Abstract: With the evolution of educational paradigms from “human-computer collaboration” to “human-intelligence collaborative co-education”,the intelligent evaluation of teaching is also facing new requirements and challenges.In recent years,the task that takes student behavior as the starting point has gained widespread attention.Aiming at the challenges of diverse student beha-viors,heavy occlusions and severe background interference in real classroom environments,a cross-stage partial vision graph network(CPViG-Net) is proposed to enhance the accuracy of student behavior detection in complex classroom settings.Based on a classic object detection framework,the model integrates the dynamic feature modeling ability of the vision GNN and constructs the partial max-relative graph convolution(PMG) module and the cross-stage partial fusion(CPF) module.The PMG module captures the neighborhood information with the greatest feature differences between nodes by embedding maximum relative graph convolution,thereby specifically addressing the issue of information loss caused by local occlusions.It also incorporates depthwise separable convolution to reduce the computational cost of the graph convolution algorithm.The CPF module reconstructs the feature structure using fully connected layers and leverages the cross-stage connection mechanism of the C2f module to achieve multi-level feature fusion,thereby enhancing the ability of the model to recognize small-scale objects.In addition,the model proposes optimization strategies for different datasets through the optimization of nearest neighbor K values.On the public dataset SCB03-S,the mAP@50 of CPViG-Net reaches 70.9%,which is a 2 percentage points improvement over the baseline model.Experiments on multiple publicly available datasets demonstrate that the model exhibits good performance and high robustness in addressing the various challenges of student behavior recognition in real classroom scenarios.

Key words: Student behavior, Max-relative graph convolution, Multi-scale object recognition, Occlusion, Depthwise separable convolution