Lightweight outdoor drowning detection based on improved YOLOv8

Despite abundant global water resources, drowning remains one of the complex challenges to tackle worldwide. To solve the problem of complicated outdoor remote water environments, where people in the water have different morphologies and scale variations, and the existing detection models are highly complex and computationally intensive, we propose a lightweight drowning detection model, YOLOv8-REH, based on YOLOv8n. First, the C2f-RVB-ELA feature extraction module (effective fusion of C2f and RVB-ELA) is designed to improve the C2f module of the YOLOv8n backbone network, reducing the model's parameters and computation effectively. Second, in the Neck section, we incorporate the ELA-HSFPN feature fusion module, which consists of the Hierarchical Scale-based Feature Pyramid Network (HSFPN) module and the Efficient Local Attention (ELA) mechanism. This helps us gather multi-scale and spatial information perception more comprehensively and efficiently, enhancing the feature fusion capability of the model. Then, we introduce the Powerful-IoUv2 loss function to enhance bounding box regression along the effective path, consequently improving the model's convergence speed and detection performance. Finally, we use a pruning method based on layer-adaptive magnitude-based pruning scoring to prune and remove unimportant redundant parameters from the improved model, further compressing the model complexity and achieving a better lightweight effect. The final compressed YOLOv8-REH model is compared with the current mainstream algorithms for comparison experiments as well as ablation experiments. The experimental results indicate that the YOLOv8-REH model sustains an average detection accuracy while the computational volume, parameter count, and model size reach 3.7GFLOPs, 0.72 M, and 1.8 MB, and the FPS is improved by 22.9, which achieves a significant improvement in the model's lightweight performance compared with the existing methods.