In modern smart education, accurate recognition of student classroom behavior is paramount. However, the inherent complexity of classroom environments, marked by a high concentration of students and limitations in computational resources of applicable devices, presents significant challenges for accurate behavior recognition. Existing methods often fall short regarding recognition accuracy in addressing these challenges. This paper proposes an innovative lightweight detection network for student behavior in classroom scenario (LDSBC) designed to detect student classroom behavior through object detection. We construct an efficient and lightweight feature extraction architecture that 1) replaces computationally intensive components in traditional structures with lighter convolution, substantially reduces the model parameters, and 2) incorporates an efficient multi-scale attention mechanism in the initial part of the deep feature extraction network, significantly enhancing the network’s feature extraction capabilities. Furthermore, LDSBC introduces a more precise Intersection over Union (IoU) loss function strategy, enhancing the network’s detection capabilities in dense scenarios. Experimental results demonstrate that our LDSBC, compared to the baseline model, sustains detection accuracy without compromise, achieves a 23.5% reduction in the number of parameters, and reduces algorithmic complexity by 19.8%.