摘要
深度学习已成功地应用于合成孔径雷达(synthetic aperture radar,SAR)图像的解释中,并取得了最新的成果。然而,目前的深度学习算法受SAR图像斑点噪声和俯仰角变化较大的影响,解释精度不佳。本文提出一种基于胶囊神经网络的SAR图像目标识别方法。该方法首先对MSTAR数据集使用灰度化和双线性插值的中心裁剪方法统一图像尺寸以完成对数据的预处理。然后利用两个卷积层生成胶囊基本单元,第一个卷积层将像素强度转换为局部特征探测器的活动并将其用作初级胶囊的输入,第二个卷积层生成主胶囊。最后通过基于路由算法的全连接层实现了分类任务。经过科学实验和对比,该方法网络结构简单,参数量小,准确率高。
Deep learning has been successfully applied to the interpretation of synthetic aperture radar(SAR) images and made the latest achievements. However, the current deep learning algorithm is affected by the large variation of speckle noise and pitch angle in SAR image, and the interpretation accuracy is poor. This paper presents a method for SAR image target recognition based on capsule neural network. In this method, the image size of MSTAR data set is unified by the center clipping method of graying and bilinear interpolation to complete the data preprocessing. Then, two convolutional layers are used to generate the basic unit of the capsule. The first convolutional layer converts the pixel intensity into the activity of the local feature detector and uses it as the input of the primary capsule. The second convolutional layer generates the main capsule. Finally, ten classification tasks are realized through the full connection layer based on routing algorithm. Through scientific experiments and comparison, the network structure of this method is simple, the number of parameters is small, and the accuracy is high.
作者
王璐
温显斌
WANG Lu;WEN Xianbin(School of Computer Science and Engineering,Tianjin University of Technology,Tianjin 300384,China)
出处
《天津理工大学学报》
2021年第6期36-40,共5页
Journal of Tianjin University of Technology
基金
天津市研究生科研创新项目(2019YJSS049)。
关键词
胶囊神经网络
路由算法
合成孔径雷达图像
目标识别
capsule neural network
routing algorithm
synthetic aperture radar(SAR)image
target recognition
