behavioral sequences may be split at the window boundaries, causing the model to observe only partial behavior segments. This fragmentation leads to incomplete feature representations and increases the likelihood of misclassification or missed detections. b) Temporal modeling and contextual information: The proposed D-T2D-Conv module relies on adequate contextual frames to effectively model temporal dynamics. A moderate overlap ratio ensures that a substantial number of frames are shared across adjacent windows. Consequently, even when a behavior spans across window boundaries, a relatively complete contextual representation is preserved within multiple overlapping windows. More comprehensive information from preceding and succeeding frames is utilized when predicting behaviors near the center of each sliding window, thereby enhancing the accuracy of assessing the complete behavioral process and mitigating the adverse effects of boundary segmentation. c) Computational redundancy and noise amplification: While high overlap ratios theoretically minimize boundary segmentation issues, they also introduce significant redundancy due to repeated frames across adjacent windows. This redundancy leads to increased computational cost from repeated feature extraction and inference across highly similar windows, thereby reducing processing efficiency. Moreover, the high degree of content similarity among overlapping windows may amplify transient or local noise or mispredictions during result fusion, rather than providing diverse and complementary information. In some cases, this effect may even slightly degrade overall performance or result in diminishing returns.

Considering both recognition accuracy and computational efficiency, an overlap ratio of 25% to 33% is recommended. This configuration strikes an effective balance between minimizing behavior boundary fragmentation, ensuring sufficient contextual representation, and controlling computational redundancy. This parameter serves as a critical factor in maximizing the performance of the proposed multi-scale temporal modeling network. Furthermore, the findings reinforce the importance of effectively leveraging temporal context information to enhance the robustness of behavior recognition.

Table 5. Comparison of testing results on the actual classroom video dataset with other mainstream models

Based on the comparison of the test results from the actual classroom video dataset with other mainstream models provided in Table 5, the proposed classroom behavior recognition network demonstrates superior performance across multiple evaluation metrics. The model achieves a mean intersection over union (mIoU) of 83.21, significantly outperforming other leading network models, such as the two-stream networks (62.26) and SlowFast networks (72.15). On the mAP@0.5IoU metric, the proposed network also leads all comparison models with a score of 96.36, especially surpassing the inflated 3D ConvNet (57.54). The model continues to maintain a leading position in mAP@0.7IoU with a score of 81.26, well above SlowFast networks (61.25) and temporal segment networks (64.58). Additionally, in terms of frames per second (FPS), the processing speed of the proposed classroom behavior recognition network is comparable to that of the two-stream networks, both achieving 75 FPS, indicating high real-time performance suitable for real-time classroom behavior analysis applications. The comparison results clearly indicate that the proposed network excels in both accuracy and efficiency. Particularly in the mIoU, mAP@0.5IoU, and mAP@0.7IoU metrics, the model significantly outperforms other mainstream models, demonstrating a strong ability to accurately recognize classroom behaviors and more precisely segment and identify behavior-related time intervals. Furthermore, the network’s real-time performance is excellent, with a processing speed of 75 FPS, matching that of the two-stream networks, ensuring efficient real-time feedback capability. In contrast, the inflated 3D ConvNet performs poorly across multiple metrics, particularly in mAP@0.7IoU, where it scores only 9.36, highlighting its deficiencies in both accuracy and robustness. Overall, the proposed network not only offers significant advantages in recognition accuracy but also excels in computational efficiency, showing greater potential for practical applications, especially in meeting the high efficiency and real-time requirements of actual classroom behavior analysis. Figure 6 provides an example of classroom behavior tracking and judgment.

6.png

Figure 6. Example of classroom behavior tracking and judgment

To evaluate the temporal and spatial computational efficiency, a comparative experiment was conducted using 3D convolution as the baseline method on the ClassroomAction dataset. As presented in Table 6, the T2D-Conv module achieved a 68.3% reduction in FLOPs, a 52.7% decrease in the number of model parameters, and a 41.2% improvement in per-frame inference time relative to the standard 3D convolution. These gains can be attributed to the structural optimization offered by the T2D-Conv, wherein the spatial and temporal convolution operations are decoupled, thereby preserving the temporal feature modeling capacity while significantly enhancing computational efficiency. Such lightweight architectural design is particularly advantageous for real-time analysis in classroom scenarios, as it reduces the computational burden on edge devices and facilitates practical system deployment.

Table 6. Comparative evaluation of temporal and spatial computational complexity

To further validate the effectiveness of the AFG mechanism, a pair of comparative experiments was designed, consisting of Model A with the mechanism enabled and Model B without it. On the standard test set of the ClassroomAction dataset, Model A achieved an average recognition accuracy of 89.7%, representing a 4.5 percentage point improvement over Model B, which yielded an accuracy of 85.2%, as detailed in Table 7. Under robustness evaluation conditions with 10% Gaussian noise added to the input images, the performance of Model A decreased to 82.3%, while Model B dropped more sharply to 75.1%. These results demonstrate the AFG mechanism’s capacity to suppress noise interference effectively. By dynamically adjusting the weight distribution across spatiotemporal features, the mechanism enables more precise focus on salient behavioral cues while mitigating the impact of background noise and irrelevant information. Consequently, the mechanism enhances the practical applicability of the model in real-world classroom environments.

Table 7. Comparative results for the AFG mechanism