3D Image Modeling and Visual Presentation Technologies for Education

2.1 Overall network architecture

In the field of education, especially when it involves the design of intelligent facilities and gerontechnology education, a deep understanding of complex structures and processes often requires students to observe continuous changes and microscopic details. This means that 3D images for educational purposes not only need to have high clarity to ensure all details are clearly presented but should also be able to show these details' continuous changes in real environments, allowing students to understand the entire process of change smoothly, like watching a movie. This is particularly important in teaching smart home design, as students need to understand how devices operate continuously under different conditions and how they adapt to the specific needs of elderly users. However, due to cost and technological limitations, especially in the field of gerontechnology, it is often difficult to directly obtain a sufficient number of continuous images to meet this need. For example, when teaching how to design smart devices suitable for the elderly, continuous dynamic presentations can help students understand how different devices and user interfaces can adaptively adjust as the user's behavior and health status change.

For this reason, the 3D educational image modeling method based on inter-layer feature learning proposed in this paper uses inter-layer interpolation technology, as shown in Figure 1. Inter-layer interpolation generates intermediate images between two existing adjacent images, not only greatly enriching the image sequence but also enhancing the learning experience of students, helping them understand complex 3D structures and dynamic changes in a more continuous and intuitive way. Specifically, this method first analyzes two adjacent 3D images input, extracting the features and deformation trends of the images, and then estimates the transition features between layers through the generative model obtained by training. In this way, the generated intermediate images maintain a high degree of visual consistency with the original images and can reveal more subtle detail changes that are difficult to observe. Assuming the generation of inter-layer images is represented by U_MID, the formula is as follows:

$U_{M \mathbb{D}}=d\left(U_0, U_2\right)$                (1)

11.png

Figure 1. Principle of 3D educational image modeling method based on inter-layer feature learning

The appropriate function d for generating inter-layer images is crucial for achieving high-quality 3D educational image layer interpolation. For this challenge, this paper proposes an inter-layer interpolation network for 3D educational images. This network consists of three main parts: 1) The first part is preprocessing and feature extraction, that is, using a three-layer convolutional network to process the two input 3D educational images, aiming to extract the basic features and texture information of the images. These features provide the necessary raw information for the subsequent network's inter-layer generation task, ensuring that the generated images maintain visual coherence with the original images and retain the key visual information of the educational content. 2) The second part is the interleaved residual generation network. After feature extraction, the extracted features are fed into an interleaved residual network. An intermediate result is generated by learning the residuals between the input images. This intermediate result represents the predicted changes between layers, enhancing the original data. 3) The third part is the inter-layer information fusion module. By integrating the initially extracted features and the intermediate generation results, the effective information between image layers is strengthened, and the overall visual effect of the images is enhanced, ensuring that the generated inter-layer images are more accurate in anatomical structure and functional presentation, thus providing better visual support for understanding complex 3D structures in the field of education. Specifically, assuming the intermediate result output by the interleaved residual generation network is represented by U'_MID, the image features extracted by the three convolutional layers are represented by a, and the interleaved residual generation network is represented by Y(.), the formula is:

$U_{M D}{ }^{\prime}=Y(a)$               (2)

Assuming the inter-layer information fusion module is represented by D(.), using D(.) to integrate U'_MID can generate the final inter-layer image, represented by U_MID, as follows:

$\begin{gathered}U_{M I D}=d\left(U_0, U_2\right)=D\left(Y(x), U_0, U_2\right)  =D\left(U_{M I D}^{\prime}, U_0, U_2\right)\end{gathered}$               (3)

