【作者】 刘丹丹；

【导师】 袁赣南；

【作者基本信息】 哈尔滨工程大学 ， 导航、制导与控制， 2011， 博士

【摘要】 伴随着海洋矿产资源开发、海洋工程、海洋开发等领域的日新月异的发展,工作在水下机器人、遥控潜水器等多种载体上的声纳系统除了需要胜任极端环境下的工作外,还需要高分辨率的成像、目标识别能力以及多个声纳图像的匹配及识别能力,以满足区域的高分辨率地形地貌测量、准确分辨海底起伏剧烈区域以及沉底小目标和水中目标的探测等需求。由于声纳设备成像的非线性和水下声场环境的复杂性,所采集到的水下声纳图像具有背景噪声包含的灰度级比较丰富,声纳图像的目标区域灰度级相对较少等特点,这些特点对后续的声纳图像目标检测与定位跟踪等工作带来了很大的难度。多分辨率分析属于多尺度细化分析,通过基函数的平移、伸缩等运算对信号进行局域时频分析,从信号中能有效地提取信息,是信号处理的研究热点和前沿课题,是目前国际公认的信号与信息处理领域的高新技术。它在信号滤波、图像去噪、图像分割、图像识别等领域的应用越来越多地受到人们的重视。本论文主要研究多分辨率分析及其在2D声图像及3D序列处理中的应用,主要在声纳图像的去噪增强等预处理、图像区域分割、运动目标分割、目标及敏感区域特征提取、目标跟踪定位和多声纳场景目标匹配等方面进行了研究。首先,对常用的多分辨率变换方法及传统的图像去噪方法进行了概括性的介绍与分析。阐述了声纳的成像机理及水下目标声图像的统计特性。针对水下目标声纳图像不易判别边缘及细节、对比度差等特点,提出了三种基于多分辨率工具的声图像去噪方法,其中包括基于抽样矩阵的Surfacelet变换水下目标声图像去噪方法、水下目标声图像分块自适应降噪方法、基于三维上下文模型的水下目标声图像降噪方法。这些方法具有平移不变性,多向性,图像及序列的空域与时域相关信息利用率高,并且通过仿真实验,验证了这些去噪方法的优越性和有效性。其次,对经典的图像分割方法进行了概括性的介绍与分析。针对高分辨声纳图像富含大量随机噪声、目标区域的边缘不清且断续等特点,基于多分辨率工具以边缘检测算法进行声图像区域识别及修复,提出了基于双边滤波和Surfacelet的边缘检测的声图像分割方法。分割实验表明本文算法能有效去除混响等噪声区域的同时,显著地提高了图像的视觉效果,尤其是在边缘、细节保持方面有一定程度的提高。与马尔科夫随机场(MRF)模型分割方法、小波聚类分割方法比较,获得了更为准确的分割结果。再次,对传统的基于图像及图像序列的运动目标检测及跟踪技术方法进行了概括性的介绍与分析。针对高分辨声纳图像特点,提出了静态背景下基于LIP模型及光流法的运动目标检测算法、复杂背景下基于Surfacelet的声图像序列运动目标检测算法、基于可变图像模版匹配及Surfacelet变换的声图像序列目标跟踪算法。实验表明本文算法在声图像存在较多混响等噪声区域的条件下,能有效检测序列中运动目标。基于可变图像模版匹配及Surfacelet变换的声图像序列目标跟踪算法,与mean shift跟踪算法、SIFT特征匹配的跟踪算法比较,获得了更为准确的跟踪结果。在匹配稳定的情况下,能够及时利用尺度不变特征估计目标尺度变化,克服传统的基于核的跟踪方法不能有效估计目标的尺度及旋转量的缺陷。最后,对常用的多摄像机图像目标匹配方法进行了概括性的介绍与分析。针对高分辨声纳图像序列富含大量信道噪声、散射噪声,不易判别边缘及细节、对比度差等特点,依据人眼视觉系统在图像的不同空间频率及不同区域敏感度特性,提出了基于区域SIFT描述子及Surfacelet的声纳目标匹配算法。实验表明,本文算法能有效提取声图像目标SIFT描述子特征,能够实现多声纳场景目标匹配。与全局应用SIFT描述子进行声图像目标匹配方法相比,基于区域SIFT描述子及Surfacelet的多声纳场景目标匹配算法SIFT描述子数据总量(SIFT特征点数及SIFT描述子维数)较低,计算复杂度有较大降低,并且具有较高的匹配精度。综上所述,本文研究了基于多分辨率方法在2D声图像及3D序列处理中的应用,并针对目前该领域中存在的不足及需求,设计相应算法进行改进与功能实现。仿真实验证实,本文所提出的改进方案和应用的算法,均能够获得很好的效果。更多还原

