Intelligent Analysis and Optimization of Adaptability in Interdisciplinary Learning Environments Using Image Recognition Technology

In the context of global educational reform, interdisciplinary learning has gradually become an important approach to cultivating innovative talents [1-4]. Interdisciplinary learning environments break the boundaries of traditional disciplines, requiring students to possess abilities such as integrating diverse knowledge, collaborating with others, and engaging in self-directed exploration [5-8]. However, how to assess students’ adaptability in such complex environments remains a significant challenge in educational research. With the development of artificial intelligence and computer vision technologies, the use of image recognition technology to analyze students' behavior in interdisciplinary learning environments provides new technical means for adaptability assessment [9-11]. This approach not only helps educators monitor students' learning status in real-time but also provides data support for improving teaching effectiveness.

Research on adaptability in interdisciplinary learning environments holds important educational significance. First, understanding students' adaptability can help educators identify their strengths and weaknesses in the learning process, allowing them to adjust teaching strategies to promote students' overall development. Second, studying adaptability in interdisciplinary environments can support personalized education and precise teaching, helping students overcome challenges in interdisciplinary learning and enhancing their self-directed learning ability and innovative mindset [12-16]. Therefore, constructing an intelligent adaptability analysis system plays an essential role in optimizing interdisciplinary learning environments and improving students' learning outcomes.

Although some research has been conducted on adaptability in interdisciplinary learning, most of it relies on traditional methods such as questionnaires and interviews. These methods are time-consuming and labor-intensive, and the results are susceptible to bias from subjective factors [17, 18]. Moreover, existing applications of image recognition technology mostly focus on limited areas of behavior recognition, lacking a comprehensive analysis of multidimensional behavioral characteristics in complex learning environments [19-24]. Thus, current methods cannot fully meet the needs of adaptability assessment in interdisciplinary environments, making it difficult to achieve real-time, dynamic, and precise analysis of students' behavior.

This study focuses on three main aspects. First, based on the characteristics of interdisciplinary learning environments, it defines key behaviors that reflect students' adaptability and constructs a clear system of behavioral indicators. Second, it develops an adaptability behavior detection algorithm to identify specific behavioral characteristics through image recognition technology. Finally, it conducts a comprehensive evaluation of students' adaptability in interdisciplinary environments based on the behavior detection results. This research provides essential data support for intelligent analysis and optimization in interdisciplinary learning environments, aiming to offer educators scientific, real-time methods for assessing students' adaptability to support personalized teaching and enhance learning outcomes.

3.1 Encoder

Figure 2 illustrates the principles of the adaptability behavior detection algorithm for interdisciplinary learning environments. In the algorithm for detecting students' adaptability behaviors in interdisciplinary settings, a multi-scale separable spatiotemporal feature encoder has been designed for more efficient and accurate capture of students' spatiotemporal behavioral characteristics. The core of this encoder lies in the independent processing of temporal and spatial dimension information within student behaviors, thereby reducing the computational burden associated with global self-attention. Students' behavioral adaptability typically encompasses variations in bodily actions and trends in continuous behaviors, namely subtle spatial positioning information and temporal sequence information. In the multi-scale separable spatiotemporal feature encoder, the attention mechanisms for space and time are separated, implemented through two submodules: Temporal Pooling Attention (TPA) and Spatial Pooling Attention (SPA).

In the separable spatiotemporal pooling attention module, the input data a^{^} is a sequence of vectors with M dimensions of F, specifically representing the encoded features of a sequence of S frames of collected image segments at a spatial resolution of G×Q. This module is divided into the SPA and TPA submodules, which are used to extract spatial and temporal features from interdisciplinary behavioral data to reveal potential spatiotemporal adaptability changes in student behavior.

