基于压缩域多特征融合的图像分割算法研究

发布时间：2018-11-20 18:16

【摘要】：二十一世纪以来,伴随着科学技术的迅速发展,人们每天都要处理大量的信息数据(例如图像、视频以及文档资料等)以便进一步地分析和研究。图像作为常用信息的载体之一,在人们接收、传递和处理信息的过程中起着至关重要的作用。所谓图像分割是指将图像中具有相同颜色、亮度、纹理等“特殊意义”的不同区域分割开来,并使这些区域互不相交,同时每个区域应满足特定区域的一致性条件[1]。现有的图像分割技术在实际使用过程中存在着较大的局限性,如分割算法复杂度高、鲁棒性较差、人工干预过多、复杂背景下分割目标不准确等,因此,图像分割是机器视觉技术中的难题之一,也是近年来国内外学者研究的热点。伴随着数字图像采集设备的迅猛发展,当前数字图像的分辨率越来越高,图像的尺寸也越来越大。因此,一种快速、高效的图像分割方法显得尤其重要。本文的主要研究内容是基于图论与谱聚类的图像分割方法,利用改进的离散余弦变换(Discrete Cosine Transform,DCT)对图像预处理得到基于DCT的方形结构块(DCT-SBS),使用得到的DCT-SBS构造图节点,然后提取每个节点的颜色信息、纹理信息以及位置信息等,利用新的多特征融合方法计算图边权值,最后使用本文提出的基于DCT-SBS的图像分割方法实现图像分割。本文的具体工作如下:1.介绍了基于图论的图像分割基本方法,针对传统分割方法使用图像像素构造图节点时计算复杂度随着节点增加大大增加的问题,本文使用DCT-SBS构造图节点,不仅降低复杂度,同时节点保留了原始图像数据的结构信息。2.在计算图边权值时,本文提出了新的基于压缩域下多特征信息融合的计算方法,有效地将颜色信息、纹理信息以及位置信息用于计算图边权值。3.在传统谱聚类算法的基础上提出了一个通用的谱聚类算法框架及其求解过程,且该框架可以通过调整参数转变到不同的图像分割算法。4.将本文提出的算法与8种不同的图像分割算法进行实验对比。为了验证本文提出的图边权值计算方法的有效性以及谱聚类算法框架的综合性能,本文在Corel1000数据集、MSRA10K数据集上对本文提出的分割算法进行了验证,实验结果表明,本文提出的图像分割方法具有较高的分割准确率以及算法效率高等特点。在现实情况下具有一定的理论意义和应用价值。
[Abstract]:Since the 21 century, with the rapid development of science and technology, people have to deal with a large number of information data (such as images, videos and documents) every day for further analysis and research. As one of the commonly used information carriers, image plays an important role in the process of receiving, transmitting and processing information. The so-called image segmentation refers to the segmentation of different regions with the same color, brightness, texture and other "special meanings" in the image, and makes these regions disjoint, at the same time, each region should satisfy the consistency condition of a particular region [1]. The existing image segmentation techniques have some limitations in actual use, such as high complexity, poor robustness, too much manual intervention, inaccurate segmentation targets in complex background, and so on. Image segmentation is one of the difficult problems in machine vision technology. With the rapid development of digital image acquisition equipment, the resolution of digital image is becoming higher and higher, and the size of image is becoming larger and larger. Therefore, a fast and efficient image segmentation method is particularly important. The main content of this paper is the image segmentation method based on graph theory and spectral clustering, using the improved discrete cosine transform (Discrete Cosine Transform,DCT) to preprocess the image to obtain the square structure block (DCT-SBS) based on DCT. Using the resulting DCT-SBS to construct the map nodes, then extracting the color information, texture information and location information of each node, and using a new multi-feature fusion method to calculate the edge weights. Finally, the image segmentation method based on DCT-SBS proposed in this paper is used to realize image segmentation. The specific work of this paper is as follows: 1. This paper introduces the basic method of image segmentation based on graph theory, aiming at the problem that the computational complexity increases greatly with the increase of nodes when traditional segmentation methods use image pixels to construct graph nodes. In this paper, the complexity of constructing graph nodes by using DCT-SBS is not only reduced. At the same time, the node retains the structural information of the original image data. 2. In this paper, a new method based on multi-feature information fusion in compressed domain is proposed to calculate the edge weight of the graph. The color information, texture information and position information are effectively used to calculate the edge weight of the graph. Based on the traditional spectral clustering algorithm, a general spectral clustering algorithm framework and its solution process are proposed, and the framework can be changed to different image segmentation algorithms by adjusting the parameters. 4. The proposed algorithm is compared with eight different image segmentation algorithms. In order to verify the validity of the proposed method and the comprehensive performance of the spectral clustering algorithm framework, the proposed segmentation algorithm is validated on Corel1000 data sets and MSRA10K datasets in this paper. The image segmentation method proposed in this paper has the characteristics of high segmentation accuracy and high algorithm efficiency. In reality, it has certain theoretical significance and application value.
【学位授予单位】：西南大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】