【作者】 王曦；

【导师】 汪学刚；

【作者基本信息】 电子科技大学 ， 信息与通信工程， 2022， 硕士

【摘要】 合成孔径雷达(Synthetic Aperture Radar,SAR)具备全天时、全天候捕获高分辨遥感数据的能力。因此,已成为军事应用场景中重要的探测工具。然而,SAR图像的丰富特征信息会由于对目标位置敏感、噪声干扰、样本匮乏等多种因素而变得难以获取。SAR图像解译的关键则在于从各种复杂场景中提取出有用的SAR图像特征信息以完成特定的目标任务。自动目标识别(Automatic Target Recognition,ATR)技术在SAR图像解译方面具有至关重要的作用。随着对深度学习理论的深入研究,ATR技术在实际应用方面有了更近一步的发展。一个标准的SAR ATR系统主要包含检测、判别以及分类三个阶段,而分类是ATR技术中重要的研究部分。为了提高分类性能,特征的提取和分类器的选择已成为SAR图像目标识别中探究的热点和难点。本文主要从目标特征提取的角度出发,针对多种扩展场景、相干斑噪声干扰场景以及训练样本匮乏场景,对基于深度学习的SAR图像目标识别方法进行了研究,同时,开展了相关实验对模型性能进行评估。主要研究工作和成果如下:(1)针对多种扩展场景下的SAR图像目标识别问题,提出了一种基于动态感知注意力网络的SAR图像目标识别方法。该网络在不增加网络深度和宽度的情况下,通过动态感知卷积层以及通道和空间的注意力机制来提升基于卷积神经网络模型的特征表现能力,从而可实现多种扩展场景下的稳健性目标特征提取。(2)针对相干斑噪声干扰场景下的SAR图像目标识别问题,提出了一种基于去噪任务辅助的深度注意SAR图像目标识别方法。该网络是一种去噪和分类相耦合的多尺度残差注意网络。在去噪子网络中,通过利用空洞卷积实现多尺度特征提取,并通过注意力机制进行自适应特征通道挑选,从而可以有效地学习噪声信息;为了提取更多具有判别力的目标特征,在分类子网络中嵌入了一种跨维度交互的注意力机制,以进一步提升相干斑噪声干扰条件下的目标识别性能。(3)针对小样本场景下的SAR图像目标识别问题,提出了一种基于多任务表征学习的小样本SAR图像目标识别方法。一方面,该网络模型旨在同时具有感知输入变换、认识自身身份以及定义类别边界的能力;另一方面,该网络模型也可以提取目标的形态学特征,并通过通道注意力机制实现特征精炼。该网络的强大特征学习能力为小样本条件下的特征提取提供了保障。更多还原

【Abstract】 Synthetic aperture radar(SAR)has the capability to capture high-resolution remote sensing data in day-and-night,all-weather,thus having become an important detection tool in military application scenarios.However,the rich feature information of SAR images will become difficult to obtain due to various factors,such as sensitivity to target location,noise corruption,and insufficient samples.The key to SAR image interpretation is to extract useful SAR image feature information from various complex scenes to accomplish specific targeted tasks.Automatic target recognition(ATR)technology plays a vital role in SAR image interpretation.With the intensive research of deep learning theory,ATR technology has been further developed in practical applications.A standard SAR ATR system mainly includes three stages: detection,discrimination and classification,in which classification is an important research stage in ATR technology.In order to enhance the classification performance,feature extraction and classifier selection have become hot and difficult points in SAR image target recognition.From the perspective of target feature extraction,this thesis mainly studies the deep learning-based SAR image target recognition methods for a variety of extended scenes,speckle noise interference scenes and few-shot scenes.In the meanwhile,relevant experiments are carried out to evaluate the performance of the models.The main research work and achievements are as follows:(1)Aiming at the problem of SAR image target recognition in various extended scenes,a SAR image target recognition method based on dynamic perception attention network is proposed.The network improves the feature representation ability of the convolutional neural network-based model through the dynamic perception convolution layer and spatial and channel-wise attention mechanism without increasing the depth and width of the network,so as to realize robust target feature extraction in a variety of extended scenarios.(2)With regard to the problem of SAR image target recognition under speckle noise corruption,a deep attention SAR image target recognition method assisted by despeckling task is proposed.The network is a despeckling and classification cascaded multi-scale residual attention network.In the despeckling sub-network,the multi-scale feature extraction is realized by using dilated convolution,and the adaptive feature channel selection is performed by the attention mechanism,thereby learning the noise information effectively.In order to extract more discriminative target features,a cross-dimensional interactive attention mechanism is embedded in the classification sub-network to further improve the target recognition performance under speckle noise corruption.(3)Considering the problem of SAR image target recognition under the condition of a few samples,a few-shot SAR image target recognition method based on multi-task representation learning is proposed.On the one hand,the network aims to have the ability to perceive the input transformation,identity itself and class discrimination simultaneously.On the other hand,the network can also extract the morphological features of the target and achieve feature refinement through the channel attention mechanism.The powerful feature learning ability of the network provides a guarantee for feature extraction under the condition of small samples.更多还原