2.png

Figure 2. Principle of the adaptive behavior detection algorithm in interdisciplinary learning environments

3.png

Figure 3. Structure of the SPA module

The function of the SPA submodule is to compute the attention weights between various spatial positions within the same time frame, thereby extracting multi-scale spatial features within each frame. As shown in Figure 3, the SPA processes all input vector sequences within each frame, obtaining the corresponding query, key, and value matrices through linear transformations. Unlike traditional self-attention mechanisms, SPA reduces computational complexity and extracts multi-scale features by applying pooling operations to downsample the query, key, and value matrices. This pooling process reduces the spatial dimensions, allowing for multi-scale integration of feature information from each frame, enabling the model to focus on more prominent local features. The pooling results are then used to calculate attention through dot products, producing a pooled spatial attention matrix that captures the relationships among spatial blocks. The pooled spatial attention matrix effectively represents the multi-scale spatial structural information within the frame, forming multi-scale spatial features for subsequent temporal feature extraction. Assuming layer normalization is denoted by LN, the learnable tensor of dimensions F×F is represented by Q_Ws, Q_Js, and Q_Ns, the spatial position variable is denoted by r_t. The specific transformation process can be represented as:

${{W}_{s}}={{Q}_{{{W}_{s}}}}LN\left( {{{\hat{A}}}_{s}}+{{r}_{t}} \right)$     (2)

${{J}_{s}}={{Q}_{{{J}_{s}}}}LN\left( {{{\hat{A}}}_{s}}+{{r}_{t}} \right)$      (3)

${{N}_{s}}={{Q}_{{{N}_{s}}}}LN\left( {{{\hat{A}}}_{s}}+{{r}_{t}} \right)$     (4)

Assuming the scale transformation factor is represented by 1/(f)^1/2, the pooling operation can be expressed as:

${{\hat{W}}_{s}}=POOL\left( {{W}_{s}};{{O}_{W}} \right)$     (5)

${{\hat{J}}_{s}}=POOL\left( {{J}_{s}};{{O}_{J}} \right)$       (6)

${{\hat{N}}_{s}}=POOL\left( {{N}_{s}};{{O}_{N}} \right)$      (7)

The attention calculation formula for the pooled results is given by:

$ATT\left( {{W}_{s}},{{J}_{s}},{{N}_{s}} \right)=\text{softmax}\left( \frac{{{{\hat{W}}}_{s}}{{{\hat{J}}}_{s}}^{S}}{\sqrt{f}} \right){{\hat{N}}_{s}}$      (8)

The pooling residual connection operation calculation formula is as follows:

${{B}_{s}}=ATT\left( {{W}_{s}},{{J}_{s}},{{N}_{s}} \right)+POOL\left( LN\left( {{{\hat{A}}}_{s}}+{{r}_{t}} \right);{{O}_{W}} \right)$     (9)

The role of the TPA is to extract multi-scale behavioral features along the temporal dimension from different frames, detecting changes in student behavior over time. Specifically, TPA takes the output BÎR^M'×F from the SPA module as input and performs temporal pooling operations on the features of the same spatial position in a chronological order. By pooling along the temporal dimension, TPA reduces the dimensionality of the temporal feature representation and captures behavioral changes over long sequences, eliminating subtle noise. As a result, the features extracted by TPA better reflect the ongoing adaptation process of students in interdisciplinary learning environments, identifying behavioral patterns and trends as students adapt to interdisciplinary tasks. Through dot product calculations of the pooled temporal attention, the output CÎR^M''×F represents the attention feature matrix after temporal pooling, where M''=S'×G'×Q', which serves as the output of the final multi-scale spatiotemporal encoder. Assuming W_t=Q_WtMV(B_t+r_s), J_t=Q_JtMV(B_t+r_s), and N_t=Q_NtMV(B_t+r_s), with all vectors at temporal position t represented by B_t and the temporal position vector represented by r_s, the formula is:

$\begin{align}  & {{C}_{t}}=\text{softmax}\left( \frac{POOL\left( {{W}_{T}};{{O}_{W}} \right)POOL{{\left( {{J}_{T}};{{O}_{J}} \right)}^{S}}}{\sqrt{f}} \right)POOL\left( {{N}_{T}};{{O}_{N}} \right) +POOL\left( LN\left( {{B}_{T}}+{{r}_{s}} \right);{{O}_{W}} \right) \\\end{align}$      (10)

In interdisciplinary learning environments, students' behavioral adaptability often manifests as alternating changes across different time periods and spatial locations. For instance, in an interdisciplinary task combining scientific experimentation with artistic creation, students may initially exhibit fluctuations in attention over short periods, while over time, their behavior evolves into a global adaptability to the entire task process. The multi-scale spatiotemporal feature encoder needs to capture the behavioral patterns of students transitioning from local adaptation to global adaptation through these gradual changes, aiding in the intelligent analysis of their adaptability status. In the initial phase, the encoder focuses on local detail features, with more refined temporal and spatial feature representations, that is, larger feature dimensions S, G, Q, indicating that more detailed information is preserved in both time and space, while the number of feature channels F is smaller, representing weaker expressive ability. This stage primarily captures local changes in student behavior, such as micro-behaviors or short-term emotional fluctuations. As the model progresses, spatiotemporal features gradually become more abstract, with the encoder placing greater emphasis on global behavioral patterns; the feature dimensions S, G, Q decrease, indicating global changes in time and space, while the number of feature channels F increases, providing a stronger representational capability.

