基于YOLOv8的轻量化水下目标检测算法  

Lightweight Underwater Target Detection Algorithm Based on YOLOv8

作　　者：梁秀满 赵佳阳 于海峰 LIANG Xiuman;ZHAO Jiayang;YU Haifeng(College of Electrical Engineering,North China University of Science and Technology,Tangshan 063210,China)

机构地区：[1]华北理工大学电气工程学院,河北唐山063210

出　　处：《红外技术》2024年第9期1015-1024,共10页Infrared Technology

基　　金：河北省自然科学基金资助(D2024209006);河北省教育厅科学研究项目资助(QN2024147)。

摘　　要：针对复杂水下环境导致水下目标检测时出现误检、漏检以及检测效率低等问题,提出了一种改进YOLOv8模型的轻量化水下目标检测算法。首先,为了改善颈部网络特征融合不足的问题,将YOLOv8的颈部网络融合(Bidirectional Feature Pyramid Network,BiFPN)双向特征金字塔结构,提高小目标层的检测效果;其次,针对网络中卷积模块参数量大和计算复杂度高的问题,设计了一种自适应注意力下采样(Adaptive-Attention Down-Sampling,AADS)模块,将主干网络中的卷积模块替换为AADS模块,降低模型参数量和计算量;最后,引入大可分离核注意力机制(Large Separable Kernel Attention,LSKA),强化特征提取能力,使模型能够更精确地关注重要信息,提高目标检测精度。将改进的网络在水下目标检测数据集中进行实验,改进后的算法与YOLOv8相比,平均检测精度提升了1.4%,模型计算复杂度降低了15.9%,模型参数量减少了43.3%,使检测精度和检测速度之间达到了很好的平衡。To address the problems of misdetection,omission detection,and low detection efficiency when detecting underwater targets due to the complex underwater environment,a lightweight underwater target detection algorithm with an improved YOLOv8 model is proposed.First,to ameliorate the problem of insufficient feature fusion in the neck network,the neck network of YOLOv8 is fused with a BiFPN bidirectional feature pyramid structure to improve the detection of the small target layer.Second,to address the problem of the large number of parameters of the convolution module in the network and high computational complexity,an Adaptive-Attention Down-Sampling(AADS)module is designed to replace the convolution module in the backbone network to reduce the number of model parameters and amount of computation.Finally,Large Separable Kernel Attention(LSKA)is introduced to strengthen the feature extraction capability such that the model can focus on important information more accurately and improve target detection accuracy.The experimental results show that in the underwater target detection dataset,the improved algorithm improves the average detection accuracy by 1.4%,reduces the number of model parameters by 43.3%,and reduces the computational complexity of the model by 15.9%when compared with YOLOv8.This realizes a good balance between detection accuracy and detection speed.

关 键 词：YOLOv8 水下目标检测 大可分离核注意力机制 轻量化 多尺度特征融合 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]