A Student behavior Recognition Algorithm Based on Improved MobileNetV2

Student behavior recognition is crucial for detecting learning activities within the classroom. Addressing the high complexity, poor real-time performance, and low accuracy of existing recognition models, we develop a student behavior classification dataset and propose an algorithm for student behavior recognition based on improved MobileNetV2. The algorithm diminishes model complexity and accelerates inference speed by reducing the width factor and improves model accuracy by improving classifier of the model and embedding coordinate attention block. The ablation experiment demonstrates the algorithm’s superior performance, achieving an average accuracy of 92.45% on the dataset, a 2.64% improvement over the initial model. The parameter is also significantly reduced to 0.56M, marking a 75% decrease. Computational requirements are lowered to 1.67 GFLOPs, representing a 68% reduction. Furthermore, the inference speed is enhanced to 244 FPS, a 106.7% increase compared to the initial model. Comparative experiment indicates that our algorithm outperforms other lightweight neural networks, including EfficientNet, DenseNet, ShuffleNetV2, and MobileNetV3, in terms of model complexity, inference speed, and accuracy.