基于正则化的非线性扩散模型的超分辨率方法
发布时间：2020-10-27 03:15
由于噪声和硬件的限制,低端图像设备采集到的图像和视频并不理想。因此,许多文献中都对这个问题提出了解决方法。超分辨率技术就是其中一种将图像或视频由低质量重构成高质量的一种内在的适定性问题。大多数已有的超分辨率重构算法都不能完全保留一些重要的图像特征,比如物体的边缘和轮廓等,然而事实上人眼会对这些边缘很敏感,并且它们也在目标检测等计算机视觉应用领域会起到很大的作用。为解决过去方法中的问题,本文中我们提出了几种基于非线性扩散泛函的超分辨率算法。新方法能根据图像特征自动调整正则化水平。具体来说,正则化在平坦区域较强以消除噪声,在边缘区域较弱以保护重要的图像信息。这种基于图像特征的方法使得我们的模型重建后的图像信息更详细。首先,我们的超分辨率算法基于Perona-Malik光滑泛函,其中的扩散性部分含有以空间为变量的指数项,它随标准化变化而变化。第二,我们引入了一种改进的Charbonnier模型来描述超分辨率的适定性问题。这种方法能适应诸如线性等方向扩散,全偏差以及Charbonnier等不同的正则化模型,并且具有灵活性,并且能产生可观的超分辨结果。第三,为了能同时提高图像的空间分辨率和重构频率成分,我们引入Papoulis-Gerchberg算法。最后,对于超分辨问题我们得到一个新的正则化势函数。为保证势函数的凸性、光滑性和单调性,我们在参数中加了适当的约束条件。这种势函数可以使我们的超分辨模型达到更高的分辨率,这在以往的模型中是达不到的。新的重构算法有很广泛的应用。例如,可以应用于改进医学上血涂片的图像质量,准确检测并诊断疟疾等疾病。本文中,可以将任意一种超分辨率算法嵌入到低端图像采集设备(采集的图像是低分辨率图像)中,来增强输入图像的质量,这样既避免了昂贵的显微镜设备,同时保证了高准确性。而传统的自动检测诊断方法需要昂贵的硬件,许多人都无法支付。实验结果显示,本文中的模型较最先进的其它经典方法更高级。通过多种图像、视频的仿真,本文方法的视觉效果和性能指标(噪声信号峰值比、边缘和结构相似性)更理想。
【学位单位】：哈尔滨工业大学
【学位级别】：博士
【学位年份】：2015
【中图分类】：TP391.41
【文章目录】：
摘要
Abstract
Chapter 1 Introduction
1.1 Synopsis
1.2 Background of the Study
1.2.1 What is resolution?
1.2.2 Super-resolution imaging
1.3 Related works and their limitations
1.4 Objectives of the Research
1.5 Significance of the Study
1.6 Thesis outline and contributions
Chapter 2 Multiframe super-resolution image degradation model
2.1 Introduction
2.2 Image degradation model
2.3 Regularization of the multiframe super-resolution problem
2.3.1 Basics of inverse problems
2.3.2 Regularization
2.4 Comparisons of the classical regularizing functionals
2.5 Summary
Chapter 3 Super-resolution methods based on the variable exponent nonlin-ear diffusion models
3.1 Introduction
3.2 Motion estimation
3.3 Proposed methods
3.3.1 Super-resolution method based on the adaptive Perona-Malik diffu-sion model
3.3.2 Super-resolution method based on the adaptive Charbonnier diffusionmodel
3.3.3 Super-resolution method based on the non-standard anisotropic diffu-sion model
3.3.4 Super-resolution method based on the adaptive Perona-Malik modeland Papoulis-Gerchberg algorithm
3.4 Experiments
3.4.1 Preliminaries
3.4.2 Experiment 1: Edge detection
3.4.3 Experiment 2: Image denoising
3.4.4 Experiment 3: Super-resolution image reconstruction
3.5 Results and discussions
3.5.1 Experiment 1: Edge detection
3.5.2 Experiment 2: Image denoising
3.5.3 Experiment 3: Super-resolution image reconstruction
3.6 Summary
Chapter 4 A noise suppressing and edge-preserving multiframe super-resolutionmethod
4.1 Introduction
4.2 Motion estimation
4.3 Proposed smoothing energy functional
4.3.1 Derivations and important properties
4.3.2 Multiframe super-resolution process
4.3.3 Invariance and the regularizing parameter
4.4 Numerical implementation details
4.4.1 Explicit scheme
4.4.2 Additive Operator Splitting （AOS） scheme
4.5 Experiments
4.5.1 Preliminaries
4.5.2 Experiment 1: Edge detection
4.5.3 Experiment 2: Image denoising
4.5.4 Experiment 3: Super-resolution image reconstruction
4.6 Results and discussions
4.6.1 Experiment 1: Edge detection
4.6.2 Experiment 2: Image denoising
4.6.3 Experiment 3: Super-resolution image reconstruction
4.7 Summary
Chapter 5 Practical applications of the super-resolution methods
5.1 Introduction
5.2 Practical applications of the super-resolution methods
5.2.1 Fusion of images
5.2.2 Improving the spatial resolution of mammograms in X-Ray imaging
