基于非参数估计理论的多帧低照度图像去噪方法研究

Image denoising problem has been considered as one of the hottest issues in both digital image processing and computer vision fields. With the development of image processing technology and the popularity of the image and video capture devices, the image denosing problem under specific scenarios has gained much more attention. Under the guidance of the non-parametric estimation theory, this project focuses on the research of multi-image denosing method with multi short-time exposure images captured under low illumination. The main innovation include, with analyzing the statistical properties of the noise sources in the image capture device under low illumination, a mathematical model reflecting the low light image noise statistics characteristics is established, and the parameters of the noise model are estimated with multi short-time exposure images. Based on the non-parametric estimation theory, the three-dimensional kernel function and non-local kernel function is constructed, which reflecting the local structure characteristics in the spatial-temporal domain of the multi-images. By incorporating both the three-dimensional kernel function and the non-local kernel function under a variational framework, the multi-image denoising method based on joint spatial-temporal domain filtering is implemented. Finally, based on the results of the two studies mentioned above, the multi-image denoising method under low illumination scenarios is developed, which is adaptive to the noise intensity. This research could be considered as a complete solution for image and video denoising under low illumination scenarios, and will provide potential new insight and method for image segmentation, object tracking, pattern recognition research under low illumination condition.

图像去噪问题一直以来都是图像处理与计算机视觉等领域所关注的热点问题。随着图像处理技术的发展和图像采集设备的普及，针对特定场景下的图像去噪问题获得了越来越广泛的关注。本项目主要针对低照度这一应用场景，在非参数估计理论的指导下，研究基于多帧短时曝光图像的图像去噪方法。主要创新点包括：分析低照度场景下，成像设备中噪声源的统计特性，建立反映低照度图像中噪声统计特性的数学模型，设计和实现基于多帧的噪声模型参数估计方法；基于非参数估计理论，构建反映多帧图像空时域局部结构特性的三维核函数和非局部均值核函数，并将两者纳入变分框架下，实现基于空时域联合滤波的多帧图像去噪；并基于上述两项研究成果，实现针对低照度场景的，噪声强度自适应的多帧图像去噪方法。相关研究成果可以作为低照度场景下，图像和视频去噪的完整解决方案，也可以为低照度场景下的图像分割、目标跟踪、模式识别等研究提供新的解决思路和技术。

图像去噪问题一直以来都是图像处理与计算机视觉等领域所关注的热点问题之一。随着图像采集设备的普及和图像处理技术的发展，针对低照度场景下的图像质量增强获得了越来越广泛的关注。本项目针对低照度场景下的图像质量增强问题，研究利用针对同一场景内容的多帧短时曝光图像实现图像去噪。针对该问题，项目组首先构建包含室内、室外场景的多帧短时曝光低照度图像序列数据库，并建立了反映低照度图像中噪声统计特性的数学模型，实现了对噪声模型的参数估计；在非参数估计理论指导下，构建了反映多帧低照度图像空时域局部结构特性的三维核函数，设计和实现了一种基于加权核回归的多帧图像去噪算法；在考虑低照度图像中噪声分布不均匀这一客观实际的基础上，通过将反映多帧低照度图像空时域特性的三维核函数和反映多帧低照度图像空时域局部相似性的非局部均值滤波纳入到变分框架下，设计和实现了一种噪声强度自适应的多帧图像去噪算法。针对所构建的低照度图像数据库的仿真结果表明，相比于已有的多帧图像去噪算法，所提出的噪声强度自适应的多帧图像去噪算法在抑制噪声对图像质量影响的同时，可以有效的保留图像中的细节信息，从而显著的提升了去噪图像的主观质量。针对低照度场景下的图像质量增强问题，本项目在一定程度上揭示了低照度场景下噪声的统计特性，同时为利用多帧图像中的空时域冗余来实现低照度图像质量增强，提供了一种新的思路。从应用前景上看，所提出的噪声强度自适应的多帧图像去噪算法，在进一步优化其计算复杂度的基础上，可以应用于包括安防监控、交通工程、卫星遥感在内的多个领域，具有较为广泛的应用前景。