Based On The Content Of Color Image Segmentation

With the rapid development of intelligent information technology, industrial automation, video surveillance, remote education, human-computer interaction technology, video tracking and authentication of security and so on, more and more effort has been taken on the processing of color images. This article use a noparameter clustering arithmetic:Mean shift, In the very beginning, the Mean Shift generally provides a fast search for a point to reach its object controlled by a quantity similar to the minimum decreasing gradient. Given a local region of interest and the corresponding centre of mass (CM), a mean shift vector from a point to the CM can be defined to identify the direction and distance to the CM, and then the point and the local region and is"moved"to the CM. This procedure continues until the densest region is located, as a result modes in a set of data samples can be found and the underlying probability density function can be manifested. The Mean Shift is a sort of non-parametric density gradient estimations. The Mean Shift can bypass the priori hypothesis on the type of entire probability density function for a sample. Not only has it higher performance such as reliability, robustness and flexibility, but also is it of strict convergence. Furthermore, it can partition the image space into a combination of various shapes in an effective way through a series of the iterative process of modes searching. At present, this kind of approaches has been widely studied under a variety of applications, and many basic solutions for a vast of image processing and pattern recognitions has been successfully provided up to now.Based on the specific features of color images, in this thesis, I systematically describe the Mean Shift method, make a strict proof on the convergence of Mean shift sequence of moving sample points, and take a close parametric analysis and sensitive experiments on the design and performance of the Mean Shift algorithm:(1) To explore relevant theoretical basis: the Mean Shift method and its principle, the algorithm description and the convergence proof;(2) To select kernel function: The kernel function in Mean Shift approach is used to distinguish sample points in terms of how much influence a point make to the mean shift vector, meanwhile the kernel has impact on the number of iterations and the clustering accuracy. Hence, I propose to solve the problem using ;(3) To deal with background phagocyte (melt-down) phenomenon: The Mean Shift is able to highlight the objects partitioned from the image space with sharp difference between regions but clear similarity within regions. As an object get very close to its background in terms of certain image properties such as color, texture etc., the phenomenon of background phagocytes will appear in most cases, therefore I make most advantage of the region-overlapping method to tackle this issue.(4) To loose the bandwidth: The Mean Shift can not only determine the number of sampling points, but also it will affect the algorithm on its convergence speed and accuracy. From the point of view of mathematics, in fact, it is more important than the kernel function. However the option of fixed bandwidth could easily lead to under-segmentation or over-segmentation, in his thesis I choose the adaptive bandwidth method for this particular purpose.