【作者】 刘波；

【导师】 林焰；

【作者基本信息】 大连理工大学 ， 船舶与海洋结构物设计制造， 2013， 博士

【摘要】 机器视觉在海洋工程中用于对被观测物体进行视觉监控、精密定位和非接触测量。随着海洋开发和研究的深入,以及国防安全的需要,作为海洋工程高新技术研究的重要组成部分的智能水下探测器得到了广泛的应用。智能水下探测器在环境恶劣且复杂多变的海洋里作业时,常常要进行定位、识别、避障与路径规划等动作行为,因而使得其视觉系统显得尤为重要。对水下机器视觉系统的研究是一项具有挑战性的课题,具有重要的理论意义和实用价值。特征检测是图像处理中的一项基本技术。根据图像匹配、合成及对象辨识的要求,特征检测方法需要较精确的特征定位,并能同时检测出角点、边缘等图像中不同的显著性特征。本文对机器视觉水中图像(包括水下图像和水面图像)特征提取与匹配技术中下述重点内容进行了研究。(1)基于蚁群优化技术的水中图像分割算法。图像分割是图像理解与图像识别的基础,其分割质量的好坏对后续图像处理的效果会有直接的影响。就水中图像的特点来说,其图像具有模糊性与信息的不可加性,传统的图像分割方法不能满足要求。因此,本文基于蚁群算法,设计了一种智能化的水中图像分割方法,该方法利用图像分割技术的本质,将水中图像里各个像素进行自动分类,最终达到分割图像的目的。在像素分类的过程中,通过引入信息熵和均值聚类等概念,对基本蚁群算法进行了改进,从而使得本方法在水中图像分割中具有自适应性、鲁棒性、并行性和快速收敛性等特点。(2)经验模式分解算法和相位信息相结合的水中图像检测分析技术。二维经验模式分解算法可以实现图像的多尺度结构分析,能够对图像进行融合、降噪、边界特征提取和图像压缩等方面的处理；相位信息是图像中最稳定、最重要的特性之一。因而本文在对经验模式分解算法和相位信息进行分析与综合的基础上,提出了一种用于水中图像特征检测分析的EP模型。该模型充分继承了上述两种方法的优点,可以用于水中图像的处理与分析,实现了水中图像的多尺度、多像素边缘特征提取,提高了图像中目标的匹配定位精度,还可以完成图像的多尺度分割。(3)基于尺度不变特征检测的水中图像匹配技术。针对特征点匹配对于尺度变化比较敏感的问题,本文基于SIFT特征对图像的旋转和尺度的不变性,以及对于噪声、光照变化和视角改变等具有良好鲁棒性等优点,并考虑到水中弱光环境特征和不同的实验场合,提出了一种改进的基于SIFT特征的水中图像配准策略。该方法有效地提高了水中图像匹配的精度和速度,较好地解决了尺度变化给图像配准带来的影响,使得在水中进行较大尺度变化的图像匹配拼接成为可能。(4)基于纹理特征的图像型船舶尾流分类辨识技术。通常来说,船舶尾流的诸多特性与船舶的船体线型、主机和螺旋桨的布置及参数等因素有关,并与船舶航行时的浮态与速度、所航行海域的海况、温度、盐度和海水密度分布等信息也有关。本文以图像型船舶尾流为研究对象,采用局部二进制模式和灰度共生矩阵方法,检测海面尾流的自然纹理形貌和共生统计特征,并将提取出来的特征用作BP神经网络的分类输入向量,建立了图像型船舶尾流自动分类辨识系统,经过对五种航速下的尾流图像的识别测试,实验结果表明用该方法进行的尾流目标检测平均正确识别率可达到80%以上。更多还原

【Abstract】 Machine vision is used for visual monitoring, precision positioning and non-contact measurement of the observed objects in ocean engineering. With the further advance of ocean research and the national security needs, the intelligent underwater detector has been widely used as an important part of ocean engineering high-tech research. The actions such as object location, identification, barrier-avoiding and path planning are often taken by the smart underwater detector in the underwater surroundings, so complexities and uncertainties of working environments make the vision system of the detector stand out especially, further upgrading the performance of machine vision system for underwater research is not only a challenging task, but also has important theoretical significance and practical value.The extraction of the features is the basic technique of the image processing. According to the need of the matching and object recognition and image synthesis, the method of feature detection can detect the different features of corners and edges and precision of the characteristics location are relatively good. This dissertation deals with the following key issues in machine vision underwater and surface image feature extraction and matching techniques research.(1) Underwater image segmentation based on an ant colony optimization algorithm. Image segmentation is the basis of the image understanding and recognition, and the segmentation quality can directly affect on the results of the subsequent image processing. However, as far as the underwater image is concerned, the fuzzy and non-additive information of the image make the traditional method of image segmentation hardly meet the requirements. So a method of image segmentation based on the intelligent ant colony algorithm is designed in this paper, each pixel of an image is classified by analyzing the nature of image segmentation, and ultimately achieving the purpose of image segmentation. In the classification process, the basic ant colony algorithm has been improved by adopting the concepts of the entropy and clustering method, which making the underwater image segmentation program with self-adaptability, robustness, parallelism and fast convergence and so on.(2) The detection of underwater image based on the empirical mode decomposition algorithm and the phase information. Two-dimensional empirical mode decomposition algorithm can achieve multi-scale image structure analysis, and deal with some problems of image such as the image fusion and noise reduction and feature extraction and the image compression and so on; in addition, the phase information is one of the most stable and important features of an image. Therefore, the EP model is proposed based on the feature detection for underwater image analysis by analyzing and synthesizing the empirical mode decomposition method and the phase information. This model is fully inherited the advantages of the two methods, and can be used for underwater image processing and analysis. The multi-scale and multi-pixel edge detection is achieved for an underwater image. And the positioning accuracy of the matching target is also improved. The multi-scale image segmentation can be completed based on this model.(3) The underwater image matching technology based on the scale-invariant feature detection. Feature points matching algorithm is sensitive to the image scale change. In order to overcome this problem, an improved SIFT-based image registration scheme is proposed. The improved registration strategy can solve the above problem by using the rotation and scaling invariant property of the SIFT feature points as well as its robustness to added noise, illumination change and viewpoint change, and taking into account underwater environment with the low-light characteristic and different experimental situations. The proposed algorithm is effective in improving the accuracy and the speed of the image matching, as well as solves the scale changes to the influence of image registration, which makes it possible to achieve successful underwater image registration and mosaic when large scale change occurs.(4) The problem of ship wakes identify based on the image texture feature. In general, many features of the ship wakes relate to the information of the ship’s hull, geometric scale, the host’s position in the boat, the propeller geometry parameter and its working conditions and other factors. In addition, the ship’s speed and the navigation direction, as well as the information of the salinity, temperature, and density of sea water are also relevant with the ship wakes. In a word, the ship wake is widely studied in the ship design, marine environment monitoring and remote sensing areas, as well as in naval air force reconnaissance. It plays important roles in actual practice and military. This paper, taking a ship wakes image as an object of research, using the local binary patterns and the GLCM method to detect the natural texture and the statistical characteristics of the symbiotic as the input vectors of the BP neural network, establishes an automatic identification system for the ship wakes image. This method is applied to sort and recognize the ship wakes of five different speeds images, the result shows that the detection accuracy is satisfied as expected, the average correctness rates of wakes target recognition at the five speeds may be achieved over80%.更多还原