Scholarlink

<p>Marine debris poses a significant threat to marine ecosystems, necessitating efficient and accurate detection methods. Autonomous underwater vehicles (AUVs) offer a promising solution, but they require lightweight, real-time algorithms for debris detection. This paper introduces DRPI-DEIM, a novel lightweight debris detection model based on the DEIM framework, incorporating four architectural modules plus a distillation-based compact variant: an Information Integration Attention (IIA) mechanism for precise spatial positioning, a Progressive-Wavelet Transform Enhancement (PWTE) Module for addressing image degradation, a Dimension-Aware Diffusion Network (DADN) for enhanced multi-scale feature representation, and a RepGhost Efficient Layer Aggregation Network (RGELAN) for reduced computational cost. Evaluated on the Trash-ICRA19 dataset, DRPI-DEIM achieved an AP of 82.2 ± 0.2% and an AP50 of 97.9 ± 0.1%, with reductions in parameters (48.1%) and computational cost (58.9%). It also achieved 120 ± 1.3 FPS on the Jetson AGX Orin and 48 ± 1.0 FPS on the Jetson Xavier NX, significantly improving energy efficiency, demonstrating its potential for deployment on resource-constrained underwater robots.</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

Enhancing underwater debris detection: a dimension-aware diffusion and progressive feature enhancement approach

  • 期刊论文

  • 发表日期：

    2026年6月16日

查看全文
  • Yongjie Yu,
  • Hui Chen,
  • Shunxiang Zhang,
  • Bin Ge,
  • Tao Fu

摘要

Marine debris poses a significant threat to marine ecosystems, necessitating efficient and accurate detection methods. Autonomous underwater vehicles (AUVs) offer a promising solution, but they require lightweight, real-time algorithms for debris detection. This paper introduces DRPI-DEIM, a novel lightweight debris detection model based on the DEIM framework, incorporating four architectural modules plus a distillation-based compact variant: an Information Integration Attention (IIA) mechanism for precise spatial positioning, a Progressive-Wavelet Transform Enhancement (PWTE) Module for addressing image degradation, a Dimension-Aware Diffusion Network (DADN) for enhanced multi-scale feature representation, and a RepGhost Efficient Layer Aggregation Network (RGELAN) for reduced computational cost. Evaluated on the Trash-ICRA19 dataset, DRPI-DEIM achieved an AP of 82.2 ± 0.2% and an AP50 of 97.9 ± 0.1%, with reductions in parameters (48.1%) and computational cost (58.9%). It also achieved 120 ± 1.3 FPS on the Jetson AGX Orin and 48 ± 1.0 FPS on the Jetson Xavier NX, significantly improving energy efficiency, demonstrating its potential for deployment on resource-constrained underwater robots.