图像处理中去噪与超像素生成算法研究

发布时间：2017-12-28 00:28

本文关键词：图像处理中去噪与超像素生成算法研究　出处：《山东大学》2017年博士论文　论文类型：学位论文

【摘要】：视觉是人类获取信息的主要方式之一。图像在人类工作和生活中充当着异常重要的角色。图像处理就是指从图像中提取信息的关键技术,在工业和生活等各个方面都有广泛的应用,如农业生产、石油勘探、生物医学等领域。图像去噪与超像素生成(过分割)作为图像处理中的两个基本任务,为图像分析和理解提供了坚实的基础。在图像获取、压缩、传输阶段,由于环境、传输信道等多种因素的影响,图像会受到噪声的干扰,而使图像信息丢失,产生失真。对失真的图像进行处理,势必会影响图像处理的结果,降低提取信息的准确度,进而干扰据此做出的各种判断和决策。图像去噪即是要从含有噪声的图像信号中去除噪声干扰,从而恢复出图像真实信号,进而保证进一步图像处理与分析结果的准确性。因为噪声种类的多样性,目前去噪算法还存在一些问题。一方面,图像去噪算法通常都是针对某一种特殊的噪声模型提出的,而算法效果很大程度上取决于所处理噪声和算法所采用的噪声模型的匹配程度。绝大多数已经存在的去噪算法都是针对高斯噪声而设计,对其他噪声,如椒盐噪声,则去噪效果微乎其微。另一方面,针对单幅图像提出的去噪算法用于视频序列去噪时,是对每一帧图像依次分别进行处理,忽略了视频的时域特征。图像超像素是将具有相似属性的像素点聚集成一个区域,代替像素对图像进行表示,从而减少表示图像原子结构的数量级,进而降低后续图像处理算法的复杂度,为图像处理算法的实时性提供可能。由于图像本身种类繁多、内容复杂,目前超像素生成算法还存在一些问题和挑战。一方面,超像素的边界应该与图像边缘紧密贴合。而在需要考虑紧致性的情形下,由于边界贴合率与紧致性二者之间的相互制约关系,使得同时获得既能较好保持边界又具有高紧致性的超像素难度较大。另一方面,大多数算法都试图生成大小一致的超像素,而由于图像内容分布的不均匀性,在超像素数目一定的情况下,如果平坦区域可以采用少量尺寸较大的超像素表示,而细节丰富的区域采用数量相对较多、尺寸较小的超像素进行表示,那么图像就能获得更准确的表达,这样的超像素能保持更多的图像信息。本文围绕上述图像去噪与超像素生成中存在的难点问题展开研究,并提出了多种解决方案。本文的主要工作包括:1.提出了基于低秩矩阵恢复的去除视频椒盐噪声的方法。通过观察分析,视频相邻帧图像之间具有很大的结构相似性,也就是说视频序列具有极高的时间冗余性。采用非局部处理方法的基本思想,将每一帧图像当作一个图像块,相邻帧图像作为相似图像块,将其各自排列成列向量并构成矩阵,该矩阵具有低秩性。这样就将去除视频椒盐噪声的问题转换为了低秩矩阵恢复问题。该算法可以较好的去除视频信号中的椒盐噪声,并获得令人满意的量化结果。2.提出了保持边界的图像超像素生成算法。为了使生成的超像素在兼顾紧致性的同时能更好的贴合图像边界,该算法重新定义了一个集超像素边界信息、紧致性和内部一致性于一体的距离度量函数对每个像素点进行归类。同时通过设置参数,调节三者之间的比重,实现边界贴合率与紧致性之间的可调平衡。为了能更好的刻画图像边界,该算法还提出了一种利用邻域像素信息度量像素点位于图像边界概率的计算公式。为了使算法在每次迭代的过程中的结果能更加准确,算法中还提出了基于3-sigma原则的更新策略。实验结果证明,该算法生成的超像素不仅能很好的贴合图像边界,而且可以在边界贴合率和紧致性之间实现可调节的平衡。3.提出了基于二次误差测度(QEM)的超像素生成算法。图像是从三维空间中经过采样得到的。如果将二维图像反映射回三维空间,那么三维空间中的算法就可以用于图像处理领域。基于此种想法,本文将图像网格化,并拓展到三维空间,从而将图像超像素生成问题转换成了网格简化问题。为了使网格简化方法更好的适用于图像处理领域,该算法对传统QEM算法进行了改进,重新定义了边收缩代价的度量,并选择了多种图像属性进行计算。通过多次实验验证,表明该方法是有效的且能生成多种尺度的超像素,并能较好的贴合图像边界。
[Abstract]:Vision is one of the main ways people get information. Images play an important role in human work and life. Image processing is the key technology to extract information from images. It has been widely applied in all aspects of industry and life, such as agricultural production, petroleum exploration, biomedicine and so on. Image denoising and super - pixel generation (over segmentation), as two basic tasks in image processing, provide a solid foundation for image analysis and understanding. In the stage of image acquisition, compression and transmission, due to various factors such as environment, transmission channel and so on, the image will be disturbed by noise, and the image information will be lost and distorted. The processing of distorted images is bound to affect the results of image processing, reduce the accuracy of information extraction, and interfere with the various judgments and decisions made accordingly. Image denoising is to remove noise interference from noisy image, and restore the real signal of image, so as to ensure the accuracy of further image processing and analysis. Because of the variety of noise, there are still some problems in the current denoising algorithm. On the one hand, image denoising algorithms are usually aimed at a particular noise model, and the algorithm depends largely on the matching degree of the noise and the noise models adopted by the algorithm. Most of the existing denoising algorithms are designed for Gauss noise. For other noises, such as salt and pepper noise, the denoising effect is negligible. On the other hand, the denoising algorithm proposed for single image is used to process every frame image in order to process the video sequence denoising, ignoring the temporal characteristics of the video. Image super pixels are the pixels with similar attributes gathered in a region, instead of pixels to represent the image, thus reducing the number of said level image of atomic structure, thereby reducing the complexity of subsequent image processing algorithms, providing real-time image processing algorithms. Because of the wide variety and complex content of the