Object detection techniques in computer vision are increasingly employed in education to analyze student performance and evaluate teaching effectiveness. However, current classroom behavior recognition methods have problems such as decreased accuracy, high memory usage, and slow detection speed. In order to address these challenges, we design a lightweight object detection network (SCB-LEDN) based on the classroom monitoring perspective to enhance the detection accuracy of objects while reducing the model’s parameters and accelerating real-time inference. First, we create the student classroom behavior dataset (SCB-E), which includes five behaviors: raising hands, reading, sleeping, writing, and using a mobile phone. Secondly, We propose the partial detection head (PHead), effectively reducing parameter count and improving detection speed. We introduce the diverse branch block (DBB) in the feature extraction layer and create the C2DB module, enhancing student classroom behavior recognition accuracy. We introduce the EfficientRepBiPAN structure in the neck layer, enabling practical preservation and fusion of spatial information. The experimental results show that our method achieves an average precision (mAP) of 91.5% on the SCB-E dataset and reduces the parameters by 20% compared to the baseline method.