2.2 Residual and inter-layer information fusion

In the field of modeling and visual presentation of educational 3D images, the application of deep neural networks can greatly enrich the feature expression of images, improving the accuracy of modeling and the realism of visual effects. However, simply increasing the depth of neural networks often encounters problems of gradient vanishing or exploding, leading to poor training effects, a phenomenon known as network degradation. This is particularly prominent for complex educational 3D images, as these images often require deeper networks to extract more subtle structural information. To enhance network performance and overcome the degradation problem, this paper designs a new network architecture that can effectively extract rich hierarchical features while avoiding network degradation. The necessity of this architecture lies in its ability to allow the network to deeply learn the complex structures of 3D images while maintaining training stability, thus better serving educational purposes, such as providing more accurate and vivid visual teaching materials. Specifically, an interleaved residual generation network structure is adopted, utilizing skip connections to achieve identity mapping, allowing gradients to directly flow through some layers. These skip connections help alleviate the problem of gradient vanishing because they provide a shortcut for gradients, maintaining good information and gradient flow even in very deep networks. Moreover, through carefully designed interleaving patterns, the information transfer between lower and higher layers is optimized, achieving effective propagation of information in deep networks.

Assuming the input of the residual network block is represented by u, the number of layers by m, and the stacked network layers by G(u), the formula is as follows:

$u_m=G\left(u_{m-1}\right)+u_{m-1}$                 (4)

The backbone network block for processing image feature information is set with five convolution layers Z0, Z1, Z2, Z3, Z4. To fully utilize feature information and better preserve image details, convolution layers use small kernels of size 3*3, with three interleaved skip connections T0, T1, T2. Based on the above formula, the inputs for Z2 and Z4 can be derived:

$u_2=G\left(u_1\right)+u_0$                 (5)

$u_4=u_0+G\left(u_1\right)+G\left(u_3\right)$                 (6)

This paper constructs a deep interleaved residual generation network composed of 10 network blocks. This network architecture uniquely introduces skip connections within and between network blocks, which not only enhances feature reuse but also preserves key information of the image and effectively mitigates the degradation problem during network training. Specifically, skip connections are set from block 0 to block 9, from block 0 to block 5, from block 2 to block 7, and from block 5 to block 9. Such a tiered design allows feature information to be effectively transmitted and fused between layers of different depths, helping to capture both the detailed information and overall contour of the image simultaneously. In the actual construction of this network structure, special attention was paid to how to deal with the dimension matching problem when concatenating feature maps of different dimensions in skip connections. For this purpose, a special convolution layer is set at each concatenation point of the skip connection to adjust and match the dimensions of the feature maps output by different network blocks. This design ensures the seamless fusion of features from different levels, optimizing information flow, and providing a more refined feature representation for the task of inter-layer interpolation of 3D images.

In the current state of application of 3D educational image modeling and visual presentation, traditional methods of inter-layer interpolation often rely on direct pixel-level comparison between adjacent images in time or space dimensions. This approach tends to produce inaccurate motion estimation when dealing with images that have complex movements or structural changes. This is particularly disadvantageous in the field of education, where teaching content often requires precise and coherent visual representation so that students can clearly understand and master knowledge. Therefore, to provide a more stable and coherent visual experience in continuous image sequences, this paper proposes an inter-layer information fusion module, as shown in Figure 2. This module specifically processes the inter-layer structural information and change patterns output by the interleaved residual generation network that generates the intermediate result U_MID. The core of the module is composed of ConvGRU units, combining the capabilities of Convolutional Neural Networks (CNNs) in extracting spatial features from images with the advantages of Gated Recurrent Units (GRUs) in processing sequential data. Such a design allows the network to better simulate continuous motion and changes in the real world when processing educational 3D images, providing a highly realistic and coherent visual environment for teaching.

2.png

Figure 2. Structure of the inter-layer information fusion module

Assuming the input hidden state is represented by g, the activation function by δ, the current input to the ConvGRU unit by t_s, and the network-learned weight matrices by Q and I. The ConvGRU unit internally has a reset gate E and an update gate C, with the formulas as follows:

$E=\delta\left(Q_E t_s+I g_{s-1}\right)$             (7)

$C=\delta\left(Q_C t_s+I g_{s-1}\right)$             (8)

Assuming the current input to the ConvGRU unit is represented by U_MID, the previous hidden state by g_s-1, and the intermediate quantity that includes the current state U'_MID by g'. E is used to selectively fuse the inter-layer deformation features implied in the two consecutive input images U₀ and U₂ onto U_MID, that is, adding the information of g_s-1 to t_s, with the calculation formula as follows:

$g_s=(1-C) g_{s-1}+C g^{\prime}$             (9)

2.3 Weighted fusion loss function

Considering the roles and importance of the interleaved residual generation network and the inter-layer information fusion module during the generation process, this paper opts to introduce a weighted fusion loss function. This allows for adjusting the influence of these two modules on the final output according to different educational content and visual requirements, aiming to produce superior 3D visual effects.

