单幅图像超分辨技术研究

发布时间：2018-01-19 15:07

本文关键词： 超分辨率低秩稀疏非局部相似性 l_(1/2)正则迭代反投影　出处：《华东师范大学》2017年硕士论文　论文类型：学位论文

【摘要】：作为计算机视觉领域中的经典问题,图像超分辨率重建旨在通过一幅或者多幅低分辨率图像,恢复对应的较高分辨率的图像。但是,给定一幅低分辨率的输入图像,超分辨问题存在多重解,因而我们需要借助一些合适的先验知识来克服上述难题。目前,研究者提出的许多算法都可以取得不错的效果,但这些算法的结果可能会受到异常值的影响,产生一些不属于原始图像的额外细节。在本文中,针对图像超分辨率问题,我们主要阐述以下两大解决方案:(a)基于非局部稀疏和低秩正则的单幅图像超分辨重建算法。其主要通过非局部冗余性来恢复图像的潜在特征。在该算法中,我们对每个图像块提取相似局部结构,然后向量化形成矩阵。借助于将上述矩阵分解成低秩和稀疏两部分,来有效地正则图像超分辨率重建问题的不适定性。在没有异常值以及图像块差异的情况下,可以利用低秩矩阵来近似相似图像块矩阵。然而,相似图像块之间必定存在细微不同,而且在重建过程中可能会受到异常值的干扰,因此我们将上述矩阵分解成了低秩成份和稀疏成份两部分。我们再结合与低分辨率图像逼近的保真项,最终形成了超分辨模型。(b)基于l_(1/2)和非局部低秩稀疏正则的图像超分辨重建算法。鉴于l1/2范数约束所获得的解比l1范数更稀疏的特性,本算法以稀疏表达算法为基础,运用l1/2范数正则表达系数的稀疏性,结合图像块对字典联合训练,获得了包含丰富细节内容的图像重建初始值。为了进一步提高图像重建质量,我们以非局部稀疏和低秩正则项作为约束,运用迭代反投影思想,把初始恢复值反投影到退化图像的解空间中。在重建更多细节的同时,较好地保留了输入图像的结构特征。大量测试数据显示,所提出的算法获得了具有竞争性的超分辨率结果,能够重建出更多的图像细节,拥有较清晰的图像边缘信息,同时有效地抑制了伪影以及异常值。
[Abstract]:As a classical problem in the field of computer vision, image super-resolution reconstruction aims to restore the corresponding high-resolution image through one or more low-resolution images. Given a low-resolution input image, the super-resolution problem has multiple solutions, so we need to use some appropriate prior knowledge to overcome the above problem. Many of the algorithms proposed by the researchers can achieve good results, but the results of these algorithms may be affected by the outliers, resulting in some additional details that do not belong to the original image. Aiming at the problem of image super-resolution. Let's focus on the following two major solutions:. The super-resolution reconstruction algorithm based on non-local sparse and low-rank regularization is mainly used to restore the potential features of the image by non-local redundancy. We extract similar local structures for each image block, then vectorize to form a matrix, which is decomposed into two parts, low rank and sparse. In the case of no outliers and block differences, the low rank matrix can be used to approximate the similar image block matrix. There must be subtle differences between similar image blocks and may be disturbed by outliers during reconstruction. So we decompose the matrix into two parts: low rank component and sparse component. We combine the fidelity term with the approximation of low resolution image. The superresolution model. And non-local low-rank sparse canonical image super-resolution reconstruction algorithm. The solution obtained by L _ 1 / 2 norm constraint is more sparse than l _ 1 norm. Based on sparse representation algorithm, this algorithm uses the sparsity of L / 2 norm canonical expression coefficients and combines image blocks to train dictionaries. In order to further improve the image reconstruction quality, we use the non-local sparse and low-rank regular terms as constraints, and use the iterative back-projection idea. The initial restoration value is projected back into the solution space of the degraded image. While more details are reconstructed, the structural features of the input image are well preserved. A large number of test data are shown. The proposed algorithm can obtain competitive super-resolution results, can reconstruct more image details, have clear image edge information, and effectively suppress artifacts and outliers.
【学位授予单位】：华东师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】