UF-SIENet: Frequency-selection based image enhancement for underwater object detection
Haoyu Wang, Jinlong Li, Weidong Ji, Lin Zheng, Aodong Zhang · Computer Vision and Image Understanding · 2026
Underwater object detection (UOD) plays a vital role in marine ecological monitoring, facility inspection, and resource exploration. However, underwater images often suffer from blurriness, noise, and color distortion, severely degrading detection performance. Traditional enhancement methods prioritize visual aesthetics but neglect the needs of detection models. To bridge this gap, we propose UF-SIENet, a dedicated enhancement network for underwater detection. Central to our method is the Frequency Enhanced Low-level Knowledge Aggregation (FELKA) module, which estimates the transmission map by decomposing features into high- and low-frequency components using learnable low-pass filters. It then adaptively fuses these components to enhance structural and semantic consistency, improving detail preservation under challenging underwater conditions. To estimate background light, we integrate Underwater Background Attention Module (UBAM), which applies both channel and spatial attention, allowing the network to concentrate on informative regions while suppressing background interference. This attention-guided mechanism improves estimation robustness in scenes with uneven illumination. We further propose Blur-Guided Data Augmentation (BGDA), which utilizes blurred-region priors to guide the detection model’s attention toward ambiguous areas, thereby increasing robustness to various forms of visual degradation. Extensive experiments on the DUO and TrashCan datasets demonstrate that UF-SIENet consistently improves detection accuracy across various models, with up to 3.1% AP gain on YOLOV10-S.