SI‐FloatDet: A Visual Inspection Method for Water Surface Cleaning Robots Based on Shallow Information Injection and Adaptive Spatial Refinement

The detection of floating garbage on the water surface significantly aids unmanned surface vessels in quickly perceiving their surrounding environment, which is crucial for the development of water surface garbage monitoring and automated debris collection. However, the relatively small size of the detection target compared to the water surface background, along with its susceptibility to noise interference such as light, water waves, and reflections, significantly increases the difficulty of detection. To address the above challenges, this paper proposes a shallow information‐injected pyramidal network, SI‐FPN, and integrates it with the YOLOv11 target detection network to create the SI‐FloatDet framework for complex water surface scenarios. Firstly, to better capture the detailed features of small targets, we design a plug‐and‐play pyramid network (SI‐FPN) that can help solve the problem of information interaction between neighboring feature layers. Secondly, to suppress the noise interference on surface targets, we develop an adaptive spatial refinement module (ASRM). We conduct experiments on the Flow‐Img dataset, which contains a large number of small targets. The results show that compared to the original YOLOv11 model, SI‐FloatDet improves by 6.1% and 4.6% in mAP@0.5 and mAP@0.5:0.95, respectively, and outperforms the existing model in detecting both small and medium targets. Additionally, field experiments were conducted on a water surface trash‐cleaning robot equipped with a vision system. The results show that SI‐FloatDet maintains high accuracy in complex scenarios (e.g., bright light, reflective interference), verifying its reliability and effectiveness in practical applications. This method provides an efficient and reliable solution for detecting water surface litter.