The structure of the multi-scale spatiotemporal feature encoder consists of multiple stacked multi-scale spatiotemporal attention modules, each extracting multi-scale features of student behavior through the separable spatiotemporal pooling attention mechanism. Each module contains several separable spatiotemporal pooling attention mechanism modules, each performing attention calculations and pooling operations for the spatial and temporal dimensions separately, gradually extracting student behavioral features across different temporal and spatial scales. Specifically, for the input A_uk of the k-th separable spatiotemporal pooling attention mechanism module in the u-th module, assuming spatial pooling attention is represented by SPA, temporal pooling attention by TPA, and multilayer perceptron by MLP, the calculation process for the output C_uk is as follows:

${{B}_{uk}}=TPA\left( SPA\left( {{A}_{uk}} \right) \right)$      (11)

${{C}_{uk}}=MLP\left( LN\left( {{B}_{uk}} \right) \right)+{{B}_{uk}}$     (12)

3.2 Decoder

In the student adaptability behavior detection algorithm for interdisciplinary learning environments, the introduction of the multi-scale spatiotemporal feature decoder aims to restore more abstract global features to specific spatiotemporal detail features, thereby better analyzing and understanding students' adaptive behaviors. Unlike the encoder, the decoder achieves a multi-layered, fine-grained analysis of students' adaptive behaviors through a stepwise decoding process that combines feature representations at different scales. The structure is shown in Figure 4.

4.png

Figure 4. Decoder structure

In this multi-scale spatiotemporal feature decoder, a behavior micro-tube query mechanism is introduced, generating a set of dynamic micro-tube queries for each specific adaptive behavior, which are used to track and decode students' adaptation details during interdisciplinary learning processes. For example, in interdisciplinary tasks, students may need to switch between “analysis” and “synthesis” cognitive modes; this behavioral change will be gradually decoded and recorded by specific micro-tube queries, forming a dynamic trajectory that includes the adaptation process, thereby helping the model better understand students' adaptive behavior characteristics.

Here, the behavior micro-tube query WÎR^V×Z is a set of representation vectors used to capture students' adaptive behaviors in interdisciplinary learning contexts. Here, V represents the number of queries, indicating the possible instances of adaptive behaviors to be detected, and Z represents the feature dimension of each query, which includes both spatial and temporal dimension information of student behaviors. Specifically, for each query W_u, it can be understood as a behavior micro-tube that contains the spatial coordinates of that behavior in time and space, allowing for the gradual recording and tracking of students' adaptation processes across different times and subject tasks.

The introduction of a behavior micro-tube spatiotemporal self-attention module in the decoder further enhances the extraction and decoding effect of spatiotemporal features in behavior detection. In the spatial dimension, the attention mechanism focuses on the relationships between different behavior self-labels at the same point in time, capturing the feature interactions among these behavior self-labels to identify the details of students' adaptive behaviors in a particular subject task. In the temporal dimension, the attention mechanism focuses on the feature interactions of the same human target at different time points. This mechanism can track changes in students' behaviors in interdisciplinary contexts, such as the transition from a divergent thinking mode in the literature subject to a concentrated thinking state in mathematical logic; this interdisciplinary adaptive behavior can be captured and interpreted through temporal feature associations. During the feature decoding process, the behavior micro-tube spatiotemporal self-attention module employs upsampling operations instead of pooling operations to elevate the spatiotemporal resolution of micro-tube queries to align with the features of the encoder. Due to the multi-scale feature distribution of adaptive behaviors in different subject environments, directly merging the original low-resolution micro-tube features with encoder features may result in a loss of detail. Through upsampling, micro-tube queries can be accurately aligned with the multi-scale features of the encoder at a higher resolution, ensuring that the details of adaptive behaviors are preserved within the micro-tube queries, thereby providing higher spatiotemporal decoding accuracy.

