【作者】 张菁；

【导师】 刘群；

【作者基本信息】 哈尔滨工程大学 ， 计算机应用技术， 2005， 博士

【摘要】 地雷探测过程包括以下步骤:首先是通过探测器对地下的地雷进行探测,然后对探测结果进行分析生成图像,最后进行图像处理与分析将地雷从背景环境中识别出来。因此地雷图像处理与目标识别是地雷探测系统中的一个重要组成部分。从地雷的组成材料来说,地雷主要分为金属地雷和非金属地雷两种。传统的金属地雷探测器对金属地雷探测效果好。探地雷达对塑料或橡胶地雷的探测效果好。但是总的来说,无论采用何种地雷探测器,如果不借助于图像处理与目标识别技术,地雷的识别率都不高。有的甚至是一千次预报,只有一次是正确的。因此,在现有的地雷探测器的基础上要提高识别地雷的能力,地雷图像处理与目标识别研究无疑具有重要的理论意义和实际应用价值。 近年来,地雷图像处理与目标识别的研究引起了国内外许多学者的关注,但是仍然存在许多问题,主要有以下几个方面:(1)由于塑料或橡胶地雷对探地雷达波的反射特征与石头和土地等其他物质的反射特征相似,如何从其他物质中将地雷准确地识别出来是地雷识别中需要解决的问题。(2)由于探地雷达波长的限制和地面对雷达波的散射使得探地雷达图像中地雷的尺寸要大于其实际尺寸,有时甚至是其实际尺寸的二倍。这给识别塑料地雷的型号增加了困难。(3)在探测地雷之前地下埋藏地雷的种类是无法预知的,因此在实际探测地雷时是将金属和探地雷达探测器同时使用的。处理图像时就必须将两种探测器的图像融合起来,此时要提高地雷的识别率,就涉及到如何提高融合图像质量的问题。 针对以上问题,本文的主要研究工作如下: 1.在研究探地雷达图像特点的基础上,总结出地雷目标的特征,从而提出了基于支持向量机的地雷识别方法。试验表明该方法能够从多个目标中将地雷识别出来,为地雷识别开辟了新的途径。 2.针对地雷图像中目标尺寸大于其实际尺寸,提出一种缩小图像目标尺寸的图像分割方法。并且提出了改进竞争学习的图像分割算法——用权值代替图像目标边缘点坐标值。该方法包括两个步骤:首先是图像的粗分割,获更多还原

【Abstract】 Landmine image processing and analysis are very important in the landmine system. Traditional metal detector works well in metal landmine detection, but it is not useful to the non-metal or minimum-metal landmines, such as plastic landmines. In addition, this kind of landmines is mostly harmful after World War Ⅱ. The new detector-ground penetrating radar and the image processing of ground penetrating radar are needed. Though there are many countries which pay attention to it with many people and lot of money, the result is not satisfied.The main contribution in this thesis is as follows. Firstly a new method of recognition of landmines is put forward by using the support vector machine. Then the competition learning algorithm is improved to obtain the landmine image segmentation. Furthermore, a new method of registration is given for image fusion. Finally, all these methods are put into an image model.The detail researches are1. Landmines recognition by support vector machine. The ground penetrating radar image of landmine has its own characteristic that it is formed by three hyperbola, the others are one or two hyperbola. From this point of view, a new method of landmine recognition is put forward by support vector machine.2. Image segment is a primary step in image analysis of landmines detection by ground penetrating radar sensor which is accompanied with a lot of noises and other elements that affect the recognition of real target size. In this thesis a new theory brought forward, that is, I look the weight sets as target vector sets which are the new cues in semi-automatic segmentation to form the final image segmentation. The experiment results show that the measure size of target with our method is much smaller than the size with other methods and closes to the real size of target.3. Landmine image registration approach by metal sensor and ground penetrating radar sensor is presented. I go forward an approach which is based oncompetition learning and support vector machine. Firstly, I describe the feature of landmine image by the weights of competition learning in order to decrease the vector numbers of being processed and pay attention to the interesting region. Secondly, put the weights into support vector machine as training vectors, then to obtain the support vectors in a cycle way. I demonstrate the efficiency of this approach by applying this method to three group of registration and fusion of metal sensor images and ground penetrating radar sensor images. The results show that the approach is feasible and would be the base for further image processing of landmine.4. On the base of the analysis of landmine image processing and object recognizing, a landmine image processing and object recognizing model is proposed.Methods here are especially for landmine image processing which are all according to the feathers of landmine and ground penetrating radar detector. They would be useful to the theory and method of landmine detection system.更多还原