基于动态特征增强的水下鱼类目标检测

打开文本图片集

关键词：水下检测；动态增强；深度学习；卷积神经网络；YOLOv8n

中图分类号：TN911.7;TP391 文献标识码：A doi：10.37188/OPE.20263403.0481 CSTR：32169.14.OPE.20263403.0481

Dynamic feature-augmented for detection of underwater fish targets

ZHU Xiaolong 1，2* ， CHEN Yuwei1，WANG Jiayu1，GUO Haitao 1 ， CHEN Xiangzi 1，2 （1. Yazhou Bay Inmovation Institute， Hainan Tropical Ocean University， Sanya 572O22， China; 2. College of Marine Science and Technology， Hainan Tropical Ocean University， Sanya 572022，China） * Corresponding author，E-mail： zhuxiaolonglong@sina. cn

Abstract： Eficient monitoring of underwater fish species is essential for marine ecosystem conservation， biodiversity assessment，and the sustainable management of aquatic resources. To address reductions in detection robustness and eficiency under complex underwater conditions，a dynamic feature enhancement model，termed Fish Detection Network YOLO （FDN-YOLO），is proposed based on the YOLOv8n framework. First，a Multi-scale Deformable Receptive Field （MDRF） module is incorporated to adaptively regulate the efective receptive field，thereby improving backbone representations of fish targets with diverse shapes and scales. Second，a lightweight down-sampling module， Lite Space-to-Depth Depthwise Separable （Lite SPD-DS），is designed to preserve fine-grained spatial cues during subsampling while maintaining low computational cost. Third，an Adaptive IoU-aware Varifocal Loss （AIVF Loss） is intro duced by integrating adaptive IoU weighting with Varifocal Loss to strengthen the learning of high-quality localization samples and mitigate training bias caused by class and sample imbalance.Experiments on the TF-DET dataset show that FDN-YOLO increases mAP50 and mAP50：95 by 2.8% and 2.1% ，respectively，while reducing parameters and computational complexity by 13.3% and 16.0% . Additional comparative and generalization experiments further confirm that FDN-YOLO achieves a favorable trade-off among accuracy，eficiency，and robustness，highlighting its potential for ecological monitoring and datadriven marine resource management.

Key words： underwater detection；dynamic enhancement; deep learning；convolutional neural network; YOLOv8n

1引言

在智能化海洋观测系统中，水下鱼类检测作为核心任务之一，广泛应用于生态调查、水产养殖管理及环境评估等领域[1-2]。（剩余19717字）