基于压缩感知的图像超分辨率重构算法研究
发布时间：2019-03-20 18:20
【摘要】：众所周知，在人们的日常生活中数字图像随处可见，诸如遥感成像、安全监控、医学成像等。随着对数字图像的需求越来越多，随之而来的对数字图像质量的要求也越来越高，能够体现更多细节信息的高分辨率图像成为图像重构技术中研究的核心内容。图像超分辨率重构技术就是利用同一场景不同环境下获得的多幅低分辨率图像来重构一幅高分辨率图像。由于成像设备、光照、拍摄物与成像设备间的相对位移等因素的干扰，获取的图像分辨率往往都比较低。但这些由于不同因素发生降质的图像中包含了更丰富的图像信息，将这些信息融合到一幅图像中，再经过图像重构技术对其进行恢复，所获得的图像具有更高的分辨率。图像超分辨率重构技术成为获取高分辨率图像的关键技术之一。 压缩感知理论作为一个新的采样理论，它可以在远小于奈奎斯特采样率的条件下获取信号的离散样本，保证信号的无失真重构。该理论在现代信号处理领域有着广阔的发展空间和实用性。本文的主要工作内容为以下几项： 第一、对图像超分辨率技术的研究现状进行了简单的介绍，并针对该技术的数学特性进行了详细地分析。概述了该技术的三项核心研究内容：图像配准、图像融合及图像重构，归纳总结了几种图像质量评价标准。 第二、针对不同的应用范围，首先对图像配准技术进行了分类总结，其中针对基于特征的SIFT图像配准进行了详细的数学算法分析，并进行了仿真实验，验证了其旋转尺度不变性及模糊不变性。对于拉普拉斯金字塔图像融合技术进行了整理与分析，并进行了仿真实验。 第三、对图像超分辨率重构算法进行了分类介绍，总结了各类算法的优缺点。对其中空域法中的非均匀样本内插法、迭代反向投影法、最大后验概率估计法和凸集投影法进行了详细的数学分析，对比分析了优缺点。并针对凸集投影法提出了一种优化算法，该算法首先将图像进行区域划分插值，然后利用二维非线性滤波加强图像边缘，仿真实验验证了优化算法的重构图像含有更多的细节信息。 第四、已有的研究证明，图像在小波域具有高度可压缩的性质，可以利用压缩感知理论对单幅图像进行精确地超分辨率重构。重构算法需要将低通滤波器加入到测量矩阵里使得图像超分辨率重构问题能够满足压缩感知理论的约束有限等距性质。针对正交匹配追踪重构算法进行了分析及优化，，优化算法采用每次选取稀疏度K个原子来更新支撑集，并采用二分坐标下降迭代法加快收敛速度，仿真实验验证了优化算法在重构质量和重构时间上的优越性。
[Abstract]:It is well known that digital images can be seen everywhere in people's daily life, such as remote sensing imaging, security monitoring, medical imaging and so on. With the increasing demand for digital image, the quality of digital image becomes higher and higher. High-resolution image, which can reflect more detail information, has become the core of image reconstruction technology. Image super-resolution reconstruction technique is to reconstruct a high-resolution image by using multiple low-resolution images obtained in different environments of the same scene. Because of the interference of imaging equipment, illumination, relative displacement between camera and imaging equipment, the resolution of the obtained image is usually low. However, these images, which are degraded by different factors, contain more abundant image information. They are fused into a single image, and then restored by image reconstruction technology. The obtained images have higher resolution. Super-resolution reconstruction has become one of the key technologies for obtaining high-resolution images. As a new sampling theory, compressed sensing theory can obtain discrete samples of signals far less than Nyquist sampling rate, and guarantee signal reconstruction without distortion. The theory has wide development space and practicability in the field of modern signal processing. The main contents of this paper are as follows: firstly, the research status of image super-resolution technology is briefly introduced, and the mathematical characteristics of this technology are analyzed in detail. This paper summarizes three core research contents of this technology: image registration, image fusion and image reconstruction, and summarizes several evaluation criteria of image quality. Secondly, aiming at the different application areas, the image registration technology is classified and summarized firstly, among which, the mathematical algorithm analysis of feature-based SIFT image registration is carried out in detail, and the simulation experiment is carried out. The rotation scale invariance and fuzzy invariance are verified. The image fusion technology of Laplacian pyramid is sorted and analyzed, and the simulation experiment is carried out. Thirdly, the classification of image super-resolution reconstruction algorithm is introduced, and the advantages and disadvantages of all kinds of algorithms are summarized. In this paper, the non-uniform sample interpolation method, iterative back projection method, maximum posterior probability estimation method and convex set projection method are analyzed in detail, and their advantages and disadvantages are compared and analyzed. An optimization algorithm for convex set projection method is proposed. Firstly, the image is divided into regions and interpolated, and then two-dimensional nonlinear filtering is used to enhance the edge of the image. Simulation results show that the reconstructed image contains more detailed information. Fourth, it has been proved that the image is highly compressible in wavelet domain, and the compression sensing theory can be used to reconstruct the super-resolution of a single image accurately. The reconstruction algorithm needs to add the low-pass filter to the measurement matrix so that the super-resolution reconstruction of the image can satisfy the constrained finite isometric property of the compression sensing theory. The orthogonal matching tracking reconstruction algorithm is analyzed and optimized. In the optimization algorithm, K atoms are selected each time to update the support set, and the binary coordinate descent iteration method is used to accelerate the convergence speed. Simulation results show the superiority of the optimization algorithm in reconstruction quality and reconstruction time.
