.png){kind=icon}
.png){kind=icon}
Return
Enhancing underwater object detection through hybrid sparse-annotation optimization
.png){kind=icon}
PRE
AI
DOI：10.1007/s00530-026-02315-9.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Underwater image detection is crucial for marine ecological monitoring, yet it faces challenges due to sparse annotations caused by difficult and labor-intensive manual labeling conditions in real underwater environments. The resulting lack of reliable supervision severely limits the performance of modern detectors. To address this problem, we propose the Underwater Sparse-annotation Hybrid Enhancement for Detection framework (U-SHED), which jointly improves detection performance through pseudo-label optimization and feature enhancement. Our Dual-stage Pseudo-label Optimization mechanism (DSPO) performs reliable pseudo-label selection by adjusting temporal and spatial thresholds through a Dual-gate Pseudo-label Filtering module (DPF), while its Multi-dimensional Pseudo-label Optimization module (MPO) expands missed targets through controlled pseudo-label enhancement. In parallel, the Self-supervised Feature Enhancement module (SFE) leverages cross-view consistency to stabilize feature representations, mitigating the adverse effects of underwater image degradation. Experiments on the DUO dataset and a hydroid ecological monitoring dataset across multiple annotation sparsity levels show that U-SHED consistently outperforms all compared methods. In particular, under the Sparse-50p setting, U-SHED achieves mean average precision improvements of 3.75% on DUO and 2.59% on the hydroid dataset, demonstrating its effectiveness in addressing sparse annotation challenges in real underwater monitoring applications. The code is available at: https://github.com/hheeaavveenn/USHED .
Keywords:
Underwater object detection
Sparse annotation
Pseudo-label generation
Self-supervised feature enhancement
Hydroid ecological monitoring
Journal
IF:
3.1
Papers:
2.7K
Citations:
2.7K