Inter-layer interpolation of educational 3D images requires pixel-level generation, and both local loss functions use the L2 loss function, with the calculation formula as follows:

$M_2(H, O)=\sum_{u=1}^v\left(H_u-O_u\right)^2$             (10)

Let the loss functions corresponding to the interleaved residual generation network and the inter-layer information fusion module be represented by M_Y and M_D, respectively. The number of image pixels is represented by v, the ground truth by H, and the generated image by O. Thus, the fusion loss function M is defined as:

$M=\beta M_Y+\alpha M_D$             (11)

An analysis of the comparison of inter-layer information fusion results under different weights as shown in Figure 6 indicates that using 3D educational image modeling technology based on inter-layer feature learning, in conjunction with the inter-layer information module network and interleaved residual generation network, can effectively enhance image quality. From the Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) metrics, it is observed that as the first weight in the weight arrangement increases from 0.1 to 0.5, PSNR remains relatively stable while SSIM slightly decreases. This suggests that the overall brightness and contrast of the image have been optimized to some extent, but there is a slight loss in structural fidelity. As the first weight in the weight arrangement increases from 0.6 to 0.9, both PSNR and SSIM show a downward trend, indicating that within this weight range, the model's ability to retain image details weakens, possibly due to an overemphasis on the first feature. A comprehensive analysis of experimental data shows that when the weights of the two features are relatively balanced (e.g., a weight arrangement of (0.5, 0.5)), the model can better balance overall brightness/contrast and structural fidelity, resulting in a more balanced outcome. However, when the weight is biased towards the features of the inter-layer information module network, the image's detail representation capability declines. This demonstrates that the weighted fusion loss function proposed in this paper can better adapt to the complexity and diversity of educational content under specific weight distributions, especially when the weights of the two features are relatively balanced. Therefore, the 3D educational image modeling method based on inter-layer feature learning proposed in this paper, by adjusting the weight distribution appropriately, can effectively enhance the quality of 3D educational images, especially in achieving a good balance between maintaining image structural fidelity and overall brightness/contrast.

6.png

Figure 6. Comparison of inter-layer information fusion results under different weights

Based on the ablation study comparison results presented in Table 1, the effectiveness of the proposed 3D educational image modeling method can be comprehensively assessed. The model discussed in this paper shows the best performance across three key evaluation metrics: PSNR (Peak Signal-to-Noise Ratio), SSIM (Structural Similarity Index), and RMSE (Root Mean Square Error), achieving 32.5614, 0.9154, and 0.0465, respectively. By sequentially removing key components for ablation experiments, the impact of each component on model performance can be observed. After removing the inter-layer information module network, the model's PSNR and SSIM decreased, and RMSE increased, indicating the critical role of the inter-layer information module network in enhancing image quality. Further removal of the interleaved residual generation network led to a further decline in all metrics' performance, especially in PSNR and RMSE, emphasizing the importance of this network in improving image details and reducing reconstruction errors. Finally, when the weighted fusion loss function is not used, the model's performance drops again, particularly in the SSIM metric, highlighting the loss function's crucial role in optimizing model structure fidelity. Integrating these experimental results, it can be concluded that the 3D educational image modeling method based on inter-layer feature learning proposed in this paper achieved excellent results across the three main performance metrics. This achievement is attributed to the collaborative effect of the inter-layer information module network, interleaved residual generation network, and weighted fusion loss function. The inter-layer information module network optimized feature extraction and fusion, the interleaved residual generation network enhanced the model's detail capturing ability, and the weighted fusion loss function balanced the importance of different features during model training, making the model better adapted to the complexity and diversity of educational content. Therefore, the integration of these components significantly improved the modeling quality of 3D educational images, validating the effectiveness of the proposed method.

