【作者】 李响；

【导师】 叶秀芬；

【作者基本信息】 哈尔滨工程大学 ， 控制科学与工程， 2020， 硕士

【摘要】 随着声纳系统的不断完善以及水下智能机器人的出现,水下目标的识别与分类领域的研究也在不断深入。在民用和军事领域水下声纳图像的分类均得到了广泛的应用,在军事上有助于发现潜艇,失事飞机等。在民用上有助于发现鱼群,检测坝底的完好程度等。传统的水下目标分类方法主要是用特征提取的方法来进行分类,但这类方法往往存在应用局限性较高的缺点。本文通过对声纳成像原理以及声纳图像的特点进行分析,采用深度学习以及迁移学习的方法对声纳图像进行分类,并根据深度学习需要大量样本的特性,将迁移学习方法与深度神经网络相结合,同时根据迁移学习方法中,领域距离越近越有利于知识的迁移这一基本原则,采用了图形风格迁移以及生成对抗网络的方法生成与侧扫声呐图像相似的仿侧扫声呐图像,作为源领域,随后再使用少量真实样本进行迁移,从而达到高准确率的侧扫声呐图像目标分类的目的。论文的具体工作如下:(1)对传统的目标分类方法进行了研究。首先对声纳图像的特点进行了分析,得知声纳图像普遍有清晰度差,分辨率低,噪声严重的问题,传统的水下声纳图像分类方法主要是针对声纳图像的特点,由常规光学图像方法改进而来,主要包括预处理,特征提取,特征分类等几个步骤,对声纳图像的去噪以及灰度校正风发进行了实验及分析,进一步提升了图像的质量。分别采用了高斯马尔科夫随机场以及水平集的方法进行了图像的特征提取,通过提取的特征进而对图像进行分类。(2)针对声纳图像的特点以及传统方法的缺点,采用了深度学习的方法对图像的特征进行自动提取,对经典的卷积神经网络进行分析及对比,同时结合了迁移学习的方法,分析了领域间距离对迁移效率的影响,提出了基于缩小样本领域距离的迁移学习方法。(3)针对源领域样本仿声呐化处理,提出一种融合显著性检测及风格迁移网络的转换方法。方法通过风格迁移网络将常规的光学图像转换成侧扫声纳图像风格,并针对转换效果较差的图像进行了处理,主要采用了显著性检测算法,将该算法与风格迁移方法相融合,解决了由于背景亮度过高导致的风格迁移效果差的问题,将通过风格迁移方法生成的仿声纳图像作为源领域样本训练分类网络,在此基础上再使用迁移学习方法将其迁移到真实侧扫声呐图像样本分类任务中,最后给出了实验分析结论。(4)同时,针对源领域样本仿声呐化处理,还提出一种基于改进WGAN的仿真侧扫声呐图像生成方法,不依赖于常规光学图像,直接从真实侧扫声呐图像学习数据中像素分布,从而进行图像的生成,由于侧扫声呐图像数据集样本数量稀少,数据分布可能存在较大偏差,因而直接应用WGAN时,可能存在因参数设置不当导致的生成图像真实性差的问题,本文提出将梯度正则项融入到WGAN的损失函数中,自动的进行参数范围的确定,保证图像生成质量。通过迁移实验,证明本方法有效性。更多还原

【Abstract】 With the continuous improvement of sonar systems and the emergence of underwater intelligent robots,research in the area of underwater target recognition and classification is also continuously deepening.The classification of underwater sonar images in the civilian and military fields has been widely used,and it is helpful to find submarines and crashed aircraft in the military.For civilian use,it is helpful to find fish schools and detect the integrity of the dam bottom.Traditional underwater target classification methods mainly use feature extraction methods for classification,but such methods often have the disadvantage of high application limitations.This article analyzes the principles of sonar imaging and the characteristics of sonar images,uses deep learning and transfer learning to classify sonar images,and according to the characteristics of deep learning requiring a large number of samples,transfer learning methods are compared with deep neural networks.In combination,according to the basic principle of transfer learning methods,the closer the field distance is,the more conducive to the transfer of knowledge.The method of graphic style transfer and the generation of adversarial networks is used to generate a pseudo-side-scan sonar image similar to the side-scan sonar image.Source domain,and then use a small number of real samples to migrate,so as to achieve the purpose of high-accuracy side-scan sonar image target classification.The specific work of the paper is as follows:(1)The traditional target classification method is studied.Firstly,the characteristics of the sonar image are analyzed.It is learned that the sonar image generally has the problems of poor definition,low resolution,and serious noise.The traditional underwater sonar image classification method mainly focuses on the characteristics of the sonar image.The conventional optical image method is improved,mainly including several steps such as preprocessing,feature extraction,feature classification,etc.The experiments and analysis on the denoising of the sonar image and the gray correction wind are carried out to further improve the image quality.Gauss-Markov random field and level set methods were used to extract image features,and the images were classified based on the extracted features.(2)Aiming at the characteristics of sonar images and the shortcomings of traditional methods,a deep learning method is used to automatically extract the features of the images,and the classic convolutional neural network is analyzed and compared.At the same time,the method of transfer learning is used to analyze The influence of the distance between domains on the transfer efficiency is proposed,and a transfer learning method based on reducing the domain distance of the sample is proposed.(3)Aiming at the sonar-like processing of source domain samples,a conversion method is proposed which combines saliency detection and style transfer network.Methods The conventional optical image was converted into a side-scan sonar image style through a style transfer network,and the images with poor conversion effects were processed.The saliency detection algorithm was mainly used.This algorithm was combined with the style transfer method to solve the problem.In order to solve the problem of poor style transfer caused by too high background brightness,the imitation sonar image generated by the style transfer method is used as the source domain sample to train the classification network,and then transfer learning method is used to migrate it to the real side scan.In the sonar image sample classification task,the experimental analysis conclusion is finally given.(4)At the same time,for the simulation of sonarization in the source domain samples,an improved method for generating side scan sonar images based on improved WGAN is also proposed,which does not rely on conventional optical images and directly learns the pixel distribution in the real side scan sonar images.In order to generate images,due to the small number of samples in the side-scan sonar image data set,there may be large deviations in the data distribution.Therefore,when directly applying WGAN,there may be a problem of poor authenticity of the generated images due to improper parameter settings.The regular term is integrated into the loss function of WGAN,and the parameter range is automatically determined to ensure the image generation quality.The migration experiment proves the effectiveness of this method.更多还原