5.2.3 Improving the quality of hyperspectral images
5.2.4 Resolution enhancement of scenes on the web
5.2.5 Zooming of regions of interest （ROI） in the scene
5.2.6 Lowering the transmission costs of videos from television broadcast-ing stations
5.2.7 Improving the quality of consumer images and videos
5.3 Summary
结论
Conclusion
References
List of Publications
Acknowledgement
Resume
本文编号：2857947
【学位单位】：哈尔滨工业大学
【学位级别】：博士
【学位年份】：2015
【中图分类】：TP391.41
【文章目录】：
摘要
Abstract
Chapter 1 Introduction
1.1 Synopsis
1.2 Background of the Study
1.2.1 What is resolution?
1.2.2 Super-resolution imaging
1.3 Related works and their limitations
1.4 Objectives of the Research
1.5 Significance of the Study
1.6 Thesis outline and contributions
Chapter 2 Multiframe super-resolution image degradation model
2.1 Introduction
2.2 Image degradation model
2.3 Regularization of the multiframe super-resolution problem
2.3.1 Basics of inverse problems
2.3.2 Regularization
2.4 Comparisons of the classical regularizing functionals
2.5 Summary
Chapter 3 Super-resolution methods based on the variable exponent nonlin-ear diffusion models
3.1 Introduction
3.2 Motion estimation
3.3 Proposed methods
3.3.1 Super-resolution method based on the adaptive Perona-Malik diffu-sion model
3.3.2 Super-resolution method based on the adaptive Charbonnier diffusionmodel
3.3.3 Super-resolution method based on the non-standard anisotropic diffu-sion model
3.3.4 Super-resolution method based on the adaptive Perona-Malik modeland Papoulis-Gerchberg algorithm
3.4 Experiments
3.4.1 Preliminaries
3.4.2 Experiment 1: Edge detection
3.4.3 Experiment 2: Image denoising
3.4.4 Experiment 3: Super-resolution image reconstruction
3.5 Results and discussions
3.5.1 Experiment 1: Edge detection
3.5.2 Experiment 2: Image denoising
3.5.3 Experiment 3: Super-resolution image reconstruction
3.6 Summary
Chapter 4 A noise suppressing and edge-preserving multiframe super-resolutionmethod
4.1 Introduction
4.2 Motion estimation
4.3 Proposed smoothing energy functional
4.3.1 Derivations and important properties
4.3.2 Multiframe super-resolution process
4.3.3 Invariance and the regularizing parameter
4.4 Numerical implementation details
4.4.1 Explicit scheme
4.4.2 Additive Operator Splitting （AOS） scheme
4.5 Experiments
4.5.1 Preliminaries
4.5.2 Experiment 1: Edge detection
4.5.3 Experiment 2: Image denoising
4.5.4 Experiment 3: Super-resolution image reconstruction
4.6 Results and discussions
4.6.1 Experiment 1: Edge detection
4.6.2 Experiment 2: Image denoising
4.6.3 Experiment 3: Super-resolution image reconstruction
4.7 Summary
Chapter 5 Practical applications of the super-resolution methods
5.1 Introduction
5.2 Practical applications of the super-resolution methods
5.2.1 Fusion of images
5.2.2 Improving the spatial resolution of mammograms in X-Ray imaging
5.2.3 Improving the quality of hyperspectral images
5.2.4 Resolution enhancement of scenes on the web
5.2.5 Zooming of regions of interest （ROI） in the scene
5.2.6 Lowering the transmission costs of videos from television broadcast-ing stations
5.2.7 Improving the quality of consumer images and videos
5.3 Summary
结论
Conclusion
References
List of Publications
Acknowledgement
Resume
本文编号：2857947
本文链接：https://www.wllwen.com/shoufeilunwen/xxkjbs/2857947.html
