【作者】 徐曼；

【导师】 王乘；

【作者基本信息】 华中科技大学 ， 系统分析与集成， 2010， 硕士

【摘要】 图像分割是图像处理过程中基础且重要的阶段。传统的图像分割技术包括阈值分割法,边缘检测法,区域分割法等三种常见的方法。近年来,各种新技术和新理论也被广泛应用于图像分割,产生了许多种新型的图像分割法。现有的图像分割技术不断改进,一是为了提高图像分割结果的精确度;二是为了提高图像分割效率;三是为了使图像分割方法具有更广的适用性。分水岭算法因原理直观,算法效率高,检测边缘线连续封闭,能响应微弱边缘等优点在图像分割领域占有非常重要的地位。本文深入研究和对比了两种分水岭算法的实现过程,并在效率更高的快速分水岭算法上进行了多方面的改进,提高了分水岭算法的分割精度和分割效率,改进了现有分水岭算法图像分割技术。本文基于快速分水岭算法进行了以下四方面的改进:(1)改进了快速分水岭算法中错误积水盆地的识别方法和纠正方法,将标记号不同的两相邻积水盆地的合并情况分为三种。(2)将基于四邻域的浸没改进为基于八邻域的浸没,避免了对角方向上同一个区域划分为两个区域。基于八邻域的浸没分割结果更准确,并且分割区域数少于基于四邻域的浸没。(3)将逐层浸没改进成多层浸没。比较了逐层、5层、10层三种浸没的分割结果,5层浸没的分割结果不仅保持了图像中的边缘,且分割区域数明显少于逐层浸没。5层浸没很大程度上缩短了后期区域合并时间。(4)区域合并过程中利用区域色度均值信息和相邻区域间分水岭两侧梯度均值信息确定两相邻区域相似度。最后,本文对V-S分水岭算法,快速分水岭算法,改进的快速分水岭算法三种图像分割算法做了对比实验,对多幅图像进行仿真和分析,获得对比结果。更多还原

【Abstract】 Image segmentation is an important part in image processing. Traditional methods of image segmentation can be partitioned into three main types: thresholding, edge detection, region segmentation and so on. More and more new techniques and special theories have been applied to improve image segmentation recently. The purposes of improving these methods are: higher accurate on edge position, faster and broader applications.The watershed transform algorithm raised by Vincent and Sollie, briefly named as V-S algorithm, have received wild attention for its easy principle, efficient algorithm and continuous and integrated border lines. There are three kinds of improving methods of watershed algorithm: using filters to preprocess, marking before watershed transform and region merging after watershed transform.Fast watershed algorithm has the same principle of V-S algorithm, but the former is obviously faster than the latter, because the former improves the second step of V-S algorithm. Four following improvements on fast watershed algorithm were made:(1) Concluded three cases of correcting of error minimum areas, which made the process of correcting clearer.(2) Immersion process was based on eight neighborhood, not on four neighborhood any more, which could segment more accurately and region number was smaller.(3) Former immersion was based on 1 layer, and the improved immersion process was based on multilayer, such as 5 layers and 10 layers. Compared three kinds of immersions, multilayer immersion could efficiently relieve the phenomenon of over-segmentation.(4) Region merging: combining mean chromaticity information and edge both sides mean gradient information to calculate Regional similarity.Finally, based on the improved fast watershed algorithm, an efficient, real-time image-segmentation system was designed, which can achieve good segmentation results on different types of images.更多还原