【作者】 王雷；

【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， 2013， 博士

【摘要】 侧扫声纳自诞生以来在民用和军事领域均得到了广泛应用。随着计算机技术的迅猛发展及数字化的侧扫声纳出现，侧扫声纳图像目标的自动分割和自主识别技术促进了水下智能设备的发展。由于水下复杂环境和其他因素影响，侧扫声纳图像具有噪声污染严重，对比度差等特点。研究人员利用各种预处理方法对侧扫声纳图像进行处理，并研究各种分割算法对侧扫声纳图像进行分割，取得了一定的成果，但依然没有一种成熟的方法。在实际的侧扫声纳图像分割应用中，对算法的稳定性、分割精度和分割速度有着不同的要求。本文深入研究模糊聚类算法和水平集算法改进方法，提高分割精度和算法的稳定性。同时提出改进的分层马尔科夫模型分割算法和利用辅助信息的特定目标分割算法，实现对侧扫声纳图像快速准确分割。通过阅读大量文献，对国内外现状的分析，确定了本文的主要工作：（1）对侧扫声纳图像预处理方法进行研究。通过对侧扫声纳数据进行解码，在原始数据的基础上，提出航向角优化模型和航向角校正算法。对图像进行几何校正，并建立了图像坐标和地球坐标的转换规则。对灰度校正方法进行了研究，进一步提高图像质量。对侧扫声纳图像的纹理描述方法进行了介绍，并提取了侧扫声纳图像的GMRF纹理和Gabor纹理。对侧扫声纳图像滤波算法进行研究，提出一种改进的BEMD的图像滤波算法。为后续的分割工作开展打下基础。（2）对基于聚类的侧扫声纳图像分割算法进行研究。利用常用的几种分割算法对侧扫声纳图像进行处理，找出了其不足之处。利用纹理特征对算法进行改进，分析了初始聚类中心选定的规则，重写了隶属度函数，获得了一定的成果。然后进一步结合改进的BEMD滤波方法，对算法进行融合。通过大量侧扫声纳图片分割实验证明，该算法具有稳定的分割性能，对不同图片均由很强的适应能力。（3）对基于水平集的侧扫声纳图像分割算法进行研究。对CV模型、四相水平集模型和分层水平集模型这些基本模型进行了介绍，并利用这些模型进行了分割实验。在此基础上，结合图像纹理信息，进一步研究了基于GMRF纹理能量驱动的水平集模型和Gabor纹理驱动的模型。通过对这些模型缺点的分析，提出了改进的四相水平集分割模型和快速分层的水平集模型。能够获得更好的分割结果，并一定程度提高算法的分割速度。（4）提出基于分层MRF的侧扫声纳图像快速分割算法。介绍了平面MRF模型和分层MRF模型。针对分层MRF，通过将传统的侧扫声纳图像的目标区和阴影区归为一类，减少了MRF模型参数。利用灰度统计的方法描述图像灰度分布，减少了利用灰度分布模型描述时的参数估算的计算量。通过分割实验验证了算法的快速性和有效性。为了进一步提高分割速度和分割精度，提出了专家数据库辅助的分割算法。获得了良好的效果。（5）对全文提出的算法应用进行了说明，提出了特定目标的分割算法。利用区域检验法对侧扫声纳图像中孤立目标进行分割，通过对特定形状目标的几何空间特征进行分析，提出了区域判断法，结合前文内容，对特定尺寸的目标分割方法进行了说明。最后对前几章中提到的各种分割算法的特点和应用环境进行了说明。本文最后对全文的创新点和研究成果进行了总结，对还需要解决的问题和接下来要做的工作进行了说明和展望。更多还原

【Abstract】 Since the birth of the side-scan sonar, it has been widely used in civil and military fields.With the rapid development of computer technology and the digitization of side-scan sonar,side-scan sonar image target automatic segmentation and self-recognition technology arepromoting the development of intelligent underwater devices. For the reason of complexunderwater environment and other factors, sides-can sonar has some defaults such as imagewith serious noise pollution and poor contrast ratio. Researchers process scan-sonar imagesby using various pretreatment methods and segmentation algorithm to improve the results,and have achieved certain results, but there is still not a mature approach yet. In differentactual side-scan sonar image segmentation applications, the requirements of accuracy,adaptability and speed degree are different. This article focuses on in-depth study of fuzzyclustering algorithm and level set algorithm improved method to improve the accuracy andadaptability. The improved hierarchical Markov model segmentation algorithm and thespecific target segmentation algorithm using auxiliary information, which can quickly andaccurately get the segmentation images, are proposed. After reading a lot of literature andanalyzing domestic and international status, the main work of this paper is determined asfollows:(1) Research on side-scan sonar image preprocessing methods. After decoding side-scansonar data, on the basis of the original data, the heading angle optimization model andcorrection algorithm are proposed. The images are geometrically corrected, and the imagecoordinate and the Earth coordinate conversion rules are proposed. A study on the gradationcorrection method is done to further improve image quality. Side-scan sonar image texturedescription method is introduced, and GMRF texture image and Gabor texture image areextracted. After researching on side-scan sonar image filtering algorithm, an improved BEMDimage filtering algorithm is proposed. They are basis of following segmentation work.(2) Research on side-scan sonar image clustering segmentation algorithm. Scan sonarimages are processed using the several common segmentation algorithms, finding out theinadequacies and insufficiencies, and improving the algorithm by using texture features. Afteranalyzing the initial cluster centers selected rules, the membership function is rewritten toobtain certain results. Then combined with the improved BEMD filtering method, thealgorithm is improved. Experiment results have proved that the algorithm has stableperformance of the segmentation and has a strong ability to adapt to different pictures.(3) Research on level set algorithms of side-scan sonar image segmentation. After the basic model of the CV model, four-phase level set model and hierarchical level set model areintroduced. The segmentation experiments are done using these models. On this basis, aftercollecting image texture information, further studies of level set models driven by GMRFtextures energy and Gabor texture energy are done. Through the analysis of the shortcomingsof these models, improved four-phase level set segmentation model and rapid hierarchicallevel set model are proposed. These models are able to obtain better segmentation results, andhigher segmentation speed.(4) The quickly hierarchical MRF side-scan sonar image segmentation algorithm isproposed. Plane MRF model and hierarchical MRF model are introduced. Through makingthe target area and shaded areas in a class, the MRF model parameters are reduced. As the useof the grayscale statistical methods, the calculation of the parameter estimates described bygray distribution model is reduced. Segmentation experiments verify that the algorithm is fastand effective. In order to further improve the segmentation results and the accuracy ofsegmentation, database of experts assisted segmentation algorithm is proposed. It can obtaingood results.To be continued, the application of the algorithms is described and the specific targetsegmentation algorithm is proposed. Regional inspection method is used to detect isolatedtargets in sonar images. After analyzing the geometric space characteristics of particularshapes, the regional judge method is proposed.(5) The full text of the innovation and research are summarized, and the researchdirection of the next step is described. Combined with previous content, specific size targetsegmentation method is described. Finally, a summary of innovation and research result inthis paper is provided, and the problems needing to be solved and future works needing to bedone are described also.更多还原