摘要
针对水下目标检测中目标对比度低以及水下图像多尺度问题,提出了高频增强网络与特征金字塔(FPN)融合的水下目标检测算法,以提高对水下目标边缘、轮廓信息以及目标底层信息的提取。首先引入八度卷积将卷积层的输出特征按频率分解,将主干网络提取到的特征图进行高、低频信息分离,鉴于水下目标的轮廓信息和噪声信息均包含于高频特征中,在高频信息通道中引入通道信息具有自适应增强特点的通道注意力机制,形成了一种高频增强卷积,以达到增强有用轮廓特征信息和抑制噪声的目的;其次,将增强的高频特征分量融入FPN的浅层网络中,提高原FPN对水下多尺度目标的特征表示能力,缓解多尺度目标漏检问题。最后,将所提方法与基线算法Faster R-CNN融合,在全国水下机器人大赛提供的数据集中进行实验。结果表明:改进算法识别准确率达到78.83%,相比基线提升2.61%,与其他类型目标检测算法相比,依然具备精度和实时检测优势,证明了从特征图频域角度提升前景和背景对比度的有效性。
Aiming at the problem of low target contrast and multi-scale underwater images in underwater target detection,an underwater target detection algorithm based on the fusion of high-frequency enhanced network and Feature Pyramid Networks(FPN)is proposed.The algorithm improves the extraction of underwater target edge,contour information and target underlying information.Firstly,octave convolution is introduced to decompose the output features of the convolution layer by frequency,and the feature maps extracted by the backbone network are separated from high-frequency and low-frequency information.Since the contour information and noise information of underwater targets are contained in high-frequency features,Squeeze-and-Excitation Network with adaptive enhancement characteristics is introduced into the high-frequency information channel,and a high-frequency enhanced convolution is formed.It can achieve the purpose of enhancing useful contour feature information and suppressing noise.Secondly,the enhanced high-frequency feature components are integrated into the shallow network of FPN.It improves the feature representation ability of the original FPN for underwater multi-scale targets and alleviates the problem of missed detection of multi-scale targets.Finally,the proposed method is fused with the baseline algorithm Faster R-CNN,and the experiment is carried out on the dataset provided by the National Underwater Robot Competition.The results show that the recognition accuracy of the improved algorithm reaches 78.83%,which is 2.61%higher than the baseline.Compared with other types of target detection algorithms,it still has advantages of accuracy and real-time detection.The effectiveness of improving foreground and background contrast from the perspective of feature map frequency domain is demonstrated.
作者
乔美英
赵岩
史建柯
史有强
Qiao Meiying;Zhao Yan;Shi Jianke;Shi Youqiang(School of Electrical and Engineering,Henan Polytechnic University,Jiaozuo 454000,China)
出处
《电子测量技术》
北大核心
2023年第13期146-154,共9页
Electronic Measurement Technology
基金
国家自然科学基金(41672363)
河南省科技攻关项目(222102220076)资助
关键词
深度学习
水下目标检测
小目标检测
特征金字塔
八度卷积
通道注意力
deep learning
underwater target detection
small target detection
feature pyramid network
octave convolution
channel attention mechanism
作者简介
乔美英,博士,副教授,主要研究方向为目标识别、故障诊断、MWD数据分析与处理。E-mail:qiaomy@hpu.edu.cn;赵岩,硕士研究生,主要研究方向为计算机视觉,目标检测。E-mail:zhaoyanjoin@gmail.com;史建柯,硕士研究生,主要研究方向为计算机视觉,目标检测。E-mail:abettertomorrow2@foxmail.com
