基于卷积神经网络的SAR图像舰船分类  

SHIP CLASSIFICATION OF SAR IMAGE BASED ONCONVOLUTIONAL NEURAL NETWORKS

作　　者：陈玮 刘坤[1] Chen Wei;Liu Kun(School of Information Engineering,Shanghai Maritime University,Shanghai 201306,China)

机构地区：[1]上海海事大学信息工程学院,上海201306

出　　处：《计算机应用与软件》2024年第7期159-164,183,共7页Computer Applications and Software

基　　金：航空科学基金项目(201955015001)。

摘　　要：针对合成孔径雷达图像中斑点噪声导致图像分类准确率低的问题,提出一种基于改进VGG16的分类算法。在卷积层中加入一层注意力层,专注于重要特征,抑制不重要特征,从而抑制斑点噪声。在目标函数中引入Fisher损失函数,用该函数对特征的类内距离和类间距离进行约束,从而使得由于斑点噪声所造成的分类错误减少。通过实验可知,相比于改进前的网络,分类准确率提高了5.63百分点,有效改善了因为斑点噪声所造成的分类准确率低的问题。In view of the problem that speckled noise in synthetic aperture radar(SAR)image leads to low accuracy of image classification,a classification algorithm based on improved VGG16 is proposed.A layer of attention was added to the convolution layer to focus on important features and suppress unimportant features,so as to suppress speckled noise.The Fisher loss function was introduced in the objective function,which was used to restrain the within-class and between-class distance of the feature,so as to reduce the classification errors caused by speckle noise.Through the experiments,it can be seen that the classification accuracy is improved by 5.63 percentage points,compared with the original network,which can effectively improve the problem of low classification accuracy caused by speckled noise.

关 键 词：卷积神经网络 图像分类 注意力机制 Fisher线性判别准则 合成孔径雷达 斑点噪声 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]