Scholarlink

<p>A You Only Look Once version 8 (YOLOv8)-based detection model optimized via pruning and knowledge distillation is proposed for the challenging task of detecting dynamic responses of small underwater bubbles. Underwater bubble experiments were conducted to collect dynamic image sequences, and the improved YOLOv8 model was used to detect bubbles and extract their nonlinear dynamic responses. Through model pruning and distillation, the optimized model considerably reduces model parameters and accelerates inference while maintaining high detection accuracy. The experimental detection results were validated by a coupled Eulerian-Lagrangian computational fluid dynamics simulation, which showed high consistency with the observed bubble dynamics. The proposed approach demonstrates excellent performance in recognizing small targets with nonlinear dynamic responses, shows strong robustness in complex underwater scenarios, and offers high inference efficiency and ease of deployment.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An Integrated Lightweight YOLO Framework Combining Structured Pruning, Knowledge Distillation, and Multi-Frame Linear Interpolation for Robust Real-Time Nonlinear Dynamic Response Analysis of Small-Scale Underwater Targets under Bubble Loading

  • 期刊论文

  • 发表日期：

    2025年7月22日

查看全文
  • Jiuqiang Wang,
  • Dongyan Shi,
  • Yuxin Gou,
  • Haifeng Zhang,
  • Xiongwei Cui

摘要

A You Only Look Once version 8 (YOLOv8)-based detection model optimized via pruning and knowledge distillation is proposed for the challenging task of detecting dynamic responses of small underwater bubbles. Underwater bubble experiments were conducted to collect dynamic image sequences, and the improved YOLOv8 model was used to detect bubbles and extract their nonlinear dynamic responses. Through model pruning and distillation, the optimized model considerably reduces model parameters and accelerates inference while maintaining high detection accuracy. The experimental detection results were validated by a coupled Eulerian-Lagrangian computational fluid dynamics simulation, which showed high consistency with the observed bubble dynamics. The proposed approach demonstrates excellent performance in recognizing small targets with nonlinear dynamic responses, shows strong robustness in complex underwater scenarios, and offers high inference efficiency and ease of deployment.