基于低秩稀疏分解和字典学习的图像超分辨率重建研究

发布时间：2018-03-14 12:01

本文选题：图像超分辨率　切入点：低秩矩阵稀疏分解　出处：《山东师范大学》2017年硕士论文　论文类型：学位论文

【摘要】：在科技不断取得创新和发展的大背景下,数字多媒体技术(Digital Multimedia Technology)也得到了推广及应用。其中,图像视觉信号已成为数字化多媒体传播信息的主要载体之一,大量的成像设备(如手机、数码相机和iPad等)不断被设计出来。获取高质量的图像是人们不断追求的目标。然而在成像过程中,由于受外界各种不确定因素的干扰,人们最终获得的图像会失真。为了追求获取高质量的图像,经过研究者们的不断研究与探索,图像超分辨率(Image Super Resolution,ISR)重建理论不断得到完善,一些实际的可用性成果已被研究出来。在压缩感知(Compressed Sensing,CS)理论的发展影响下,衍生出一些有效的ISR重建算法。其中,基于信号稀疏表示(Sparse Representation)的ISR重建算法以其相对较好的重建结果受到研究者们的青睐;另一个与CS理论内在相关的理论是图像矩阵的低秩稀疏分解(Low Rank Sparse Decomposition,LRSD)理论,二者是数据的两种不同表示方式。基于两种不同的信号表示理论基础,本文提出了一种新的图像超分辨率重建算法:1.从图像的结构上分析,它的某些微结构间具有相似性。本文首先利用某种特定的方式去分解图像,然后通过图像块间的欧氏距离测量,找出结构相似的图像块,并将图像信号向量化表示,再将它们组成图像矩阵。由于结构相似的图像块信号间存在相关性,所以它们组成的矩阵是天然的低秩矩阵。以上述思想为基础,本文初步重建出一种基于低秩约束的图像重建模型,初步重建出的初始高分辨率图像可以保证与退化低分辨率观测图像基本结构上的一致。2.第二步的工作目标是恢复初始重建图像中缺失的细节信息,即图像中缺失的高频成分。基于稀疏表示的图像重建算法是目前对图像细节信息恢复重建表现较好的算法,本文借鉴它的思想来达到这一节的目的。特别与传统字典训练方法不同的是,结合图像低秩稀疏分解理论本节在样本集的构建阶段提出创新,在提高字典训练效率的同时也改善了图像的重建质量。为了验证本文提出的重建算法的有效性,本文最后进行了仿真实验。通过与该领域一些已有的表现较好的重建算法相比,在标准评价方法下的结果证实了本文提出的算法更具有优越性,对图像的细节部分恢复效果表现更好。
[Abstract]:In the background of continuous innovation and development of science and technology, digital Multimedia technology has also been popularized and applied. Among them, image visual signal has become one of the main carriers of digital multimedia communication information. A large number of imaging devices (such as mobile phones, digital cameras, iPad, etc.) are constantly being designed to obtain high-quality images. However, in the imaging process, due to various uncertainties outside the interference, In order to obtain high quality images, the theory of super-resolution image Super resolution (ISR) reconstruction has been improved through the research and exploration of researchers. Some practical usability results have been studied. Under the influence of compressed sensing theory, some effective ISR reconstruction algorithms are derived. The ISR reconstruction algorithm based on sparse representation of signals is favored by researchers for its relatively good results, and the theory of low rank sparse decomposition of image matrix and low Rank Sparse decompositionLRSD) is another theory that is intrinsically related to CS theory. They are two different representations of data. Based on two different signal representation theories, a new super-resolution image reconstruction algorithm:: 1 is proposed in this paper. Some of its microstructures are similar to each other. In this paper, the image is decomposed in a certain way, and then by measuring the Euclidean distance between the image blocks, the image blocks with similar structure are found, and the image signal is vectorized. Then they form the image matrix. Because of the correlation between the signals of similar image blocks, the matrix they form is a natural low rank matrix. In this paper, an image reconstruction model based on low rank constraint is proposed. The initial high-resolution image can be consistent with the basic structure of the degraded low-resolution observation image. The second step is to restore the missing detail information in the initial reconstructed image. The image reconstruction algorithm based on sparse representation is a good algorithm for image detail information restoration and reconstruction. This paper uses its ideas for reference to achieve the purpose of this section. In particular, different from the traditional dictionary training methods, combining with the theory of image low-rank sparse decomposition, this section proposes innovations in the construction of sample sets. The efficiency of dictionary training is improved and the image reconstruction quality is improved. In order to verify the effectiveness of the reconstruction algorithm proposed in this paper, simulation experiments are carried out at the end of this paper. The results under the standard evaluation method show that the proposed algorithm is more superior and performs better in the detail restoration of the image.
【学位授予单位】：山东师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】