【Abstract】 With development of ocean resource exploitation and ocean engineer, the sonar need multi-resolution image and target recognition ability and multi-sonar matching and recognize ability besides of working at extremeness environment in order to requirement of high resolution terrain relief measurement and small target detection. Due to sonar equipment imaging no linearity and complexity of under water sound field environment, the background noise of sonar image include enough gray levels and the target area gray levels is comparatively less. The characters bring difficulty for following sonar image target detection and orientation track. Multi-resolution analysis tool can multi-scale details analysis by basis function expansion, translation and other computing, and it can effectively extract information from signals. At present, Multi-resolution analysis is international acknowledged advanced technology in the domain of information and signal processing. Meanwhile it is the front a question for discussion and study hotspot. More and more people attach importance to the application of multi-resolution analysis in the domains such as signal filtering, image denoising, image segmentation and image recognition etc.Multi-resolution analysis and its application in 2D sonar image and 3D sonar sequence processing are investigated in detail in this dissertation. The main work can be summarized on sonar image denoising, image area and movement target division, target and sensitivity area character distill, target tracking and orientation and multi-sonar scenes target matching etc.Firstly, the concepts and denoising methods in common use multi-resolution transform are recapitulative introduced and analyzed. At the same time, the imaging mechanism and the statistics characteristic of underwater acoustic image is expatiated. Through studying acoustic image of character which is bad contrast and deteriorates edges and detail, three the acoustic image denosing methods based on multi-resolution analysis are proposed. They are Sonar image denoising method based on sample matrix and Surfacelet transform, block adaptive denoising method of acoustical imaging and the sonar image denoising method based on 3D context model respectively. These methods are characterized by multi-direction option, parallel information processing, high efficiency of information using at air space and time space, and fusing the enhanced on multi-resolution. By the simulation results, the effectiveness and superiority of the methods are proved.Secondly, the classical image segmentation methods are recapitulative introduced and analyzed. Aim to character of mass random noise and target area edge blur of multi-resolution sonar image, which is recognized and restored based on multi-resolution and edge detection. The sonar image segmentation arithmetic is proposed based on spatial filter and Surfacelet transform edge detection. By the simulation results, the effectiveness and superiority of the methods are proved. It can obtain good segmentation result than MRF model segmentation and Wavelet clustering segmentation and protect edge and detail and improve visual effect.Thirdly, the classical movement target tracking methods based on image and image sequence are recapitulative introduced and analyzed. Aim to character of sonar image, the movement target tracking arithmetic based on LIP model and light flow, the sonar image sequence movement target tracking arithmetic based on Surfacelet transform at complex background, the image sequence movement target tracking arithmetic based on changeable stencil and Surfacelet transform are proposed. The more exact result is obtained the image sequence movement target tracking arithmetic based on changeable stencil and Surfacelet transform than mean shift track arithmetic. Under conditions of matching steady, it can take full use of scale fixedness character to estimate scale change, but mean shift track arithmetic can not efficiently estimate target scale and rotate value.Finally, the classical multi-cameras image matching methods are recapitulative introduced and analyzed. Aim to sonar image character, the sonar target matching arithmetic based on area SIFT description and Surfacelet transform according to sensitivity character to image different space frequency and vary areas of human eye system. The experimental results show that the arithmetic can efficiency distill sonar image target SIFT description character and realize target matching at multi-sonar scene. Its compute complexity is lower and matching precision is higher than whole application SIFT description.In conclusion, on the base of multi resolution analysis, this paper had researched on multi resolution analysis and its application in 2D sonar image and 3D sonar sequence processing. Aiming at the shortcoming of the region, the improved algorithms were proposed. Experimental results indicated that the adopted intelligent optimization algorithms and the proposed methods could attain good results.更多还原