Classroom Behavior Recognition Algorithm Based on Graph Convolutional Networks and Knowledge Distillation
Zherui Cao, Tao Xue, Qianming Zhou · 2023
Classroom behavior recognition plays a significant role in enhancing educational quality and implementing intelligent education systems.In this study, deep learning algorithms are employed to recognize the behaviors and actions of teachers and students in the classroom. Subsequently, the overall effectiveness of classroom teaching is analyzed based on the responses and actions of teachers and students during class.Classroom behavior recognition provides data support for improving teaching quality and building intelligent classrooms. However, it requires models to have fast inference speed, fewer parameters, and robustness. Existing action recognition models, while accurate, have nonlinear structures and large parameter sizes that are not conducive to deployment. To address this, this paper proposes the Spatio-Temporal Simplified Graph Convolutional Networks(STSGC), a classroom behavior recognition algorithm based on graph convolutional networks and knowledge distillation. The Simplifying Graph Convolutional(SGC) module simplifies the model structure, and in conjunction with knowledge distillation training, transfers the knowledge of the teacher model to the student network, enabling the STSGC model to have fewer parameters and faster inference speed. Compared to the baseline network STGCN++ on the NTU-RGB+D dataset, the STSGC model’s parameters are reduced by about 49.35% to 1.56, and its GFLOPS are reduced by about 99.022%. Experiments on the TSC dataset also show that the model’s accuracy is comparable to that of the teacher model. Additionally, the model’s inference speed on the CPU is reduced by 0.7ms, and on the GPU by 70.4ms.