Learning State Quantification of University Classroom Behaviors Based on Multiscale Convolution and Attention Mechanisms

Read full text Summarise with VIRAT
Authors: Xiaoxu Liu
Cited by: 0
References: 22

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Recognizing university classroom behaviors and quantifying students' learning states are essential components in the development of intelligent teaching systems, providing critical support for instructional optimization and precise learning analytics.However, existing image-based classroom behavior recognition methods often suffer from limited adaptability to target scale variations, insufficient discrimination between easily confused behaviors, and difficulties in transforming recognition results into interpretable quantitative indicators of learning states.To address these challenges, this study proposes a dual-branch spatiotemporal-frequency attention network to achieve high-precision classroom behavior recognition and interpretable learning state quantification.The network incorporates a spatial-domain branch based on multiscale deformable convolution to construct a dynamic scale-weight allocation mechanism, enabling adaptive balancing of feature representations for targets of different scales.In addition, a frequency-domain attention branch is introduced to capture fine-grained behavioral differences by integrating frequency selection and phaseaware representations.A cross-modal semantic alignment and fusion module is further designed, which leverages cross-attention to accurately integrate heterogeneous spatial and frequency features.Finally, an end-to-end learning state quantification module is developed, establishing a quantitative mapping between behavior recognition and learning states through spatiotemporal attention pooling and contextual modulation factors.Experiments conducted on a self-constructed classroom behavior dataset and a public fine-grained behavior dataset demonstrate that the proposed method outperforms existing mainstream approaches in both behavior recognition accuracy and the rationality of learning state quantification.This study enriches the technical approaches for fine-grained behavior recognition and expands the application scope of image processing technologies in intelligent education scenarios, providing reliable quantitative support for teaching decisionmaking.

Fields of Study & Subjects

Human Pose and Action Recognition Advanced Neural Network Applications Online Learning and Analytics

Publication details

Journal / source
Traitement du signal
Volume
43
Issue
1
Pages
53–63
Publication date
28 February 2026
Language
EN