Behavior Analysis of Students in Preschool Mathematics Teaching Based on Deep Learning

• Published in: Journal of Applied Science and Engineering Vol. 29; no. 3; pp. 585 - 593
• Main Author: Guangning Qin
• Format: Journal Article
• Language: English
• Published: Tamkang University Press 01.07.2025
• Subjects: behavior analysis; channel attention mechanism; deep learning; multi-scale context fusion; yolov8
• ISSN: 2708-9967; 2708-9975
• DOI: 10.6180/jase.202603_29(3).0009

Students behaviors can directly reflect the quality of the classroom. Analyzing and evaluating classroom behaviors through artificial intelligence and deep learning is conducive to improving teaching quality. The traditional methods for identifying students classroom behaviors involve that teachers directly observe students states or analyze them through surveillance videos after class. These methods are time-consuming, laborintensive, and have a low recognition rate, making it difficult to reflect the problems existing in the classroom and exams in real time. To solve this problem, this paper proposes a novel students classroom behaviors based on YOLOv8 deep learning model. Combining the channel attention mechanism with deep convolution, a dynamic channel attention convolution (DCAConv) is proposed, which can dynamically adjust the channel weights and capture key features more sensitively. It introduces multi-scale convolutional attention (MSCA) to maximize the ability of mining multi-scale convolutional features through element multiplication, and enhance the attention to spatial details. Meanwhile, a multi-scale context fusion (MSCF) module is constructed. Through convolution and self-attention mechanism, multi-scale feature fusion is enhanced. Adding a small target detection layer and extracting local features from larger-sized feature maps significantly improves the ability to recognize the behaviors of students in the back row. The experimental results show that the average recognition accuracy rate of the proposed behavior recognition method for various parts of the human body can reach up to 83.7% at most. The recognition rates for various behaviors in simple and crowded scenarios reach more than 92.1% and 86.3% respectively, and it can effectively recognize various behaviors in the classroom.