Table 1. Ablation study comparison results of 3D educational image modeling methods

Table 2. Performance comparison results of different 3D educational image modeling methods

From the performance comparison results of 3D educational image modeling methods in Table 2, the proposed model outperforms other comparative methods in three key performance indicators: PSNR, SSIM, and root mean square error (RMSE). Compared with current popular 3D image modeling methods such as VoxNet, ProGANs, DAEs, ConvLSTM, PointNet++, and DGCNN, the proposed model achieves the highest PSNR at 32.6985, reaches 0.9365 in SSIM, and lowers RMSE to 0.0412. Especially when compared with DGCNN, which has the closest SSIM value, the proposed model shows significant improvement in both PSNR and RMSE. These results demonstrate that the proposed model performs excellently in terms of reconstruction quality, image structural similarity, and error minimization. It reflects the effectiveness of inter-layer feature learning in 3D image modeling and highlights the important contribution of the introduced inter-layer information module network and interleaved residual generation network in enhancing modeling performance.

Table 3. Data statistics for a basic classroom environment scene

Table 4. Data statistics for a complex laboratory environment scene

Table 5. Data statistics for a highly detailed campus scene

Based on the data statistics from Tables 3-5, it can be observed that the LOD technology plays a key role in the visual rendering optimization of 3D educational image scenes. In the comparison among basic classroom, complex laboratory, and highly detailed campus scenes, different configurations show varying degrees of improvement in texture size, vertex count, mesh count, and space occupancy. This improvement reflects the effectiveness of LOD technology in increasing scene details and enhancing rendering quality. Especially in the highly detailed campus scene, when using a 4096x4096 texture size, compared to the default configuration, the bounding sphere and AABB configurations show significant growth in vertex and mesh counts, with space occupancy increasing from 1.1GB to 2.4GB and 2.3GB, respectively. This indicates that ensuring higher visual quality correspondingly requires more computing resources and storage space to process these details.

From the perspective of visual rendering optimization strategies, the AABB configuration demonstrates higher data efficiency across all scenes compared to the bounding sphere configuration. AABB achieves a more compact hierarchy, reducing the number of nodes in the hierarchy tree, thus decreasing the mesh and vertex counts and correspondingly reducing space occupancy. Although the AABB and bounding sphere configurations are very close in rendering performance, the optimized AABB configuration's savings on disk space occupancy highlight its superiority in rendering optimization strategies. Therefore, it can be concluded that for efficient visual rendering while maintaining a high-quality educational interactive experience, adopting the AABB configuration of LOD technology is an optimized strategy that balances resource management and visual effects. This strategy can effectively reduce the demand for storage resources without sacrificing rendering quality, making it an ideal choice for rendering 3D educational image scenes.

7.png

Figure 7. Visual rendering curve of 3D educational image scenes at different levels of detail

The experimental data shown in Figure 7 indicates significant differences in the rendering frame rates of 3D educational image scenes under various detail standards. Under high detail standards, the fluctuation in rendering frame rates is more significant due to all meshes being loaded into memory, leading to high memory usage and affecting rendering efficiency. In this case, the instability of rendering frame rates increases with scene roaming and model precision switching, reflecting the system's performance bottleneck when handling large volumes of data. Under medium detail standards, memory usage is controlled, only loading the meshes that need to be rendered and unloading them when no longer needed. This strategy not only reduces the memory burden but also improves the stability of rendering frame rates. At low detail standards, the highest rendering frame rates and stability are achieved, thanks to the application of a preloading streaming mechanism. This mechanism prepares for upcoming model switches by loading all child node meshes of the current hierarchy level, effectively reducing rendering wait times and optimizing frame rate performance.

The comprehensive experimental results conclude that for rendering optimization in 3D educational image scenes, the appropriate LOD technology should be chosen based on the specific needs of the scene and resource limitations. High detail standards, although capable of providing richer visual effects, demand higher hardware resources, which may affect rendering efficiency and system stability. Conversely, low detail standards, through effective resource management and preloading mechanisms, can maintain high frame rates while also providing stable rendering performance, suitable for situations with limited resources.