Given that behavioral features in different subject environments often exhibit different granularities and patterns in both time and space, the algorithm specifically employs consistent granularity attention, or scale-consistent attention, during the decoding process to achieve precise spatiotemporal feature fusion for features at the same scale.

In this module, the input behavior micro-tube features and the spatiotemporal features from the encoder are subjected to spatial cross-attention (SCA) and temporal cross-attention (TCA) calculations at each scale. This separable spatiotemporal attention mechanism allows for feature fusion along the spatial and temporal dimensions, avoiding potential feature entanglement that may occur when processing both dimensions simultaneously. Additionally, to ensure the scale consistency between behavior micro-tube features and encoder spatiotemporal features, the module adheres to a scale alignment strategy during the feature fusion process, satisfying the requirement u+k=M+1, where M is the number of different spatiotemporal scales possessed by both the encoder and decoder.

After completing the spatial and temporal cross-attention calculations, the spatiotemporal cross-attention module incorporates fully connected layers and residual connections to further fuse behavior micro-tube features with encoder features, yielding the output C^u_XS of the spatiotemporal cross-attention module. The calculation process can be expressed as:

$B_{XS}^{u}=TCA\left( SCA\left( A_{XS}^{u},A_{rv}^{k} \right),A_{rv}^{k} \right)$     (13)

$C_{XS}^{u}=MLP\left( LN\left( B_{XS}^{u} \right) \right)+B_{XS}^{u}$     (14)

3.3 Prediction head and loss function

The primary objective of the prediction head is to simultaneously predict behavior categories and spatiotemporal locations. In the student adaptability behavior detection task within interdisciplinary learning environments, this means not only identifying the types of adaptive behaviors exhibited by students in different subject contexts but also accurately localizing these behaviors in both temporal and spatial dimensions. To achieve this, the prediction head employs a strategy similar to spatiotemporal behavior detection, using continuous bounding box coordinates to represent the spatiotemporal locations of behaviors. These bounding boxes correspond to the start and end times of the behaviors in the temporal dimension and to specific regions of behavior features in the spatial dimension. In the prediction head, the prediction of behavior micro-tubes is treated as a set matching problem. The core of this mechanism lies in optimally matching the predicted behavior micro-tubes from the network model with the true behavior micro-tubes input into the model. In practical implementation, the set matching mechanism considers two aspects of loss: first, the classification loss of adaptive behaviors, which is the difference between the predicted behavior category and the actual category; second, the bounding box loss, which measures the error between the predicted spatiotemporal location of behavior micro-tubes and the true location. This comprehensive loss optimization strategy ensures that the prediction head can accurately localize students' adaptive behaviors in both time and space, enhancing the overall generalization ability of the model.

The input to the prediction head is denoted as D_fr​, which is the output feature from the last layer of the decoder. In the prediction head, these features undergo linear mapping through fully connected layers to produce the predicted values for behavior categories and spatiotemporal locations. Specifically, in the context of adaptive behavior detection in interdisciplinary learning environments, the categories of adaptive behaviors may exhibit significant diversity due to differences in disciplines. For example, cognitive behaviors may be displayed in scientific environments, while social behaviors may manifest in humanities contexts. Through the optimization of cross-entropy loss, the model can accurately capture the variations in adaptive behavior categories within interdisciplinary situations. Assuming V represents the number of behavior micro-tubes and Z the total number of behavior categories, with DZ ​representing the linear mapping from the fully connected layer, the calculation formulas are given as:

${{\hat{B}}_{CL}}=F{{C}_{CL}}\left( {{D}_{fr}} \right)$     (15)

${{M}_{CL}}=CE\left( {{{\hat{B}}}_{CL}},{{B}_{CL}} \right)$     (16)

The bounding box regression task aims to determine the specific spatiotemporal locations of adaptive behaviors. Unlike traditional object detection tasks that first predict all possible bounding boxes and then determine the final bounding box through methods like non-maximum suppression, this algorithm's prediction head directly outputs the final target bounding box B^{^}_CO to enhance the accuracy and efficiency of spatiotemporal localization. Let η_M1 and η_IO represent the weights for L₁ loss and IoU loss, respectively. The specific bounding box regression loss consists of a weighted sum of L₁ loss and IoU loss:

