YOLO-MRS: An efficient deep learning-based maritime object detection method for unmanned surface vehicles

Deep learning-based object detection for an unmanned surface vehicle (USV) is an important way of visual perception. However, current methods perform poorly when performing complex maritime object detection tasks. It also lacks available datasets of complex maritime objects for visual perception system of USVs. In order to solve these problems, we propose an improved maritime object detection method, called YOLO-MRS, based on lightweight YOLOv8 model in this paper. Specifically, we introduce a multi-scale cross-axis attention (MCA) mechanism into the backbone network of the model to establish long-distance dependencies between pixels to capture global feature information. In addition, we introduce Simplified Spatial Pyramid Pooling-Fast (SimSPPF) to the backbone to enhance prediction accuracy. Also, considering computational efficiency, we replace the ordinary convolutional layers in the backbone network and neck network with refocused convolutional (RefConv) layers to reduce model parameters. Especially, we construct a maritime object detection dataset, termed MODD-13, which contains over 9000 precisely annotated images. The proposed MODD-13 sufficiently considers the characteristics of object categories (13 types), image diversity, sample independence, and background confusion, and can be used as a benchmark dataset for maritime object detection. The final experimental results show that compared with the representative YOLO series models, YOLO-MRS improves the average mAP50 accuracy by 1.8%–7% and mAP50-95 by 1.1%–11.5%, and effectively balances detection accuracy and computational efficiency, thereby effectively achieving fast and accurate detection of maritime objects.