【作者】 袁野；

【导师】 吴艳霞；

【作者基本信息】 哈尔滨工程大学 ， 软件工程， 2023， 博士

【摘要】 合成孔径雷达(Synthetic Aperture Radar,SAR)是一种主动遥感观测技术。SAR不依赖外部光源,而是通过自主发射微波信号并接收回波进行观测。这使得SAR能够不受云层、雨雾或夜晚等自然条件的影响,在光学遥感无法有效观测的恶劣天气或低光条件下仍能可靠地获取地表信息。然而,由于采用的相干成像原理,SAR图像固有地受到斑点噪声的污染,影响了图像的视觉质量和信息解译。因此,抑制或去除斑点噪声对于提高SAR图像的视觉质量和应用价值至关重要。近年来,深度学习方法取得了相比传统图像处理方法更优越的SAR图像去噪表现,这归功于其强大的非线性特征表达和端到端的自动特征提取能力。然而,在实际场景中应用深度学习方法进行SAR图像去噪时,仍存在尚未完善解决的问题,包括深度神经网络对于SAR图像特征表征能力有待提高、难以获取干净标签用于监督地训练深度去噪网络、在真实场景中容易出现域间隙现象以及对于边缘和纹理等细节特征保护不足。针对上述问题,本文基于深度学习展开SAR图像的去噪方法研究,主要研究内容包括:(1)基于自校准空洞卷积网络的SAR图像去噪方法为提高对于SAR图像的深度特征表征能力,提出一种基于自校准空洞卷积网络的SAR图像去噪方法。该方法在深度去噪网络中引入自校准模块,利用分组卷积分别计算原始空间和远距离空间的上下文特征,并通过特征本身提取的权重进行校准。同时,在自校准模块中引入空洞卷积来扩大网络感受野范围,并对不同的自校准模块设置混合空洞率来提取不同尺度特征,以避免棋盘格现象。实验结果表明,相比其他流行的基于深度神经网络的SAR图像去噪方法,所提方法不仅具有较强的特征表征能力,同时避免增加过多网络参数。(2)无需干净标签的SAR图像自监督去噪方法为解决难以监督地训练SAR图像深度去噪网络的问题,提出一种无需干净标签的SAR图像自监督去噪方法。首先,基于对抗学习,构建包含噪声重分布生成器、结构保持生成器和识别器的生成框架。然后,通过对抗损失、后向循环一致性损失和总变分损失对生成框架进行联合训练,仅改变带噪图像中的斑点噪声分布而保持结构特征,从而生成高质量的带噪图像对。实验结果表明,利用这些生成的带噪图像对训练深度去噪网络,可实现有效的SAR图像自监督去噪,从而无需干净标签图像。(3)基于真实带噪样本的SAR图像自监督去噪方法为解决使用模拟斑点噪声图像带来的域间隙问题,提出一种基于真实带噪样本的SAR图像自监督去噪方法。首先,通过对真实SAR图像进行伯努利采样,得到带噪训练样本。然后,采用编码器解码器网络对伯努利采样后的真实带噪样本进行学习,并嵌入密集残差块来提升网络学习能力。接着,采用基于集成加强的推演策略进一步提升SAR图像的去噪效果。最后,通过一系列对比实验,深入探究不同伯努利采样概率和不同集成次数对去噪效果的影响。实验结果表明,相比于使用模拟斑点噪声图像作为训练数据的方法,所提方法通过直接利用真实带噪样本学习噪声分布,能够避免出现域间隙问题,即避免在去噪结果图像中出现伪影,实现可靠的图像去噪效果。(4)分割指导的语义感知型SAR图像自监督去噪方法为解决SAR图像去噪结果中对细节边缘特征保护不足的问题,提出一种分割指导的语义感知型SAR图像自监督去噪方法。首先,结合SAR图像特点设计自监督去噪损失,并引入沿水平和垂直方向的边缘感知的平滑损失,以避免过度平滑图像同时提升斑点噪声的抑制效果。然后,采用语义分割网络来指导自监督去噪网络的训练。通过连接器网络将自监督去噪网络和语义分割网络前后连接,并通过小波变换将语义信息自适应地在不同频率上由分割网络传递至自监督去噪网络。实验结果表明,这样的设计使得自监督去噪网络能够具备对图像的语义感知能力。相比于其他流行方法,所提方法在充分抑制斑点噪声的同时,能够更好地保护SAR图像中的边缘和纹理等细节特征。更多还原

