【作者】 王彪；

【导师】 董晨钟； 李淑秋； 黄海宁；

【作者基本信息】 西北师范大学 ， 原子与分子物理， 2005， 硕士

【摘要】 基于图像处理的声纳目标识别是近年来声纳信号处理领域内研究的一个热点问题和难点。随着一些成像声纳系统的研制成功,目标的识别不仅仅停留在一维回波的处理上,各国学者及工程技术人员将着眼点放在了成像声纳系统。本文主要围绕水声图像目标识别,将图像处理技术和模式识别技术应用于声纳图像中,完成了整个识别过程的算法研究,包括图像识别前的去噪、图像预处理、图像分割,目标的探测和识别技术。 本文首先对声纳图像的预处理中的重点问题——图像去噪进行了细致而深入地研究,提出两种去噪方法,即小波去噪和全变差降噪,并对两种去噪方法进行了对比研究。仿真结果表明上述两种方法在去噪的同时可以很好的保留图像的细节信息,基于微分方程地全变差图像去噪的效果要比小波去噪好,去噪后图像的信噪比高。 对于目标地探测,本文采用两种探测方法,模板匹配和分形探测。模板匹配很好地模拟了声纳图像的特性;而分形探测可以很好地将自然物体和人工目标区分开来,由于两种探测方法都有其各自的特点,将两种方法结合起来,充分利用各自的优势,可以很好地探测出可疑目标区域,利于下一步的处理。 对可疑区完成探测后,下一步工作就是完成该区域到特征空间地映射,对可疑目标区域提取尽可能反映目标特性的特征,组成特征矢量。文中针对声纳图像的物理特征提取了一组基于声纳图像的统计特征的同时,对目标发生平移、旋转或是比例伸缩,采用不变矩特征,可以很好的适应目标的这种变化,将目标准确识别。 在分类器的设计上采用了神经网络中的径向基网络,研究了径向基网络在声纳目标识别中的应用,并与传统的BP网络进行了比较,结果表明径向基网络在分类器的设计上有很大的优点。对侧扫声纳图像进行识别算法的验证,识别效果是令人满意的。更多还原

【Abstract】 Target recognition based on image processing method in sonar image is a hot and difficult issue in sonar signal processing. With the development of some imaging sonar system, target recognition can be performed not only in one dimension but also in two or three dimensions. More and more scholars and engineers pay attention to the imaging sonar system. This dissertation applied some methods of image processing and pattern recognition to solve the target recognition issue of sonar images. It includes the whole aspects in image recognition processing algorithms such as denoising, image pre-processing, image segmentation, target detection and classification.Firstly, this dissertation lucubrated the image denoising method which is the key step of image pre-processing. Two denoising methods were adopted in the dissertation, namely wavelet denoising and total variation denoising based on Partial Differential Equation (PDE). The result indicated that these two methods could preserve the image features and the method based on PDE with higher SNR performed better than wavelet denoising.Two methods were proposed in this dissertation to solve the problem of target detection. One is named Mask Matching, which made use of the characteristics of sonar image. The other is Fractal Filtering, which could classify natural targets from artificial targets. Utilizing their individual characteristics and advantages we could detect suspicious region easily.The next step is mapping the images to feature space. Through extracting features of suspicious region and constructing feature vectors we can carry out the task. According to the physical trait of sonar image, we extracted some statistic features of sonar image. Furthermore we studied the moment invariants to fit the diversification of translation invariants, scale invariants, and rotation invariants of target.Additionally the dissertation studied the RBF neutral network and applied it to classify target of sonar image. Compared with BP network, it showed that RBF neutral network has some advantages of designing classifier and its simulating result is satisfying.更多还原