Scholarlink

<p>Surface target detection plays a crucial role in situational awareness for intelligent unmanned surface systems. However, most detection methods suffer from limitations in both detection speed and accuracy. To address these issues, we replace the focus layer with a Wavelet Extraction (WE) module on YOLOv10n to enhance target feature extraction and suppress background interference. We also modify the backbone network by adding a Cross-Stage Partial Local Networks–Wavelet Multiscale Transformer (CSP-WMT) module, which strengthens connectivity within the backbone and improves global information modelling. Additionally, we have designed and integrated a Wavelet Multi-Scale Attention (WMSA) mechanism into the neck network to enhance multi-scale feature detection capability. WE-YOLO is evaluated on the Chinese Ship Dataset (CSD) and a self-built dataset, achieving around 3.7% better results than YOLOv10n. Meanwhile, it achieves a reduction of around 27.2% in parameters and a decrease of 34.5% in floating-point operations (FLOPs) compared to YOLOv10n.</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

Surface Target Detection Method Based on YOLO with Wavelet Analysis

  • 期刊论文

  • 发表日期：

    2026年6月29日

查看全文
  • Jie Shi,
  • Shichao Hu,
  • Yang Gu,
  • Yuankun Peng,
  • Weibo Zhong

摘要

Surface target detection plays a crucial role in situational awareness for intelligent unmanned surface systems. However, most detection methods suffer from limitations in both detection speed and accuracy. To address these issues, we replace the focus layer with a Wavelet Extraction (WE) module on YOLOv10n to enhance target feature extraction and suppress background interference. We also modify the backbone network by adding a Cross-Stage Partial Local Networks–Wavelet Multiscale Transformer (CSP-WMT) module, which strengthens connectivity within the backbone and improves global information modelling. Additionally, we have designed and integrated a Wavelet Multi-Scale Attention (WMSA) mechanism into the neck network to enhance multi-scale feature detection capability. WE-YOLO is evaluated on the Chinese Ship Dataset (CSD) and a self-built dataset, achieving around 3.7% better results than YOLOv10n. Meanwhile, it achieves a reduction of around 27.2% in parameters and a decrease of 34.5% in floating-point operations (FLOPs) compared to YOLOv10n.