Scholarlink

<p>To address high environmental complexity, limited model performance, and low computational efficiency in underwater image analysis, this paper proposes SGL-YOLO, a lightweight underwater object detection algorithm based on an improved YOLOv8. The backbone network incorporates a novel SCA-C2f module to capture long-range dependencies and accurately extract object directional and positional features. A Global-Local Cross-layer Aligned BiFPN (GLCA-BiFPN) is designed in the neck to enhance feature fusion and small object detection while reducing parameters and computation. A Lightweight Shared Convolutional Detection (LSCD) head is introduced for efficiency, and WIoU loss is adopted to improve bounding box regression. On the DUO dataset, SGL-YOLO reduces parameters by 45.2% and computation by 27.2%, while increasing mAP by 1.1% compared to YOLOv8. The model effectively addresses false positives, missed detections, and occlusions in complex underwater environments. Compared to existing lightweight methods, SGL-YOLO achieves higher accuracy with fewer parameters, attaining an optimal balance between detection precision and speed.</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

SGL-YOLO: Lightweight underwater object detection algorithm based on feature fusion

  • 期刊论文

  • 发表日期：

    2025年11月17日

查看全文
  • Minghui Li,
  • Huanli Pang,
  • Changhong Jiang

摘要

To address high environmental complexity, limited model performance, and low computational efficiency in underwater image analysis, this paper proposes SGL-YOLO, a lightweight underwater object detection algorithm based on an improved YOLOv8. The backbone network incorporates a novel SCA-C2f module to capture long-range dependencies and accurately extract object directional and positional features. A Global-Local Cross-layer Aligned BiFPN (GLCA-BiFPN) is designed in the neck to enhance feature fusion and small object detection while reducing parameters and computation. A Lightweight Shared Convolutional Detection (LSCD) head is introduced for efficiency, and WIoU loss is adopted to improve bounding box regression. On the DUO dataset, SGL-YOLO reduces parameters by 45.2% and computation by 27.2%, while increasing mAP by 1.1% compared to YOLOv8. The model effectively addresses false positives, missed detections, and occlusions in complex underwater environments. Compared to existing lightweight methods, SGL-YOLO achieves higher accuracy with fewer parameters, attaining an optimal balance between detection precision and speed.