图像结构、纹理和偏场协同分解方法的研究

发布时间：2018-05-07 18:32

本文选题：图像分解 + 双边滤波　；参考：《山东师范大学》2017年硕士论文

【摘要】：图像分解在计算机视觉领域中一直是一个被广泛关注的问题,该问题的研究目标是将一幅图像分解成若干个不同的分量,从而实现原图像中的主要结构与纹理细节等信息的分离。解决这一问题,对计算机视觉和医学图像等领域的许多工作具有重要意义。然而,一方面,现存的经典图像分解方法大多缺少相应的函数来定义某些复杂的结构特征,如经典的双边滤波(BF)和双边纹理滤波(BTF)缺少定义视网膜图像血管结构特征的函数。另一方面,这些方法忽视了图像中常见的偏场信息对图像造成的影响,分解出的信息常受到偏场的干扰而严重丢失。针对这些问题,本文提出了图像结构、纹理和偏场协同分解的模型,并基于图像管状结构、纹理的分解以及对图像偏场的估计提出了图像协同分解的方法。该方法能够在分解图像中复杂的管状结构与纹理细节的同时不受偏场信息的影响。具体地,本文提出了一个最优线性扩散函数(OLSF)空间核算子来提取管状结构的特征,然后将其与BTF中的纹理分解算子Patch-shift(PS)融合,用于有效分解图像中的管状结构与纹理。为了消除偏场信息的干扰,我们利用鲁棒性较强的图像梯度分布稀疏性来有效地估计图像的偏场信息。具体来讲,本文的研究和贡献主要有以下几点:(1)提出了一个图像管状结构、细节分解方法,并将其成功地应用于眼底图像的降噪任务。该方法利用OLSF有效地提取特殊管状结构特征,例如局部血管的方向、尺度等等,然后利用这些特征区分血管结构信息和背景细节信息,最终在降低图像中的噪声的同时极大的保留血管结构。大量的手工图像和视网膜图像的实验结果表面,在保留对比度较低的细血管的效果上,该方法要优于经典的BF方法。此外,该方法不仅为分解视网膜图像中的血管结构提供了可行性,并且在其他包含小尺度的、低对比度的管状结构的图像上同样有效,为下一步的管状结构、纹理分解工作提供了基础。(2)提出了一个图像管状结构、纹理分解方法,能够有效地分解图像的管状和纹理细节信息。该方法基于BF框架,融合了PS算子和提出的OLSF。其中,PS算子利用每个像素的局部统计特征来定义该像素的纹理特征,具有很好的图像的结构和纹理细节分解效果。大量的视网膜图像和自然图像的实验结果表明,利用PS算子和OLSF定义BF的滤波核,能够在消除纹理信息的同时有效地保护管状结构,且其效果要优于经典的BF和BTF图像分解方法。(3)提出了图像结构、纹理和偏场协同分解的模型,并在图像管状结构、纹理分解的基础上加入了图像偏场的估计,提出了图像协同分解模型的实现方法。本文利用了图像梯度分布的稀疏性估计图像偏场信息,同时结合管状结构-纹理滤波分解方法,最终将图像分解成管状结构、背景纹理和偏场三个分量。大量的自然图像和眼底图像的实验对比的结果表面,图像协同分解模型比传统模型更加严谨,且实用性更强。和现存的经典的BF和BTF图像分解方法更相比,图像协同分解方法的优势在与分解纹理信息的同时能够更好的保护图像管状结构,而不会受到偏场信息的影响。
[Abstract]:Image decomposition has been a widespread concern in the field of computer vision. The goal of this problem is to decompose an image into several different components, so as to separate the information such as the main structure and the texture details in the original image. To solve this problem, many fields such as computer vision and medical images are solved. However, on the one hand, the existing classical image decomposition methods mostly lack the corresponding functions to define some complex structural features, such as the classical bilateral filtering (BF) and bilateral texture filtering (BTF) lack of functions to define the vascular structural features of the retinal images. On the other hand, these methods ignore the common image in the image. In this paper, the model of image structure, texture and partial field synergetic decomposition is proposed. Based on the image tube structure, the decomposition of texture and the estimation of image partial field, the method of image synergetic decomposition is proposed. In this paper, an optimal linear diffusion function (OLSF) space accounting is proposed to extract the characteristics of the tubular structure, and then it is fused with the texture decomposition Patch-shift (PS) in the BTF to effectively decompose the image. In order to eliminate the interference of partial field information, we use robust image gradient sparsity to effectively estimate the partial field information of images. In particular, the research and contribution of this paper are as follows: (1) an image tubular structure, detail decomposition method is proposed, and it is applied successfully to the image. This method uses OLSF to effectively extract special tubular structure features, such as the direction of the local blood vessel, the scale and so on, and then use these features to distinguish the vascular structure information and the background details, and eventually reduce the noise in the image and retain the vascular structure greatly. A large number of manual images and network. The experimental results surface of the membrane image is better than the classical BF method in preserving the low contrast fine blood vessel. In addition, this method not only provides the feasibility to decompose the vascular structure in the retinal image, but also is effective in the other small scale, low contrast tubular structure images, for the next step. The tubular structure provides the basis for the texture decomposition. (2) an image tubular structure and a texture decomposition method are proposed, which can effectively decompose the tube and texture details of the image. Based on the BF framework, the PS operator and the proposed OLSF. are fused, and the PS operator uses the local statistical features of each pixel to define the pixel. Texture features, with good image structure and texture details decomposition effect. A large number of experimental results of retinal images and natural images show that using the PS operator and OLSF to define the filter kernel of BF can effectively protect the texture information and effectively protect the tubular structure, and its effect is better than the classical BF and BTF image decomposition methods. (3) The model of synergetic decomposition of image structure, texture and partial field is proposed, and the estimation of image partial field is added on the basis of image tube structure and texture decomposition. The realization method of image synergetic decomposition model is proposed. This paper uses the sparsity of image gradient distribution to estimate image partial field information and combines tubular texture filtering. Finally, the image is decomposed into a tubular structure, a background texture and a partial field of three components. A large number of natural images and eye images are compared with the results of the experiment. The image co decomposition model is more rigorous and more practical than the traditional model. Compared with the existing classical BF and BTF image decomposition methods, the image cooperative decomposing side The advantage of the method is that it can better protect the image tube structure while decomposing the texture information without being affected by the bias field information.

【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】