武铄 王晓 张丹阳 周海波 陈家儒

摘 要：针对当前沉船侧扫声呐数据样本少、无法大量获取有标记的数据、小样本数据用于卷积神经网络模型正确识别率较低的问题，提出一种联合迁移学习和深度学习的沉船侧扫声呐图像识别方法。在不采用迁移学习的情况下，用沉船声呐样本数据集训练网络，将结果保存作为对照组。应用弱相关数据集，对网络进行训练，然后将网络参数迁移到新网络中，再使用沉船声呐样本数据集训练;应用强相关数据集重复上述步骤。在AlexNet网络结构上进行试验，试验结果表明：在未迁移的情况下，正确率为70.58%;应用弱相关数据集训练，正确率为74.51%;应用强相关数据集训练，正确率为80.39%。试验证明，迁移学习算法有利于提高小样本情况下卷积神经网络的正确率、泛化性。

关键词：侧扫声呐;深度学习;迁移学习;卷积神经网络

中图分类号：S126     文献标识码：A     文章编号：1003-5168（2021）36-0036-05

Side Scan Sonar Combined With Transfer Learning and Deep Learning Shipwreck Identification Method

WU Shuo    WANG Xiao    ZHANG Danyang    ZHOU Haibo    CHEN Jiaru

（Jiangsu Ocean University，School of Marine Technology and Geomatics，Lianyungang  Jiangsu  222005）

Abstract： Aiming at the problem that the current sunken side scan sonar data samples are few and it is impossible to obtain a large amount of labeled data，the small sample data is used for the problem of low correct recognition rate of the convolutional neural network model，and a joint migration learning and deep learning sunken side scan sonar is proposed. In the case of not using migration learning，use the shipwreck sonar sample data set to train the network and save the result as a control group; use the weakly correlated data set to train the network，and then transfer the network parameters to the new network，and then use the shipwreck sonar Sample data set training; apply the strong correlation data set to repeat the above steps. Experiments on the AlexNet network structure show that the accuracy rate is 70.58% without migration; the accuracy rate is 74.51% for training with weakly correlated data sets; and 80.39% for training with strong correlated data sets. Experiments show that the transfer learning algorithm is beneficial to improve the accuracy and generalization of the convolutional neural network in the case of small samples.

Keyword： side scan sonar; deep learning; transfer learning; convolutional neural network

海洋测绘是一切水下活动的基础，是海底调查、海底矿物勘探、海洋经济发展的前提[1]。其中，侧扫声呐技术（Side Scan Sonar，SSS）作为一种高分辨率、多用途、低成本的海洋调查设备，广泛應用于海洋、港口、河流、湖泊等水域，主要应用于海道测量，水下工程选址，飞机残骸、沉船、鱼雷、水雷、集装箱、化学品桶、锚链、沉石、浮标等水下目标的探测，以及海底底质分类、海洋调查和海洋科学研究等。而目前市场上使用的数据处理软件中，水下目标探测和识别均依靠人工判读，效率低下且精度难以保证。……