【作者】 李彦鹏；

【导师】 梁甸农； 庄钊文；

【作者基本信息】 国防科学技术大学 ， 信息与通信工程， 2004， 博士

【摘要】 自动目标识别具有巨大的理论意义和应用价值,长期以来,对识别效果评估的研究相当有限,效果评估成为目标识别中最迫切需要解决的重要问题。本文将其作为研究方向,主要结合雷达目标识别效果评估展开。 针对所研究的问题,文中独创性地提出一种目标识别效果评估整体思路:在定量刻画识别系统所处条件的基础上,考察能够反映识别效果关键特性的评估指标,利用客观的、有效的评估模型得出定量的评估结果。评估模型必须具备较强的通用性,还要能够直接用于工程实践。本文研究工作分为几个部分: 首先, 找出识别效果评估解决思路,安排整体工作。 第二步,研究刻画识别过程所处条件的方法,即,评估参照信息.的选择及测度。文中选择的参照信息具备易于获取、易于测度,能够给出客观度量的特点。 第三步,提出识别率测试结果的概念,通过研究其特性得出一系列刻画识别效果各方面特性的评估指标,进而确立了识别效果评估指标体系。这一步集中地、系统地、有效地解决了评估工作中样本容量的解析解、给出评估结果置信度、通过选择样本容量控制评估过程中两类风险的水平、识别效果与外界条件的独立性判断、如何设计满足一定要求的实验这些关键难题。 第四步,基于不同的数学理论,构建了几大类识别效果评估模型: 1.将模糊综合评判运用于目标识别效果评估,着力解决了因素集的选取、评判矩阵的生成、多级模糊综合评判等问题,建立了数个应用于不同场合的评估模型。为了避免在被评估系统的识别效果发生一定变化的情况下得出错误的评估结论,完善了现有的变权模糊综合评判方法,形成基于变权模糊综合评判的评估模型。 文中结合已有的试探法求解模糊关系方程提出了迭代法求解模糊关系方程并用于处理本评估模型的自学习问题。 2.使用模糊积分可以获得较为直观的评价结果,文中建立了基于模糊积分的识别效果评估模型。这里主要解决了单个和多个待识别目标时向量的生成问题,同时,针对现有二重模糊积分方法在应用中的不足,提出修正的多重模糊积分理论并予以运用和建模。 修正的多重模糊积分在每重积分上可以同时处理多个可测函数和对应的可能性测度,因更多还原

【Abstract】 Automatic Target Recognition (ATR) has great value not only in theory but also in application, for a long time, performance evaluation of this technology is very limited, now, performance evaluation becomes the most important problem that should be solved urgently. This direction is chosen as the focus of the thesis. We arrange the content aiming at the performance evaluation of Radar Target Recognition.Under analyzing the problem thoroughly, we present an creative idea in sum: on the basis of depicting the environment in quantity, study some key indexes, draw a quantitative result through object, effective performance evaluation model. The model must be universal and can be applied in engineer directly. The work is divided into some parts:Firstly, we find the scheme and arrange the work.Secondly, selecting some measurements to describe the environment for ATR system, as is to say in the article, offering the reference information for performance evaluation. The reference information here can be acquired and measured easily, and give objective result.In the third step, we give the concept called measurement of recogmtion rate (MRR). From the characteristics of MRR, a succession of evaluation indexes is educed. We have established the system of performance indexes. Here, Such problem as sample capacity’s analytic solution, presenting the confidence extent of the evaluation conclusion, manipulating the level of two kinds of risk by different sample capacity, independence test between the recognition result and the environment, designing an experiment according to certain require are solved in a centralized, systemic and effective way.Subsequently, some evaluation model is developed based on different mathematics theory.1. We admit fuzzy comprehensive evaluation into ATR performance evaluation successfully; some evaluation models were built in conformity to different application. The key problems are settled like forming evaluation index in performance evaluation of multi-target recognition, selection of factor set, the eduction of evaluation matrix and multi-step fuzzy comprehensive evaluation. In order to avoid improper conclusion when the performance of thesystem being evaluated has changed, we perfect the weight-variable fuzzy comprehensive evaluation in existence; as a result, the performance evaluation model based on weight-variable fuzzy comprehensive evaluation is formed.During the evaluation course, the experimenter may need to settle the weight set in some situation; we advance an iterative method to work the fuzzy relation equation, on the basis of some ways in fuzzy mathematics.2. We can draw a concise result by fuzzy integration, so, performance evaluation model based o n fuzzy i ntegration i s c onstructed. We p robe t he m ethod tow ork o ut t he j udgement vector. Aiming at the deficiency of fuzzy integration in application, we put forward multi-layer fuzzy integration, then, establish an evaluation model with it. The modified multi-layer fuzzy integration is able to manage multi measurable function and their possibility measure; it can be used in dealing with complex problem directly in practice. The self-study in this kind of model is resolved.3. We study the technique to draw fuzzy resemble matrix between the objects being evaluated and fuzzy resemble matrix between different elements in the comment set; so, these means can be used in our work. Performance evaluation model based on fuzzy cluster analysis is built. It can be used to evaluate multi systems at one time quickly and draw useful conclusion.4. Performance evaluation with the help of fuzzy run theory. In order to remedy the information loss in Mini-Max operation that is included in fuzzy comprehensive evaluation, we found fuzzy run theory. It works with bottom run, variation in bottom run, and run between different dimensions, as desired by us, it analysis the information contained in the judgement matrix comprehensively. In consequence, this theory is used in work, and is illustrated through some examples.5. The fuzzy comprehensive methods may lose information in recognition results, at the same time, it can not get continuing output while the recognition result is changing in continuing way. We form some variable from the recognition result based on measure theory; in consequence, the evaluation model has been built.6. For an ATR system, it is of great importance that the performance can be stable while the situation is changing appreciably. We find the dynamic relation model of ATR system to the situation, then, with the help of Lyapunov theory, the stable condition is discussed.In addition, we analysis some typical factors related to ATR system, the equation of reference information is found.Finally, we carry out performance evaluation simulation to four ATR algorithms by the help of test data and synthesis data. It is proved by the simulation mat the idea of our work is right; the performance evaluation model can be used in practice directly, it have some characteristics: l)the result is objective, quantitive, and reflects most aspects of the recognition performance; 2)the situation is admitted in the evaluation model; 3)it can be used under a certain environment to make static state evaluation and stability evaluation, at the same time, it can be applied under changing environment to draw a dynamic evaluation; the variation interval can be all over the definition field or part of it; 4)it can compare systems in same design as well as in different design; 5)the performance model is a opening model, it can be configured according to the content being considered.From the data obtained, this is the first time to form a general quantified performance evaluation system in ATR.更多还原