Objective To address the challenges of false detection and missed detection in identifying floating bottles on the water surface in unmanned surface vehicle applications, this study proposes an improved small floating object detection algorithm based on YOLOv5s.

Method  First, data augmentation was performed on the Flow-Img dataset to expand the data and avoid model overfitting. Second, to enhance detection accuracy of the deep learning model for extremely small objects, an additional detection layer was introduced beyond the original three in YOLOv5s, while the detection head for large objects was removed to avoid anchor box allocation issues caused by data imbalance. Third, the CBAM (Convolutional Block Attention Module) was incorporated into the backbone network to address the limited feature extraction capability for detecting floating bottles on the water surface. Finally, the Normalized Wasserstein Distance (NWD) regression loss function was introduced and combined with the IoU loss function in a weighted manner to construct a comprehensive regression loss function, further enhancing detection accuracy for floating bottles on the water surface.

Results Experimental results show that the proposed algorithm achieves a mAP@0.5 of 95.7% in detecting floating bottles on the water surface. Compared to the original YOLOv5s, the improved model increases mAP@0.5 by 2.6%, mAP@0.95 by 4.5%, and reduces the number of parameters by 61.9%.

Conclusion  While maintaining a lightweight architecture, it delivers more accurate detection results for surface floating bottles, offering a valuable technical reference for small floating object detection on the water surface.