YOLO-lite: A novel lightweight model for underwater object detection using large model and graph structure

Underwater object detection is essential for ocean exploration but presents significant challenges, including poor performance in detecting small objects, environmental complexity, and high computational demands that hinder deployment on underwater robots. To tackle these issues, this study enhances YOLOv5 by introducing a graph-based multiscale feature fusion mechanism, which improves feature integration across layers. Additionally, an auxiliary training branch is incorporated, utilizing semantic segmentation images as supervision signals to enhance small object detection accuracy without increasing inference complexity. To facilitate deployment on edge devices, the model is lightweighted by pruning 70% of its parameters using a Batch Normalization (BN) layer pruning method, achieving a substantial reduction in size and a boost in inference speed with minimal accuracy loss. The optimized model achieves a 3.1 mAP improvement, runs at 100 fps (frames per second) on the Jetson Nano edge device, and satisfies real-time underwater detection requirements. This approach demonstrates potential applicability to other small object detection tasks and edge deployment scenarios.