OGRA-YOLOv8: Overlapping Gridded and Rhombus Attention for Underwater Object Detection

Underwater object detection (UOD) is vital for marine exploration, ecosystem monitoring, and aquaculture management. However, reliable UOD remains challenging due to optical degradation, low illumination, and the prevalence of small or densely distributed targets. Existing detectors often struggle to balance accuracy with real-time performance, limiting their robustness in practice. To address these issues, we propose OGRA-YOLOv8, an enhanced YOLOv8 framework tailored for UOD. We design two novel attention modules: Overlapping Gridded Attention (OGA) that strengthens local feature extraction while reducing the cost of self-attention, and Rhombus Attention (RA) to capture long-range dependencies and complement OGA. In addition, a small-object detection head (YOLOv8-P2) and WIoU v2 loss are integrated to improve bounding box regression under uneven and multi-scale target distributions. Extensive experiments on the DUO and URPC2020 datasets show that OGRA-YOLOv8 consistently surpasses state-of-the-art YOLO variants and other competitive detectors, achieving significant gains in accuracy while preserving real-time efficiency. We make the code public at: https://github.com/LYJ0126/OGRA-YOLOV8.