Scholarlink

<p>In marine resource exploration and other maritime activities, underwater object detection is critical. Although YOLO-series models dominate, they still face many challenges in complex underwater environments: feature loss of occluded small objects, feature confusion in dense areas, blurry low-contrast images from severe noise and light attenuation, and insufficient lightweight design. To address these issues, this paper proposes a <b>L</b>ightweight <b>U</b>nderwater <b>O</b>bject <b>D</b>etection model (<b>LUOD-YOLO</b>). This model proposes the <b>D</b>ynamic <b>L</b>ocal-<b>G</b>lobal Feature Fusion <b>A</b>ttention Mechanism module (<b>DLGA</b>) to enhance the feature extraction and detection capabilities for occluded small objects through multi-scale attention mechanism and dynamic local–global dual-path fusion. What’s more, this model designs lightweight <b>D</b>ual <b>P</b>ath <b>C</b>hannel Rearrangement module (<b>DPC</b>) to process input feature maps in two parts and concatenate them via channel rearrangement, thus effectively addressing feature confusion in overlapping areas of marine organisms. Moreover, the Lightweight <b>C</b>NN-based <b>C</b>ross-scale <b>F</b>eature <b>F</b>usion (<b>CCFF</b>) is introduced, which enhances the model’s ability to recognize blurry low-contrast images. Experiments were conducted on the Under water-object-detection-dataset (UOD dataset) and the Detecting Underwater Objects dataset (DUO dataset). The results on the UOD dataset show that compared with YOLOv8n, LUOD-YOLO not only improves the mAP50:95 (54.2%) by 0.2%, but also the model parameters and computational amount are 45.7% and 28.4% lower than those of YOLOv8n, respectively.&#xa0;Our code is available at <a href="https://github.com/JEFfersusu/LUOD-YOLO">https://github.com/JEFfersusu/LUOD-YOLO</a></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

LUOD-YOLO: a lightweight underwater object detection model based on dynamic feature fusion, dual path rearrangement and cross-scale integration

  • 期刊论文

  • 发表日期：

    2025年10月14日

查看全文
  • Chengze Lv,
  • Weichao Pan

摘要

In marine resource exploration and other maritime activities, underwater object detection is critical. Although YOLO-series models dominate, they still face many challenges in complex underwater environments: feature loss of occluded small objects, feature confusion in dense areas, blurry low-contrast images from severe noise and light attenuation, and insufficient lightweight design. To address these issues, this paper proposes a Lightweight Underwater Object Detection model (LUOD-YOLO). This model proposes the Dynamic Local-Global Feature Fusion Attention Mechanism module (DLGA) to enhance the feature extraction and detection capabilities for occluded small objects through multi-scale attention mechanism and dynamic local–global dual-path fusion. What’s more, this model designs lightweight Dual Path Channel Rearrangement module (DPC) to process input feature maps in two parts and concatenate them via channel rearrangement, thus effectively addressing feature confusion in overlapping areas of marine organisms. Moreover, the Lightweight CNN-based Cross-scale Feature Fusion (CCFF) is introduced, which enhances the model’s ability to recognize blurry low-contrast images. Experiments were conducted on the Under water-object-detection-dataset (UOD dataset) and the Detecting Underwater Objects dataset (DUO dataset). The results on the UOD dataset show that compared with YOLOv8n, LUOD-YOLO not only improves the mAP50:95 (54.2%) by 0.2%, but also the model parameters and computational amount are 45.7% and 28.4% lower than those of YOLOv8n, respectively. Our code is available at https://github.com/JEFfersusu/LUOD-YOLO