基于改进MobilenetV2网络的声光图像融合水下目标分类方法  

Acoustic-optical image fusion underwater target classification method based on improved MobilenetV2

作　　者：巩文静 田杰[1,3] 李宝奇[1,3] 刘纪元 GONG Wenjing;TIAN Jie;LI Baoqi;LIU Jiyuan(Institute of Acoustics,Chinese Academy of Sciences,Beijing 100190,China;University of Chinese Academy of Sciences,Beijing 100049,China;Key Laboratory of Science and Technology on Advanced Underwater Acoustic Signal Processing,Chinese Academy of Sciences,Beijing 100190,China)

机构地区：[1]中国科学院声学研究所,北京100190 [2]中国科学院大学,北京100049 [3]中国科学院先进水下信息技术重点实验室,北京100190

出　　处：《应用声学》2022年第3期462-470,共9页Journal of Applied Acoustics

基　　金：中国科学院国防科技重点实验室基金项目(CXJJ-20S035)。

摘　　要：针对小样本条件下水下目标分类准确率低、计算资源量大的问题,提出一种声光图像融合目标分类方法。首先,对MobilenetV2网络进行改进,去掉第9层之后的网络层,并将该层卷积通道数改为128,通过Flatten层进行数据降维,增加一个全连接层得到分类结果;其次,设计一种融合网络结构,将声光图像成对输入网络进行特征提取,在中间层利用通道拼接算法实现特征图融合,使用融合特征进行目标分类。在真实数据集上对网络进行训练,结果表明,改进的MobilenetV2网络对水下目标的分类性能更好,融合网络的分类准确率相比融合前有所提高,更加适用于水下目标分类任务。To solve the problems of low accuracy and high consumption of underwater target classification under the condition of small samples,an acoustic-optical image fusion classification method is proposed.Firstly,the Mobilenetv2 network is improved by removing the network layer after layer 9 and changing the channels to 128.After the dimension reduction by flatten layer,a dense layer is added to get classification results.Secondly,a fusion network structure is designed,which inputs acoustic image and optical image in pairs for feature extraction.Then,the extracted feature maps are combined in the middle layer and used for target classification.Using the real data to train the network,the results show that the improved Mobilenetv2 network has better classification performance for underwater targets,and the accuracy of fusion network is improved compared with that before fusion,indicating the applicability in underwater target classification.

关 键 词：改进MobilenetV2 声学图像 光学图像 图像融合 水下目标分类 

分 类 号：TB566[交通运输工程—水声工程]