【作者】 周振华；

【导师】 张桂林；

【作者基本信息】 华中科技大学 ， 模式识别与智能系统， 2006， 硕士

【摘要】 性能评估是通过向ATR系统输入指定图像序列,统计分析ATR系统产生的数据,并根据评估数据发现系统缺陷,从而调整系统参数,提升系统性能的有力手段。性能评估需要大量的各种条件下的图像作为ATR系统的输入,但是采集大量的图像需要耗费巨大的人力物力,这种情况促使图像合成方法被日益重视。同时必须对评估结果数据加以有效地分析,才能获得被评估系统的真实性能。本文针对已有性能评估中的图像合成方法不足,提出利用温度预测模型和纹理置换方法,模拟在指定条件下红外图像,并且对评估系统结果数据采用各种回归模型和插值方法进行了分析。本文包含的内容有:一种温度预测模型,两种随机场纹理模型及它们的参数估计方法、合成图像背景区域分布构造方法以及性能评估数据回归分析和插值方法。每一种地表的表面红外辐射度随天气时间变化而变化,不同地表的这种变化是有差异的。根据采集的红外图像,模拟指定天气条件下的红外图像,必须考虑这种变化差异。本文利用已有的一维温度预测模型预测不同地表的温度,并转换成红外辐射度,然后转换成模拟图像各像素的灰度值,从而获得指定天气时间条件下的图像。对合成图像背景某地表区域进行纹理置换,可能会遇到纹理图像尺寸不够大的问题。为了解决该问题,本文采用纹理模型对已有纹理图像进行模拟,生成所需尺寸的纹理图像。本文介绍了马尔科夫随机场模型和一般长纹理模型,并给出了这两种模型参数的估计方法。纹理置换的另一个问题是置换区域问题。要进行纹理置换,首先要获得合成图像背景区域分布,这就需要对背景图象进行区域编码。本文采用两种方式对背景图象编码,一是图像分割方法,而是人工几何方法。本文详细介绍了采用马尔科夫随机场图像分割算法和各方向灰度方差特征分类方法进行背景区域编码。评估数据的分析采用了线性回归和非线性回归两种,其中非线性回归方法采用多项式回归,插值方法则采用分段多项式以及B-样条插值方法。本文采用多项式回归、分段多项式插值及B-样条插值对评估数据进行拟合实验,取得了一定的效果。更多还原

【Abstract】 Performance evaluation is a strong method to improve ATR system performance by inputting special image data to ATR system, analyzing result data and finding disadvantages of it and adjusting system parameters. Performance evaluation needs lots of image data as the input of ATR system, but collecting image data is very costly, and this situation urge the research of the image forming technology. The result of performance evaluation must be effectively analyzed so that real performance of the system can be got. Aiming at the shortcoming of original image forming technology, this paper give the solutions by modeling temperature of the surface of different terrain regions and replacing texture to simulate IR image in given conditions and using several recursion methods to analyzing performance data. This paper include temperature of surface of different terrain regions ,texture models and parameters estimation method ,method of coding different terrain regions of background and recursion methods of result data.Radiance of different terrain regions change with weather and time and the change of those regions is different .It must be considered in simulating image of given condition from real image. This paper use temperature model of terrains to predict temperature of given condition, then transforms it to radiance and then get the gray level of result image. Replacing the texture of a region in background image will meet a problem. It is that the scale of texture image is not big enough. To overcome this problem, this paper use texture models to simulate texture in needed scale. MRF models and GLC models are introduced here and methods of parameters estimation. Another problem is that we must get the distribution of background. Here are two methods. The first one is image segmentation technology; the second one is geometric method .We introduce two image segmentation methods. One is based on MRF; another is based on features of deviation of four orientations. Linear regression and non-linear regression methods are used in performance evaluation data. Non-linear method used in this paper is multinomial. Subsection multinomial and b-spine methods are also introduced. The three methods have been used in evaluation system.更多还原