RDA-YOLOv11: Stage-wise adaptive feature enhancement for underwater object detection in complex marine environments
Year: 2026
Type: article
Abstract: As ocean observation, marine resource development, and underwater operations become increasingly automated, reliable visual perception in complex underwater environments has become essential for marine information acquisition and autonomous operations. However, light absorption and scattering in water often cause color attenuation, reduced contrast, and blurred details in underwater images. Complex backgrounds, substantial variations in object scale, and dense target distributions further weaken discriminative information and complicate classification and localization. To address the difficulty of consistently exploiting limited discriminative information during feature propagation in complex underwater environments, this study proposes RDA-YOLOv11, a stage-wise adaptive feature enhancement method for underwater object detection in complex marine environments. The method establishes a continuous adaptive enhancement process along the feature propagation pathway to mitigate the persistent effects of underwater image degradation on object features. During feature extraction, it strengthens object-relevant local responses to reduce interference from low contrast and blurred backgrounds. During spatial resampling, it dynamically adjusts the reorganization of spatial information according to feature content, better preserving spatial cues associated with small objects and indistinct boundaries. At the detection stage, it adaptively coordinates the contributions of multi-scale features at each spatial location to accommodate variations in object scale and background interference. These operations improve the extraction, preservation, and fusion of limited discriminative information in underwater visual representations. Experiments on two public underwater datasets, DUO and RUOD, show that RDA-YOLOv11 achieves mAP@50–95 values of 65.1% and 65.3%, respectively, outperforming YOLOv11s by 2.1 and 2.5 percentage points. The complete stage-wise configuration also achieves the highest Recall, mAP@50, and mAP@50–95 in the ablation experiments. While maintaining relatively high inference efficiency, the method provides a promising object-detection solution for visual perception in marine ecological monitoring, aquaculture monitoring, subsea infrastructure inspection, and intelligent underwater platforms.
Source: Applied Ocean Research
Authors: Bing Hao, Jiaming Wang, Xinlin Sun
Institution: Qiqihar University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,950
OpenAlex ID: W7235930032
