S₂Head: Small-Scale Human Head Detection Algorithm by Improved YOLOv8n Architecture

Analyzing crowd density and distribution through human head detection in surveillance footage is crucial for intelligent public safety management. Existing head detection algorithms face challenges in accurately detecting small and densely packed head targets, resulting in lower detection accuracy in complex scenarios. To address this, we propose a novel lightweight head detection algorithm named S2Head, based on an improved YOLOv8n framework. Specifically, we incorporate RepBlock into the backbone for effective multi-scale feature extraction and utilize structural reparameterization techniques to significantly reduce model complexity during inference. Additionally, a small object detection branch and a reparameterizable BiFPN (Rep-BiFPN) structure are incorporated into the neck to improve the model’s sensitivity to small-scale features. Finally, a lightweight MSBlock is also integrated into the head to reduce computational overhead and parameter count without sacrificing detection accuracy. Extensive experiments and ablation studies conducted on the SCUT-HEAD dataset, a benchmark for small-scale head detection, demonstrate that S2Head achieves a 3% improvement in mean average precision (<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\text {mAP}_{0.5}$ </tex-math></inline-formula>) compared to YOLOv8n while reducing model parameters by approximately 46%. These results highlight S2Head’s effectiveness in balancing detection accuracy and computational efficiency, particularly for small-scale head detection tasks.