深圳大学学报（理工版）2025，Vol.42Issue(4)：428-436,9.DOI:10.3724/SP.J.1249.2025.04428
基于自注意力和改进金字塔的水下小目标检测
Underwater small objects detection based on self-attention and improved pyramid network
摘要
Abstract
Addressing the challenges of limited feature information and decreased detection accuracy in underwater small object detection tasks due to underwater environmental effects,we propose an underwater small object detection algorithm SF-Bi-YOLOv8 based on self-attention mechanism and improved pyramid network.This algorithm,built upon the YOLOv8 framework,firstly enhances the global feature extraction capability using the Swin-Fa module to modify the last C2f layer of YOLOv8 backbone network,thereby improving the model's ability to detect small objects.Secondly,we replace the feature pyramid in the neck network with a simplified version of weighted bidirectional feature pyramid network(BiFPN)structure to better learn multi-scale features.Finally,we reconstruct the loss function of the detection head using linear interval mapping Focaler-IoU to focus on challenging samples and expedite bounding box regression.Experimental results demonstrate that on the detecting underwater objects(DUO)dataset,the improved algorithm achieves an PMA@0.50(mean average precision with intersection over union(IoU)threshold is 0.50)of 0.862,representing a 0.023 increase over the original YOLOv8,while PMA@[0.50:0.95](mean average precision with Idl threshold is from 0.50 to 0.95)improves by 0.036.This indicates higher detection accuracy,which can provide valuable reference for underwater target detection tasks.关键词
人工智能/水下目标检测/深度学习/YOLOv8/注意力机制/特征金字塔Key words
artificial intelligence/underwater object detection/deep learning/YOLOv8/attention mechanism/feature pyramid分类
信息技术与安全科学引用本文复制引用
杜睿山,王紫珊,孟令东,井远光..基于自注意力和改进金字塔的水下小目标检测[J].深圳大学学报（理工版）,2025,42(4):428-436,9.基金项目
National Key Research and Development Program of China(2022YFE0206800) 国家重点研发计划资助项目(2022YFE0206800) （2022YFE0206800）
