PA-SBR: student behavior recognition based on parallel convolution and attention mechanism

Student behavior recognition is an important part in smart classrooms. There are two main problems with existing student behavior recognition methods. Firstly, it is easy to overlook the behavior of students who are seated densely; Secondly, it is difficult to recognize the behavior of students who sitting in the back row or being partially obstructed by the students in front row. To address these two issues, a new university classroom student behavior recognition method named PA-SBR was proposed. Firstly, a university classroom student behavior dataset named UCSB was constructed, which consists of 1200 images sourced from real university classrooms. Secondly, C3K2-PC and SPPF-PA modules were designed to form the backbone network, and a joint loss function was designed to recognize student behaviors. Parallel convolution in C3K2-PC module is beneficial to obtain behavioral features of occluded students. Pixel-level attention mechanism in SPPF-PA module helps to focus more on key information at different scales and improve the recognition precision of multi-scale targets and student behaviors with small sample size. Meanwhile, the joint loss function can effectively cope with situations where students sit densely and some students are obstructed. Subsequently, a structured pruning technique was utilized to streamline the model, in order to achieve a certain balance between recognition accuracy and time efficiency. Finally, comparison and analysis between PA-SBR and other excellent methods was performed on UCSB and SCB-Dataset3-S datasets. PA-SBR achieved the best performance on both UCSB and SCB-Dataset3-S datasets, with mAP50 values of 0.538 and 0.757, respectively. Compared with YOLO V11 and YOLO V12 with mAP50 value of 0.517 and 0.482 on UCSB and DB-SBR with mAP50 value of 0.729 on SCB-Dataset3-S, the performance of student behavior recognition has improved.