ACM-YOLOv10: Research on Classroom Learning Behavior Recognition Algorithm Based on Improved YOLOv10

To effectively integrate the research on learning engagement with teaching practices and accurately assess and analyze students’ learning behavior participation in the classroom to improve teaching quality, this paper proposes an improved YOLOv10 algorithm model, ACM-YOLOv10, targeting the issues of insufficient detection precision, missed detection, false detection, and slow speed of traditional recognition algorithms in classroom behavior detection under multi-scale scenarios and occluded targets. The model is designed with an Asymmetric Depthwise Separable Convolution (ADSConv) module to replace the traditional convolutional layers. This module, with its lightweight design, optimizes convolution operations to reduce computational complexity and parameter quantity, thereby accelerating the model’s inference speed. Additionally, the Spatial and Channel Reconstruction Convolution (C2f_SCConv) module is embedded in the backbone and neck networks, combining the characteristics of SCConv to reduce redundant features of CNN through spatial and channel reconstruction units, which effectively enhances the model’s ability to learn local features, particularly suitable for fine-grained feature extraction in complex backgrounds. Finally, the redesigned spatial Pyramid Pool Fast Learning Separable Kernel Attention (SPPF_LSKA) module is introduced into the backbone network, improving the multi-scale feature fusion method to more effectively handle features of different scales and enhancing the model’s sensitivity and detection capability for minority class behaviors. Experimental results demonstrate that the improved ACM-YOLOv10 model achieves mAP of 82.4% for six behavior categories on the SCB-Dataset3, which is a 5.3% improvement compared to the baseline model, outperforming other mainstream detection models and meeting the practical requirements for Student Behaviors (SB) detection. Additionally, generalization experiments conducted on another SB dataset confirm that the improved algorithm model possesses good generalization performance.