Improving maritime distress target detection through modelling and simulation with YOLOv5s and Next-ViT

To additionally raise the accuracy of maritime distress target detection, an improved YOLOv5s model with Next-ViT is proposed through modelling and simulation. In this model, Next-ViT is applied to extract representations, followed by adopting a neck network with spatial context pyramid and Focal-GIoU loss to identify the targets. To validate its effectiveness, extensive experiments are conducted on the sub-dataset of the SeaDronesSee dataset. Contrasted to the primary YOLOv5s model, the proposed model has promoted recall, mAP0.5 and mAP0.5-0.95 by 9.2, 6.3 and 3.3 percentage points, respectively, demonstrating superior performance over existing models.