基于Faster-rcnn的水下目标检测算法研究  

Exploring Underwater Target Detection Algorithm Based on Faster-rcnn

作　　者：王璐 王雷欧[1] 王东辉[1] WANG Lu;WANG Leiou;WANG Donghui(Key Laboratory of Information Technology for AUVs,CAS,Institute of Acoustics,Chinese Academy of Sciences,Beijing,100190,China;University of Chinese Academy of Sciences,Beijing,100049,China)

机构地区：[1]中国科学院声学研究所,中科院水下航行器信息技术重点实验室,北京100190 [2]中国科学院大学,北京100049

出　　处：《网络新媒体技术》2021年第5期43-51,58,共10页Network New Media Technology

基　　金：国家自然科学基金(编号:61801469);中国科学院声学研究所自主部署项目(编号:ZYTS202005)。

摘　　要：对海洋资源开发的关键是实现对水下目标实时而准确的检测,但由于水介质的吸收以及悬浮粒子的散射作用,水下待测目标往往存在颜色失真、对比度低等复杂问题,这极不利于准确评估目标检测算法的性能。本文提出一种基于Faster-rcnn的水下目标检测算法,该算法以Faster-rcnn结构为主框架,将ResNet-101深度神经网络替代Faster-rcnn原本的VGG-16卷积神经网络作为特征提取和训练初始化的共享卷积网络,同时采用Water-Net网络对水下图像数据集进行增强处理,最后针对部分图像标签数据过少的问题采取了标签数据增强的方法。通过实验证明,数据集的增强性能有效提升检测算法的性能,且能满足实时检测的需求。How to achieve real-time and accurate detection of underwater targets is the key to the development of marine resources.However,due to the absorption of water media and the scattering of suspended particles,underwater targets often face problems such as color distortion and low contrast,which is extremely unfavorable for evaluating the performance of the target detection algorithm.For this reason,this paper proposes an underwater target detection algorithm based on Faster-rcnn.The algorithm takes the Faster-rcnn structure as the main framework,and then replaces the original VGG-16 convolutional neural network of Faster-rcnn with the ResNet-101 deep neural network as a shared convolutional network for feature extraction and training initialization.At the same time,Water-Net is used to enhance the underwater image data set to improve the performance of the target detection algorithm.Finally,a label enhancement method is adopted to solve the problem of too few image label data.Training experiments have proved that the performance of our algorithm has a significant improvement compared with the original data,and can meet the needs of real-time detection.

关 键 词：水下目标检测 Faster-rcnn模型 水下图像增强 Water-Net模型 数据增强 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]