【Abstract】 Synthetic aperture radar(SAR)is an active remote sensing observation technology.SAR does not rely on external light sources;instead,it autonomously emits microwave signals and receives echoes for observation.This enables SAR to acquire surface information reliably without being affected by natural conditions such as clouds,rain,fog,or nighttime,which can hinder optical remote sensing.However,due to its coherent imaging principle,SAR images are inherently contaminated by speckle noise,affecting the visual quality and information interpretation of the images.Therefore,suppressing or removing speckle noise is crucial to enhancing the visual quality and application value of SAR images.In recent years,deep learning methods have achieved superior performance in SAR image denoising compared to traditional image processing methods.This is attributed to their powerful non-linear feature representation and end-to-end automatic feature extraction capabilities.However,when applying deep learning methods for SAR image denoising in real-world scenarios,there are still unresolved issues,including the need to improve the SAR image feature representation capability of deep neural networks,difficulty in obtaining clean labels for supervised training of deep denoising networks,the occurrence of domain gap phenomena in real scenes,and inadequate protection of detailed features such as edges and textures.To address these issues,this thesis conducts research on SAR image denoising methods based on deep learning principles.The main research contents include:(1)SAR image denoising method based on self-calibrated dilated convolutional networkTo enhance the deep feature representation capability of SAR images,a SAR image denoising method based on the self-calibrated dilated convolutional network is proposed.This method introduces a self-calibrated module into the deep denoising network,which calculates context features of both the original space and distant space using grouped convolutions,and calibrates them using weights extracted from the features themselves.Meanwhile,dilated convolutions are introduced in the self-calibrated module to expand the network’s receptive field,and hybrid dilated rates are set for different self-calibrated modules to avoid the checkerboard artifact.Experimental results show that compared to other popular deep neural network-based SAR image denoising methods,the proposed method not only has strong feature representation capability but also avoids adding too many network parameters.(2)SAR image self-supervised denoising method without clean labelsTo solve the problem of difficulty in supervising the training of SAR image deep denoising network,a self-supervised denoising method for SAR images without clean labels is proposed.Firstly,based on adversarial learning,a generation framework is constructed containing a noise redistribution generator,a structure-preserving generator,and a discriminator.Then,through joint training of the generation framework using adversarial loss,backward cycle consistency loss,and total variation loss,the structure features are preserved while changing the speckle noise distribution in noisy images,thus generating high-quality noisy image pairs.Experimental results show that using these generated noisy image pairs to train deep denoising networks can achieve effective SAR image self-supervised denoising without needing clean labeled images.(3)SAR image self-supervised denoising method based on real noisy samplesTo address the domain gap issue caused by simulated speckled images,a self-supervised denoising method for SAR images based on real noisy samples is proposed.Firstly,noisy training samples are obtained by Bernoulli sampling of real SAR images.Then,an encoderdecoder network is used to learn the real noisy samples obtained from Bernoulli sampling,and dense residual blocks are embedded to enhance the network’s learning capability.Next,an inference strategy based on ensemble reinforcement is adopted to improve the denoising effect of SAR images further.Finally,through a series of comparative experiments,the influence of different Bernoulli sampling probabilities and different ensemble times on the denoising effect is explored in depth.Experimental results show that compared to the methods of using simulated speckled images as training data,the proposed method can avoid the domain gap issue by directly utilizing real noisy samples to learn noise distribution,i.e.,avoiding pseudoartifacts in denoised result images,and achieve reliable image denoising effects.(4)Segmentation-guided semantic-aware self-supervised denoising method for SAR imageTo address the issue of inadequate protection of detailed edge features in SAR image denoising results,a segmentation-guided semantic-aware self-supervised denoising method for SAR images is proposed.Firstly,a self-supervised denoising loss is designed considering the characteristics of SAR images,and an edge-aware smoothing loss along horizontal and vertical directions is introduced to avoid excessive smoothing of images while enhancing speckle noise suppression.Then,a semantic segmentation network is used to guide the training of the selfsupervised denoising network.The self-supervised denoising network and the semantic segmentation network are connected before and after through connector networks,and semantic information is adaptively transferred to the self-supervised denoising network at different frequencies using wavelet transforms.Experimental results show that this design enables the self-supervised denoising network to possess semantic awareness of images.Compared to other popular methods,the proposed method not only effectively suppresses speckle noise but also better preserves detailed features such as edges and textures in SAR images.更多还原