计算机科学
算法
人工智能
计算机视觉
噪音(视频)
跟踪(教育)
水下
信号处理
杂乱
过程(计算)
作者
Deyi Yang,Guangyu Li,Yanxin Li
出处
期刊:
日期:2026-05-15
卷期号:: 746-749
标识
DOI:10.1109/ainit70033.2026.11557730
摘要
Underwater target detection suffers from several well-defined problems: ambiguous target recognition, high false-negative rates for small targets, poor detection accuracy, and sub-optimal real-time performance in complex underwater environments. Therefore, this paper presents an improved SPED-YOLOv10n model based on YOLOv10n, with two clear, complementary improvements. First, spatial-to-depth convolution is applied in the Backbone and Neck architectures, converting spatial information into depth data so as to minimize information loss and naturally better handle small targets and low-resolution images. Second, an efficient multi-scale attention mechanism (EMA) is introduced which retains channel-specific information while keeping computational overhead low. More importantly, SPD-Conv solves the local detail dependency problem caused by down sampling-induced feature loss in EMA, and the EMA itself uses the retained details to reinforce feature representation via spatial attention. Experimental validation on the DUO underwater biological dataset clearly and convincingly shows that the improved model obtains 76.7%m AP@0.5and52.5%m AP@0.5:0.95, which are 2.4% and 1.9% improvements over the original model, respectively. Therefore, its overall performance outstrips all existing models considered here: YOLOv9t, YOLOv12n, YOLOv6n, YOLOv3-tiny and Water-YOLO. More importantly, the ablation experiments demonstrate that the synergistic optimization of SPD-Conv and EMA attention perfectly balances detection accuracy with computational efficiency, making it an extremely reliable choice for underwater applications.
科研通智能强力驱动
Strongly Powered by AbleSci AI