宇宙线缪子散射成像是一种利用天然宇宙线和缪子多重库仑散射原理的成像方法。经过十几年的发展，它已经在核材料监测等领域取得成功的应用。而在集装箱检查的安检领域，由于快速检测的需求，该方法的应用仍存在困难。在低统计性缪子测量数据的制约下，传统成像算法的单次散射密度估计值存在严重的噪声。为了提高低统计性缪子散射成像的图像质量和物质甄别能力，本文从不确定度分析的角度出发，研究实用的解决方案。 基于贝叶斯原理，建立了宇宙线缪子散射成像的统计模型，用后验概率分布描述每个体素的散射密度。根据缪子散射成像的具体流程，分析了多重库仑散射理论、宇宙线缪子动量、探测器空间分辨率和成像算法对散射密度不确定度的影响。由统计模型的分析结果，既可以给体素设定散射密度值形成图像，还可以基于概率来判断物质的材料种类。 根据宇宙线缪子散射成像的特点，本文将Bootstrap抽样、蒙特卡罗抽样和降噪卷积神经网络应用于散射密度的不确定度分析。其中，Bootstrap方法和蒙特卡罗方法通过对缪子测量数据进行随机抽样，得到每个体素的多次散射密度估计值，以这些估计值推导其概率分布；降噪卷积神经网络则利用了大量缪子散射密度图像的噪声分布规律，直接对低统计性测量下的图像进行降噪，得到剥离不确定度后的图像。 由模拟和实验平台获取缪子测量数据，在10分钟的快速检测场景下，对这3种方法在成像与物质甄别中的应用进行验证。通过PSNR评价图像质量，并统计体素材料及其真值的差异，这3种新方法均取得了优于传统的PoCA和MLSD成像算法的指标。本文提出的不确定度分析方法，降低了宇宙线缪子散射成像中体素散射密度的不确定度，适用于快速检测等低统计性的缪子散射成像应用。

Cosmic muon scattering tomography is a novel imaging method that utilizes cosmic rays and the multiple Coulomb scattering theory. It has been successfully applied in nuclear material monitoring during the development. However, the application in contraband detection of containers remains a difficult task for the fast detection requirement. Limited by the low statistical muon measurement, estimated scattering densities in traditional muon tomography algorithms have severe noise. To improve the image quality and material discrimination accuracy for low statistical muon measurement, several solutions are demonstrated in the thesis. Based on Bayes' theorem, a statistical model is proposed for cosmic muon tomography and the scattering density in each voxel is modeled with a posterior probability. Contributions of uncertainty in the scattering density from the multiple Coulomb scattering theory, the cosmic muon momentum, resolutions of detectors, and imaging algorithms are analyzed for the procedure of muon tomography. Both imaging and material discrimination tasks can be accomplished from the analysis of the statistical model. In the thesis, bootstrap sampling, Monte Carlo sampling, and denoising neural networks are applied in the uncertainty analysis for cosmic muon tomography. With both sampling algorithms, the scattering density in each voxel is calculated multiple times. Thus, the probability distribution can be derived from these estimations. Since convolutional neural networks perform well in denoising different types of images, an unsupervised denoise network is designed for muon tomography based on the uncertainty model. Applied to the 10-minute data from simulation and experiment, the imaging and material discrimination results of these methods are tested in the fast detection situation. In the tests, image qualities are evaluated by PSNR values, while predicted material labels are compared with their ground truth. These new methods achieve better performances than traditional PoCA and MLSD algorithms in quantitative comparisons of image quality and material discrimination accuracy. The proposed uncertainty analysis methods are applicable for low statistical muon tomography applications.