【学位授予单位】：吉林大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN911.73
本文编号：2444485
[Abstract]:It is well known that digital images can be seen everywhere in people's daily life, such as remote sensing imaging, security monitoring, medical imaging and so on. With the increasing demand for digital image, the quality of digital image becomes higher and higher. High-resolution image, which can reflect more detail information, has become the core of image reconstruction technology. Image super-resolution reconstruction technique is to reconstruct a high-resolution image by using multiple low-resolution images obtained in different environments of the same scene. Because of the interference of imaging equipment, illumination, relative displacement between camera and imaging equipment, the resolution of the obtained image is usually low. However, these images, which are degraded by different factors, contain more abundant image information. They are fused into a single image, and then restored by image reconstruction technology. The obtained images have higher resolution. Super-resolution reconstruction has become one of the key technologies for obtaining high-resolution images. As a new sampling theory, compressed sensing theory can obtain discrete samples of signals far less than Nyquist sampling rate, and guarantee signal reconstruction without distortion. The theory has wide development space and practicability in the field of modern signal processing. The main contents of this paper are as follows: firstly, the research status of image super-resolution technology is briefly introduced, and the mathematical characteristics of this technology are analyzed in detail. This paper summarizes three core research contents of this technology: image registration, image fusion and image reconstruction, and summarizes several evaluation criteria of image quality. Secondly, aiming at the different application areas, the image registration technology is classified and summarized firstly, among which, the mathematical algorithm analysis of feature-based SIFT image registration is carried out in detail, and the simulation experiment is carried out. The rotation scale invariance and fuzzy invariance are verified. The image fusion technology of Laplacian pyramid is sorted and analyzed, and the simulation experiment is carried out. Thirdly, the classification of image super-resolution reconstruction algorithm is introduced, and the advantages and disadvantages of all kinds of algorithms are summarized. In this paper, the non-uniform sample interpolation method, iterative back projection method, maximum posterior probability estimation method and convex set projection method are analyzed in detail, and their advantages and disadvantages are compared and analyzed. An optimization algorithm for convex set projection method is proposed. Firstly, the image is divided into regions and interpolated, and then two-dimensional nonlinear filtering is used to enhance the edge of the image. Simulation results show that the reconstructed image contains more detailed information. Fourth, it has been proved that the image is highly compressible in wavelet domain, and the compression sensing theory can be used to reconstruct the super-resolution of a single image accurately. The reconstruction algorithm needs to add the low-pass filter to the measurement matrix so that the super-resolution reconstruction of the image can satisfy the constrained finite isometric property of the compression sensing theory. The orthogonal matching tracking reconstruction algorithm is analyzed and optimized. In the optimization algorithm, K atoms are selected each time to update the support set, and the binary coordinate descent iteration method is used to accelerate the convergence speed. Simulation results show the superiority of the optimization algorithm in reconstruction quality and reconstruction time.
【学位授予单位】：吉林大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN911.73
【参考文献】
相关期刊论文 前5条
1 黄建明,李金宗,李冬冬;利用二维插值核的高分辨力图像重构[J];光电工程;2005年05期
2 李家德;张叶;贾平;;采用非局部均值的超分辨率重构[J];光学精密工程;2013年06期
3 赵康;王建国;;无源毫米波成像改进POCS超分辨率算法[J];雷达科学与技术;2011年05期
4 吴亚娟;刘晓锋;;基于线过程模型的POCS超分辨图像重构算法[J];中国民航飞行学院学报;2008年05期
5 李彦;;探究车牌识别中的超分辨率算法[J];淮海工学院学报(自然科学版);2013年02期
本文编号：2444485
本文链接：https://www.wllwen.com/kejilunwen/wltx/2444485.html