${{\hat{B}}_{BO}}=F{{C}_{BO}}\left( {{D}_{fr}} \right)$     (17)

$\begin{align}  & {{M}_{CO}}={{\eta }_{{{m}_{1}}}}{{M}_{1}}\left( {{B}_{CO}},{{{\hat{B}}}_{CO}} \right) +{{\eta }_{IO}}{{M}_{IO}}\left( {{B}_{CO}},{{{\hat{B}}}_{CO}} \right) \\\end{align}$      (18)

In the student interdisciplinary learning environment, the appearance and disappearance of adaptive behaviors may span considerable time periods. For example, when students transition from one subject task to another, their behavior may experience a brief adaptation phase. The behavior switching module can effectively detect this transitional state through the optimization of binary cross-entropy loss:

${{\hat{B}}_{xt}}=F{{C}_{xt}}\left( {{D}_{fr}} \right)$     (19)

${{M}_{xt}}=BCE\left( {{{\hat{B}}}_{xt}},{{B}_{xt}} \right)$     (20)

To comprehensively optimize the classification loss, bounding box regression loss, and behavior switching loss, the algorithm adopts a weighted sum of multi-task losses as the final optimization objective. Assuming the weights of the various loss functions are represented by η₁, η₂, η₃, and η₄​, and B^{^}_MA and B are a pair of optimal matches, the calculation formula is given by:

$M={{\eta }_{1}}{{M}_{CL}}\left( \hat{B}_{CL}^{MA},{{B}_{CL}} \right)+{{\eta }_{2}}{{M}_{1}}\left( \hat{B}_{CL}^{MA},{{B}_{CL}} \right)$$+{{\eta }_{3}}{{M}_{IO}}\left( \hat{B}_{CO}^{MA},{{B}_{CO}} \right)+{{\eta }_{4}}{{M}_{xt}}\left( \hat{B}_{CO}^{MA},{{B}_{xt}} \right)$     (21)

This study focuses on the intelligent evaluation of students' adaptive behaviors in interdisciplinary learning environments through image recognition technology, providing a novel technical pathway for adaptive analysis in interdisciplinary education. The research consists of three main parts: First, based on the complex characteristics of interdisciplinary learning, key adaptive behaviors required of students—such as active participation, hands-on operation, focused learning, and interdisciplinary communication—are defined. A corresponding behavioral feature indicator system is constructed to provide a standardized theoretical framework for behavior recognition. Second, the study develops an adaptive behavior detection algorithm that utilizes image recognition technology to monitor students' behavioral performance in real time, accurately identifying behavioral features. Experiments with different frame rates and resolutions further optimize the precision of the image recognition model. Finally, based on the results of behavior detection, a comprehensive adaptive assessment system is constructed, providing valuable analytical data and references for personalized support and instructional improvements in interdisciplinary education.

The experimental results indicate that the proposed detection method performs well under various input settings. The experiments with different spatial and temporal scales validate the impact of varying frame rates and resolutions on detection performance, revealing that higher frame rates and resolutions significantly enhance the accuracy of adaptive behavior recognition. Furthermore, the confusion matrix for testing different behavior categories demonstrates that the model exhibits high accuracy and stability in detecting active participation, hands-on operation, focused learning, and interdisciplinary communication behaviors. These results suggest that the proposed method can reliably identify student behaviors in interdisciplinary learning environments, offering targeted feedback to teachers and enhancing the quality and efficiency of interdisciplinary instruction.

The value of this research lies in advancing adaptive assessment in the field of interdisciplinary education through intelligent analytical techniques, addressing the gap in existing studies regarding the intelligent detection and evaluation of student adaptive behaviors. However, certain limitations remain, such as the model's performance showing some confusion among specific behavior categories due to experimental constraints. Additionally, as the content of interdisciplinary education evolves dynamically, the behavioral feature indicator system may need continual updates to meet the demands of different educational contexts. Future research could further expand upon this method by incorporating additional behavioral dimensions into the detection system and optimizing the model's ability to recognize complex behaviors. Specifically, in collaborative learning scenarios, exploring multimodal data fusion—such as integrating voice and text information—could enhance the accuracy and comprehensiveness of the detection system. Moreover, future efforts could focus on developing adaptive indicator systems for dynamically updated behavioral features, supporting more personalized assessment and interventions in interdisciplinary education.