摘要合成孔径雷达图像自动目标识别(SAR ATR)是合成孔径雷达遥感的重要方向,广泛应 用于国民生产各部门。在 SAR ATR 研究中,如何找到稳健可靠的特征,在有限的有标记 样本支持下,改善人造目标分类的精度是近年来国内外研究的热点之一。而深度学习理 论可以通过非监督或半监督的学习从大量无标记样本和少量有标记样本构成的混合集 中,有效的找到目标的潜在特征。本文构建了一个两层栈式稀疏自编码器(SSAE)和一 个具有两个卷积层的卷积神经网络(CNN)学习模型,并利用 MNIST 数据库验证算法的可 实施性。构建模型后,分别通过两种算法针对 MSTAR 数据集中 SAR 图像进行处理,逐 层抽象,抽取地面车辆的潜在特征。然后从特征图、分类精度和资源消耗情况三方面进 一步分析比较两种算法的特征提取性能。通过调整深度学习模型结构,获取稳健特征用 于 SAR ATR。74535
毕业论文关键词 SAR ATR 深度学习 栈式稀疏自编码器 卷积神经网络
毕 业 设 计 说 明 书 外 文 摘 要
Title Feature Analysis of High Resolution SAR Image Based on Deep Learning Theory
Abstract Synthetic aperture radar images automatic target recognition (SAR ATR) is an important direction of the synthetic aperture radar remote sensing。 And it is also widely used in various sectors of national production。 In the research field of SAR ATR, how to find reliable and robust feature with limited marked samples supporting, and improve the accuracy of man-made object classification is one of the hottest research topic at home and abroad in recent years。 By unsupervised or semi-supervised learning, deep learning can extract the potential feature of targets effectively from a mixed set of a large amount of non-labeled samples and a small amount of labeled samples。 This paper constructs a two-stack sparse auto- encoder (SSAE) and a convolutional neural network (CNN) with two convolutional layers。 And the MNIST dataset is utilized to implement the enforceability of algorithms。 Via these two models, potential feature of ground vehicles are extracted from SAR image in MSTAR dataset with layer-wise abstraction。 There’re three aspects to analyze the features extraction performance in this article: feature map, classification accuracy and resource consumption。 And by adjusting the structure of deep learning, it can access robust feature for SAR ATR。
Keywords SAR ATR Deep Learning SSAE CNN
本科毕业设计说明书 第 I 页
目 录
1 绪论 1
1。1 研究背景和意义 1
1。2 国内外研究现状 2
1。3 论文的主要工作 4
1。4 论文结构安排 5
2 栈式稀疏自编码器基本原理及算法 6
2。1 SSAE 基本原理概述 6
2。2 SSAE 算法方案设计 11
2。3 SSAE 算法模型验证 13
2。4