基于空间平台的空间目标成像及识别方法研究

With the increasing insterest in the exploration and utilization of space, it is urgent to surveille space debris and space shuttle. Research on imaging, resolution-enhancement and recognition, taking the characteristics of space platform and the space objects into consideration, are of great significance. Because it faces difficulty to increase the resolution of space image as space based camera is limited by size and power, a spatial-temporal compressive sampling based computational imaging method is proposed to break the physical limitation of the imaging system on resolution. Another degeneration of space image is the blur introduced by the relative high speed motion between platform and objects. In order to recover more details of image, a local structure based image restoration method which attempts to compose global and local structure information together is proposed. As there are similarities among local texture in spatial, temporal and scale domain, a multi-frame and multi-scale local texture descriptive model is utilized to improve the performance of super resolution under inadequate amount of frames. As the result, this method fits space applications better. Parameter learning and statistics inference approaches for graphic model under inadequate sample and feature uncertainty is proposed to improve the ability to recognize local components of space object, since the shape and pattern of space object is complicated and the amount of sample is small. As space objects sequence is not distributed dense in feature space and its feature is not continuous, a corresponding manifold learning methods is proposed to precisely classify and recognize space objects. This research can greatly improve the ability of space exploration and surveillance.

人类对近地空间探索利用日益加剧，迫切需要对空间碎片和航天器进行监视。针对空间平台和目标特点，研究其成像方式、清晰化处理和识别方法具有重要意义。受空间平台载荷和功耗等限制，改进硬件难以提高图像分辨率，提出时空压缩感知计算成像方法，突破物理成像在分辨率等方面的局限性；空间平台与目标高速运动产生图像模糊，结合空间图像全局与局部结构信息，提出基于局部结构的图像复原方法，提高图像细节恢复性能；充分利用空间目标图像局部纹理结构在时空和尺度上的相似性，建立多帧多尺度的局部纹理相似性描述模型，改善帧数不足时超分辨重建性能，更加符合空间应用需求；空间目标形态模式复杂且样本少，提出小样本和特征不可靠条件下的概率图模型学习和推理方法，提高对空间目标局部部件的识别能力；空间目标序列在高维特征空间中分布非稠密、特征非连续，提出相应的流形学习方法，实现精确分类识别。研究成果将对提高航天探测及空间监视能力具有重要作用。

随着人类对近地空间探索和利用的加剧，空间碎片日益增加，迫切需要对其进行有效监视，提升航天器自主探测和预警规避能力。基于小型、轻型的空间平台成像设备的高分辨率成像，模糊、噪声、降采样耦合因素影响下的图像清晰化重建以及小样本复杂形态模式下的空间目标识别是其瓶颈和关键科学问题。针对空间平台下，单纯通过硬件提升分辨率受到空间平台体积、重量和功耗等限制的问题，提出了基于空-时压缩感知的计算成像手段，基于测试、加工实验平台，建立了多帧互补的编码孔径成像模型。提出了结构化稀疏先验、流形结构化稀疏先验、三维全变差稀疏先验以及重加权拉普拉斯等压缩感知稀疏先验模型，实现了鲁棒的压缩感知重建。可实现硬件规模受限条件下空间目标的清晰成像。针对空间成像环境复杂，目标图像通常受到模糊、噪声、降采样等多种降质动态耦合作用的问题，构建了自适应稀疏先验、可变形字典图像块及图像多尺度自相似性等先验模型，揭示了模糊与噪声图像之间的相关性与互补性规律，提出了自动梯度激活、广义LASSO高效匹配追踪算法等优化方法。基于以上模型、规律及方法提升了空间目标图像清晰化的效果和效率。针对空间目标形态模式复杂且有效特征样本少，其为不同姿态下的目标匹配识别造成较大困难的问题，提出了最大权重双向置信传播及动态双边向置信传播方法和基于可变带宽技术的数据核密度估计方法，解决了多约束概率图模型的优化求解问题。利用此求解模型结合局部显著性，设计了深度神经网络和压缩感知等方法用于提取适于描述空间目标的有效特征，实现了目标的识别。为了综合验证成像方法与识别算法的有效性，初步建立了计算成像相机系统验证平台、空间环境仿真暗室、天文圆顶空间目标观测设施等仿真实验环境，用于验证空间目标的仿真获取及相关算法。初步推广至硬件开发板平台，为星上实时图像处理奠定一定基础。本项目基于关键科学问题进行研究，获得的成像、清晰化及识别理论研究成果可为未来空间平台高分辨率成像设备及图像处理设备提供技术储备和支撑，已推广应用至多项国家计划及多家航天单位中，对提升我国空间碎片监视能力，满足我国航天领域的迫切需求具有重大意义。