基于多帧影像的航空超分辨成像技术研究

发布时间：2018-02-10 08:40

本文关键词： 多帧影像航空成像超分辨 Papoulis-Gerchberg 自学习字典局部线性嵌入　出处：《中国科学院长春光学精密机械与物理研究所》2017年博士论文　论文类型：学位论文

【摘要】：随着航空光电载荷的高速发展,更大的画幅、更高的图像分辨率以及更远的作用距离成为航空光电载荷不断追求的目标,但受体积、重量、功耗以及光学系统成像过程中引起的欠采样、运动模糊及噪声等因素影响,航空图像分辨率不能满足实际应用的需求,因此获取高分辨率(HR)航拍图像已成为当今航空领域的热点和难点。提高图像分辨率最直接的方法是采用高分辨率CCD相机,但受工艺水平以及相机图像数据传输速率的限制,通过高分辨率CCD相机采样得到的图像分辨率的能力是十分有限的。近年来,通过信号处理方式提升图像分辨率,即超分辨技术受到广泛关注。超分辨技术即是在不改变原有硬件成像系统基础上,仅通过软件方法,也就是利用信号处理的方法将一幅或多幅包含相似信息而细节不同的低分辨率(LR)图像重构成一幅高分辨率图像。"超"即是克服传统低分辨成像系统固有衍射极限,获取超出光学系统衍射极限以外的空间频率信息,实现进一步提高分辨率的工程应用目的。本文首先介绍超分辨成像技术的研究背景和工程应用,系统的总结、分析和比较了超分辨成像技术的物理成像模型、方法类别和评价体系。在建立的成像模型基础上,针对现有超分辨算法运算复杂度,边缘模糊,图像失真等问题,以多帧航空影像为研究对象,围绕多帧图像的超分辨成像技术主题展开了深入研究。主要研究工作如下:1.为进一步提高拍摄图像的分辨率,提出一种改进的Papoulis-Gerchberg超分辨算法,新算法提出边缘检测方法,可以改善传统方法空间复杂度和重构图像边缘模糊的问题,新算法在原有的算法基础上融于边缘检测,针对多幅同一场景输入图像,在每次Papoulis-Gerchberg迭代过程加入坎尼检测,同时将每步的重构误差投影到下一步重构过程,降低了算法空间复杂度,能有效恢复丢失的边缘高频信息。与现有的经典超分辨重构方法相比,本算法反映图像质量的峰值信噪比和灰度标准差更高,去除了原始重构方法图像边缘叠影现象,有效提高了原始输入图像的分辨率。2.研究了软硬件相结合的超分辨成像技术,首先用探测器扫描获得同一场景的彼此错位亚象元像素的多帧图像作为训练样本和输入图像,然后针对传统局部线性嵌入(LLE)实例学习超分辨算法过于依赖外部训练样本,不利于光电成像系统直接处理等缺点,提出一种基于自学习的改进局部线性嵌入(LLE)算法,采用新的LLE权值计算方法以获得正数权值,同时对初始估计再次运用自学习LLE方法恢复丢失的高频细节信息,最终能获得高质量的重构图像,能满足高质量高分辨率的成像需求。3.针对基于字典学习超分辨重构方法需要大量的HR-LR图像训练冗余字典,且若选取的HR-LR训练图像不含有待重构低分辨图像的频率信息,重构出的高分辨图像会造成失真等缺点,提出自学习字典的多幅超分辨率重构方法,用待重构的多幅同一场景不同运动参数的低分辨率图像做为输入图像和训练图像,分块学习字典,重构出高一尺度图像,并加入到训练图像中,如此依次逐级构造不同尺度图像做为训练图像集,最终重构出达到或最接近目标图像尺度大小的多幅高分辨率图像。最后利用NLM思想将得到的多幅高分辨率图像融合成一幅目标图像尺度大小的最终重构高分辨率图像。仿真实验结果表明,本文算法的重构图像信噪比更高,细节细腻,能从拍摄图像中获得更清晰的高分辨图像。4.提出一种基于自适应的高性能超分辨算法,通过将基于学习与基于重构的超分辨算法相结合,充分利用两者的优点,本文不需外部训练图像,首先以输入图像做为训练图像创建字典块集,其次在训练获得的训练块集中利用自适应学习方法获取HR图像块中心点像素值,然后利用高频恢复方法重构丢失的高频边缘信息,最后结合基于重构方法,提出用边缘做为先验知识满足重构约束,获取最终的高分辨重构图像。本文算法同时解决了基于重构算法边缘模糊和基于学习算法失真的缺点,获得了高质量的高分辨率图像,对于提升航空图像分辨率具有很重要的意义。本文针对目前各种超分辨算法的失真模糊等一系列问题,围绕多帧影像超分辨成像技术进行了探索,取得了阶段性成果,这些成果为进一步的工程实践和成熟应用提供了理论基础,对航空图像超分辨成像具有一定指导意义。
[Abstract]:With the rapid development of aviation photoelectric payload, the bigger picture, the image resolution and further distance more become the constant pursuit of the goal of aviation photoelectric payload, but by the volume, weight, power consumption and undersampling caused by the optical system imaging process, motion blur and noise and other factors, can not meet the actual resolution of aerial image the needs of the application, thus obtaining high resolution aerial images (HR) has become a hot and difficult point in the field of aviation. The method of improving the image resolution is the most direct use of high resolution CCD camera, but by the technology level and the limitation of camera image data transmission rate, the ability of image resolution by high resolution CCD camera is sampled very limited. In recent years, to improve the image resolution by signal processing methods, namely super resolution technology has attracted extensive attention. Super resolution technology is not in Change the original hardware based imaging system, only through the software method, which is using the signal processing method to one or more images containing similar information and details of different low resolution (LR) image of a high resolution image. The "super" is to overcome the traditional low resolution imaging system to obtain the inherent diffraction limit. The spatial frequency information outside the optical system beyond the diffraction limit, can further improve the resolution of engineering application. This paper first introduces the research background and application of super resolution imaging technology, system summary, analysis and comparison of the physical imaging super resolution imaging model, method of category and evaluation system. Based on imaging model based on the in view of the existing super-resolution algorithm, computational complexity, fuzzy edge, image distortion and other issues, through the multi frame image as the research object, the super-resolution imaging on multi frame images The theme of technology is deeply studied. The main research work is as follows: 1. in order to further improve the image resolution, this paper proposes an improved Papoulis-Gerchberg super-resolution algorithm, the new algorithm of edge detection method, can improve the traditional method of space