Abstract
Autonomous water-cleaning vessels rely on accurate real-time perception of floating debris, yet detection in complex water-surface environments remains difficult due to small target size, scale variation, water-surface reflection, and illumination interference. We present CEC-YOLO, a lightweight detection algorithm built on YOLO11n that resolves the recurring trade-off between accuracy and inference speed in this setting. Rather than introducing new primitive operators, CEC-YOLO develops a task-specific structural integration of existing lightweight feature-extraction and calibration mechanisms within YOLO11n. CSP-EDLAN is stage-matched to the P2–P5 backbone levels to enhance multi-scale debris representation under a compact computational budget. The segmentation-derived CFC-CRB and SFC-G2 modules are reorganized into an asymmetric cascaded topology, in which semantic context is first calibrated at P5 and then used to guide spatial realignment at P3. This scale-specific coupling targets reflection-induced pseudo-textures, dense aggregation, ambiguous boundaries, and viewpoint-dependent deformation instead of treating the three modules as interchangeable plug-ins. Evaluated on a complex water-surface floating-debris dataset, CEC-YOLO attains a mean average precision (mAP@0.50) of 88.9% while sustaining 175 frames per second on an RTX 3090, surpassing the YOLO11n baseline and other mainstream detectors in the accuracy-speed balance. A cross-dataset evaluation on the public IWHR-AI-Lable-Floater-V1 floating-debris dataset further confirms that these gains are not specific to the self-built data. By delivering high precision at real-time speed in a compact model, CEC-YOLO provides a deployable perception solution for intelligent surface-cleaning vessels and contributes a practical pathway toward automated marine-pollution monitoring and mitigation.