【作者】 王兴梅；

【导师】 叶秀芬；

【作者基本信息】 哈尔滨工程大学 ， 模式识别与智能系统， 2010， 博士

【摘要】 随着海洋开发的快速发展,声纳技术作为水下探测的重要手段,在水下导航定位、目标跟踪识别和通信等方面发挥着越来越重要的作用。对水下声纳图像进行目标识别己经成为数字图像处理领域的一个重要课题。水下声纳图像目标检测和特征提取是水下声纳图像目标识别的关键步骤。基于马尔可夫随机场的检测方法能准确地描述每个像素所属类别与周围像素类别之间的重要依赖关系;基于水平集的轮廓提取方法具有很强的拓扑性,能对形状不规则以及具有空洞、重叠的水下目标进行轮廓提取。因此,本文重点对基于马尔可夫随机场的水下声纳图像目标检测和基于水平集的轮廓提取方法进行了深入的研究。主要包括如下内容:声纳图像的灰度分布模型在一定程度上,对基于模型的水下目标检测性能产生较大的影响。所以,在分析声纳图像成像原理及其特点之后,重点对声纳图像灰度分布模型进行了研究。目标高亮区灰度分布满足线性直线方程,并且这个方程只在某一范围内有意义,根据现有目标高亮区的规整化线性方程分布模型,提出一个正比例分布模型来较精确地描述目标高亮区的灰度分布。在基于马尔可夫随机场的水下声纳图像目标检测部分,针对目前三类检测中平面马尔可夫随机场模型参数集合结构复杂的问题,提出了一种既有较高精度又具有较少参数估计计算量的模型参数集合;根据目前已有的两类不完全分层马尔可夫随机场模型参数集合结构的特点,提出将新建立的三类平面马尔可夫随机场模型参数和层次间相互作用的模型参数,应用于不完全分层马尔可夫随机场三类检测中,得到最终精确的三类检测结果。此外,为了进一步提高检测精度,背景区灰度分布采用Gamma分布模型描述,提出一种基于马尔可夫随机场改进的水下声纳图像目标检测方法。在基于水平集的水下声纳图像目标轮廓提取部分,利用水平集方法对整幅声纳图像处理,可能会将背景噪声当成目标高亮区或阴影区进行了轮廓提取,因此,提出利用基于马尔可夫随机场的水下声纳图像目标检测结果确定目标演化子区域,缩小目标区域范围;为避免初始闭合曲线选取不当,不能同时提取目标高亮区和阴影区的轮廓,提出在目标检测结果中,根据目标高亮区和阴影区的位置,确定各个目标演化子区域初始闭合曲线的中心坐标,通过Vese-Chan分段常量四相水平集方法的演化函数进行目标高亮区和阴影区的轮廓提取。论文最后总结了本文的研究成果和创新之处,并对下一步工作进行了展望。更多还原

【Abstract】 With the rapid development of exploitation of marine fields, sonar technology plays an important role, in underwater navigation and positioning, object tracking and recognition, communications and so on. It has already become a significant issue in digital image processing field that the technology of object recognition of sonar image. The objects detection and feature extraction are key steps in underwater objects recognition processing. Owing that the detection method based on MRF (Markov Random Field, MRF) can precisely describe the dependent relationship between the classes of every pixel and its neighboring pixels, and the contour extraction algorithm based on level set gave a complete consideration on topological property, and it is effective even the underwater objects with hole, overlapping, and irregular shapes. This paper is committed to research on underwater objects detection based on MRF and contour extraction based on level set of the sonar image. The main research contents of this paper are shown as following:The grey distribution model of sonar image has great influence on underwater objects detection performance based on the model. On the base of analyzing sonar imaging principle and sonar image features, sonar image grey distribution model is researched. The object-highlight region grey distribution satisfying linear function which is applicable in certain range. According to current object-highlight region distribution model depicted by normalized linear equation, a proportional distribution model is proposed to accurately describe object-highlight region grey distribution.In the part of underwater sonar image objects detection based on MRF, regarding to the set of plane MRF model parameters complex problems in three-class detection ways, the new three-class plane MRF model parameters set is proposed which is high accuracy and can decrease the calculation work of parameters estimation. According to the structure property of incompletely hierarchical MRF of two-class, we propose to apply the new three-class plane MRF model parameters and model parameters of the level interaction into incompletely hierarchical MRF of three-class detection. Then a final accuracy three-class detection result will be obtained. Furthermore, in order to further improve detection precision, the Gamma distribution model is used to describe the background region gray distribution. Then an improved underwater sonar image objects detection method based on MRF is put forward.In the part of underwater sonar image contours extraction based on level set, the application of level set model to the full sonar image may cause background noise being mis-extracted as object-highlight region or shadow region due to topology of the model, so object evolution sub-region, which is determined by the MRF underwater objects detection results, is proposed to dwindle the search region. Besides the accuracy of contours extraction are also relative to the initial position of closed curves. To exactly extract the contours of object-highlight region and shadow region, the location of initial closed curves in each object evolution sub-region should be selected appropriately. In this paper, the centers coordinates are determined by the location of object-highlight and shadow regions.The contours of object-highlight and shadow regions are extracted by the four-phase piecewise constant Vese-Chan level set evolution functions. The paper summarized the research results and innovations, and look forward to the next work.更多还原