image itself, there are still some problems and challenges in the present super pixel generation algorithm. On the one hand, the boundary of the super pixel should be close to the edge of the image. In the case of compactness, due to the mutual restriction between the two constraints of the boundary and the compactness, it is difficult to obtain super pixels with good boundary and high compactness simultaneously. On the other hand, most of the algorithms to generate the same size of super pixels, but due to non-uniform distribution of the image content, the number of pixels in certain circumstances, if the flat area can be used a larger size of super pixels, and detail abundant areas with a relatively large number of smaller size, super pixel representation then, the image can get a more accurate expression, so the super pixel can keep more image information. In this paper, the difficulties in the image denoising and super pixel generation are studied, and a variety of solutions are proposed. The main work of this paper is as follows: 1. the method of removing salt and pepper noise based on low rank matrix recovery is proposed. Through observation and analysis, the video sequence has a great structural similarity, which means that the video sequence has a very high time redundancy. Using the basic idea of non local processing method, each frame image is taken as an image block, and the adjacent frame images are taken as similar image blocks, each of which is arranged into column vectors and constitutes a matrix, which has low rank property. In this way, the problem of removing the salt and pepper noise from the video is converted to the low rank matrix recovery problem. The algorithm can better remove the salt and pepper noise in the video signal and obtain satisfactory quantitative results. 2. an image super pixel generation algorithm is proposed to maintain the boundary. In order to make the generated super pixels fit the image boundary better while giving consideration to compactness, the algorithm redefines a distance metric function that integrates the information of super pixel boundary, compactness and internal consistency, and classifies each pixel. At the same time, by setting the parameters, the proportion between the three is adjusted, and the adjustable balance between the boundary rate and the tightness is realized. In order to better depict the image boundary, the algorithm also proposes a formula to measure the probability of pixels on the image boundary by using the neighborhood pixel information. In order to make the algorithm more accurate in the process of each iteration, an update strategy based on the 3-sigma principle is also proposed in the algorithm. Experimental results show that the super pixels generated by the algorithm can not only fit the image boundary well, but also achieve an adjustable balance between the boundary fitting rate and the compactness. 3. a super pixel generation algorithm based on two times error measure (QEM) is proposed. The image is sampled from the three-dimensional space. If the two-dimensional image is reflected back into the three-dimensional space, then the algorithm in the three-dimensional space can be used in the image processing field. Based on this idea, the image is meshed and extended to three dimensional space, and the problem of image super pixel generation is transformed into a mesh simplification problem. In order to make mesh simplification method better applied to image processing area, the algorithm improves the traditional QEM algorithm, redefines the measure of edge contraction cost, and selects a variety of image attributes to calculate. Through several experiments, it is proved that the method is effective and can generate a variety of super pixels, and can better fit the image boundary.
【学位授予单位】：山东大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】