A Drowning Detection Model by Integrating Multi-Scale Enhanced Receptive Fields with the YOLOv8 Framework

To prevent drowning accidents, researchers have developed various deep convolutional networks for drowning action detection. While the YOLOv8 model excels in detection accuracy and speed, it faces challenges such as limited field of view and missed small object detections. This paper introduces a drowning detection model that integrates multi-scale enhanced receptive fields with the YOLOv8 framework. Ghost convolution and partial convolution are embedded into YOLOv8 residual modules, with linear transformations applied to reduce parameters while enriching feature extraction. A lightweight detection head is designed with shared feature extraction modules for decoupled operations. Additionally, separable large kernels are employed for secondary feature fusion after spatial pyramid pooling to broaden the network’s receptive field. The proposed model improves detection accuracy by 1.87% and increases processing speed by 22.4 FPS compared to YOLOv8. Experimental results demonstrate its ability to capture broader spatial information and enhance small object detection.