complexity and reconstruction of image edge fuzzy problem, a new algorithm based on the original algorithm on the melt to the edge detection for multiple input images in the same scene, each iteration of the Papoulis-Gerchberg process to join the canny, while the projection reconstruction error of each step into the reconstruction process the next step, reduces the space complexity of the algorithm, can effectively restore the high-frequency edge information loss. Compared with the existing classical super-resolution reconstruction method, this algorithm reflects the the image quality of the peak signal-to-noise ratio and standard deviation higher, removal of the original reconstruction method of image edge aliasing phenomenon, effectively improves the original input Super resolution imaging technology for image resolution.2. of combination of hardware and software, first obtain the same scene stagger subpixel pixel multi frame images as training samples and input images with the detector scan, then the traditional local linear embedding (LLE) case study super-resolution algorithm is too dependent on external training samples, disadvantages to direct processing of photoelectric imaging system, we put forward an improved self-learning based on local linear embedding (LLE) algorithm, using LLE weighted new calculation method to obtain positive weights, and once again the use of high frequency information learning LLE method to recover the lost on the initial estimation, finally to obtain high quality images, can meet the high quality high resolution imaging.3. super-resolution reconstruction method for dictionary learning based on the need of a large number of HR-LR image training redundant dictionary, and if the selected HR-LR training Do not contain low resolution image to reconstruct the frequency information of the image, the high resolution image will cause the distortion of reconstructed defects, put forward multi frame super-resolution reconstruction method for learning the dictionary, using low resolution images to reconstruct multiple images of the same scene with different motion parameters as the input image and the training images, block learning the dictionary, to reconstruct the high resolution images, and added to the training images, so the images of different scale structure followed step by step as the training image set is reconstructed more images of higher resolution close to the target image size. The fusion of multiple images of higher resolution got into the final reconstruction of high-resolution image the size of the target images with NLM. Simulation results show that this algorithm reconstruction SNR is higher, the details, from the captured image more clear and high .4. presents a high resolution image super-resolution algorithm based on adaptive performance, the learning based super-resolution algorithm and reconstruction based on the combination, make full use of their advantages, this paper does not need external training image, first of all to the input image as the training images to create a dictionary block set, then in training to obtain the training block by adaptive learning method to get the center of the block HR image pixel values, then the high-frequency edge information recovery method using high frequency loss reconstruction, finally based on the reconstruction method is proposed for edge reconstruction constraints do meet as prior knowledge, to acquire high resolution image reconstruction in the end. This algorithm solves the reconstruction algorithm and learning algorithm of fuzzy edge distortion defects based on high resolution, high quality images were obtained, has a very important significance for improving image resolution. Aiming at the air All kinds of super resolution algorithm of fuzzy distortion of a series of problems such as multi frame image super-resolution imaging on technology are explored, and achieved initial results, these results provide a theoretical basis for further engineering practice and mature application of aerial image super-resolution imaging has certain guiding significance.

【学位授予单位】：中国科学院长春光学精密机械与物理研究所
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】