CASIA OpenIR  > 毕业生  > 博士学位论文
基于局部学习的图像编辑算法研究
Alternative TitleResearches on image editing with local learning
王颖
Subtype工学博士
Thesis Advisor潘春洪
2012-05-29
Degree Grantor中国科学院研究生院
Place of Conferral中国科学院自动化研究所
Degree Discipline计算机应用技术
Keyword局部学习 图像编辑 水平集方法 图拉普拉斯 图像融合 Local Learning Image Editing Level Set Methods Graph Laplacian Image Fusion
Abstract在图像处理和计算机视觉领域中,图像编辑具有重要的应用价值和研究意义。对于复杂的自然图像,由于其视觉模式分布的复杂性,往往难以用一个全局的统计模型对其进行建模。相对于全局复杂模式分布而言,图像局部区域内的视觉外观相对简单,统计特性更加单一。也就是说,在图像全局数据空间不满足的属性,在局部数据分布空间内可能很容易得到满足。与全局学习方法不同,局部学习关注如何从数据中获取局部有用的信息。这种立足于局部的建模方法,更加灵活,更具任务导向性。因此本文基于局部学习的观点对图像进行建模,并由此提出多种图像编辑算法。论文的主要贡献如下: 1. 提出了一种基于局部线性分类的水平集图像分割方法。该方法通过在图像局部区域引入线性分类函数来有效地区分图像的前景区域和背景区域,并通过全局误差最小化驱动水平集函数的演化。该方法不仅可以有效地处理图像亮度不均匀的情况,而且对于水平集的初始化非常鲁棒。大量的对比实验验证了该方法的有效性。 2. 提出了一种基于局部核回归的图拉普拉斯方法,并将其运用于交互式图像分割和抠图。该方法的核心思想是利用局部核映射函数和局部核岭回归实现流形正则化。局部核映射函数的优点在于它是非线性、平滑的以及可以保证局部回归的高精度性。在局部核岭回归的基础上,可以将全局回归误差表示为目标值的二次拉普拉斯形式。论文从理论上证明了基于局部核回归的模型具有图拉普拉斯的性质。实验结果表明基于局部核回归的图拉普拉斯相比于传统的图拉普拉斯方法 能得到更高精度的结果。 3. 提出了一种基于局部线性算子的图像编辑框架。由于多种图像编辑任务是通过局部操作算子来保持图像局部区域的性质,论文首先将 这些局部操作算子表示为统一的稀疏矩阵—向量乘法的形式,然后将多种图像编辑问题归入到统一的目标函数下。基于所构建的图像编辑框架,可以灵活地增加图像局部操作和改变权重因子,从而改进原有方法的不足。实验结果验证了所提出图像编辑模型的有效性。 4. 提出了一种基于多方向梯度场和加权采样的图像融合算法。所提出的算法使得融合图像在保持全色图像细节信息的同时也能保持多光谱图像的光谱信息。为了适用大规模高分辨率遥感图像的融合,论文利用图像尺度空间理论将保持多光谱图像的目标函数进行近似转换,使得目标函数可以在频率域上快速求解。融合实验结果表明,所提出的融合算法在空间分辨率和光谱分辨率上都要优于传统的融合方法。
Other AbstractIn the field of image processing and computer vision, image editing is one of the most important technologies. It is of great important value in both application and research areas. For complex natural images, due to the complex visual patterns, it is generally difficult for us to model them with global statistical models. Compared with the complexity in global region, the visual patterns and statistical properties in local region are much simpler. This means that properties which are hardly held in global region could be easily held in local region. Local learning methods can effectively explore and exploit the intrinsic information from local regions. In data modeling, local learning methods have proven to be more flexible and task-oriented than global learning methods. This motivates us to combine local learning together into the image modeling. Along this line, in this thesis, several novel image editing algorithms are developed. The main contributions of the thesis are as follows: 1. We propose a new local-region based level set model for image segmentation based on locally linear classification. In this model, locally linear classification is introduced to separate background and foreground effectively in local regions. Under this formulation, the evolution of the level set is finally driven by minimizing the total error energy. Our method can not only achieve high accuracy of segmentation, but also be robust to initialization. Comparative experiments are reported to demonstrate the effectiveness and efficiency of our model. 2. A new graph Laplacian model is proposed with local kernel regression for interactive image segmentation and matting. The key idea of this model is to perform manifold regularization on data graph via local kernel regression with nonlinear and smooth kernel-based mapping that it capable of mapping pixels with high accuracy. Based on local kernel regression, the global regression error is formulated with laplacian quadratic form of the target values. The graph Laplacian nature of this method is theoretically illustrated. Comparative experiments are conducted to validate the effectiveness and efficiency of our method. 3. We propose an image editing framework based on local linear operators. It has been demonstrated that many image editing problems can be solved via local operations. The essence of these image editing problems is utilizing local operators to preserve local image features. In our image editi...
shelfnumXWLW1714
Other Identifier200818014629093
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/6432
Collection毕业生_博士学位论文
Recommended Citation
GB/T 7714
王颖. 基于局部学习的图像编辑算法研究[D]. 中国科学院自动化研究所. 中国科学院研究生院,2012.
Files in This Item:
File Name/Size DocType Version Access License
CASIA_20081801462909(4150KB) 暂不开放CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[王颖]'s Articles
Baidu academic
Similar articles in Baidu academic
[王颖]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[王颖]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.