Automatic Student Engagement Recognition in Online Classrooms Based on Deep Image Segmentation and Attention Mechanisms

With the widespread adoption of online education, the spatial and temporal separation between teachers and students has made it difficult for educators to accurately assess student engagement in real-time.Traditional methods of manually observing engagement are no longer effective, and the complexity of online learning environments presents significant challenges for automatic engagement recognition based on images.Existing studies have several limitations, such as insufficient robustness due to reliance on single-modal features, the inability to adapt to multi-scale variations with fixed kernel convolutions, redundant calculations in attention mechanisms, and the lack of effective utilization of spatial interaction information.To address these issues, this paper proposes an automatic student engagement recognition model for online classrooms based on spatial interaction and segmentation attention.The model introduces a novel variable kernel convolution module, which dynamically adjusts the kernel size and receptive field based on the target scale to enable multi-scale feature extraction.Additionally, an improved multi-branch attention feature fusion module is constructed to process different dimensional features in parallel, strengthening the expression of important features, suppressing redundant information, and reducing computational costs.This model significantly enhances the ability to extract irregular, multi-scale, and spatially correlated engagement features in online classrooms, effectively solving core challenges in online classroom object detection, such as small sample sizes, multi-scale issues, and occlusions.It provides a new technological pathway for engagement evaluation in online education.