DDR-DETR: Real-Time Face Detection Algorithm for Classroom Scenarios

To address challenges in classroom scenarios, such as image blurriness, low resolution of distant faces, and frequent occlusions, a lightweight DDR-DETR model is designed based on the RT-DETR (Real-Time Detection Transformer) model to meet real-time requirements in these environments. The model introduces an efficient StarNet module to streamline the backbone network, optimizing real-time detection performance. A CGRLFPN structure is implemented to solve issues of retaining fine-grained details while enhancing global features during multi-scale feature fusion. Additionally, a multi-scale multi-head self-attention mechanism (MSMHSA) is employed to improve extraction of multi-scale global contextual and channel information. Experimental tests conducted on a custom dataset demonstrate a significant increase in mean average precision (mAP50-95) for face detection. Thus, this study provides an effective technical solution for real-time face detection in classroom scenarios.