Radar Target Identification Based On Range Profiles

Radar target identification (RTI) is an important developing field of modern radar technology. Modern radar technology become mature day by day makes it possible to develop a practical RTI system. The dissertation investigates the application of updating technology of pattern recognition in RTI based on range profile. The main contributions of this dissertation are as follows:In chapter one, a brief summary of basic concept and various methods of RTI including their merits and drawbacks are surveyed. The main work of this paper is outlined.In chapter two, the basic concept and obtained way of range profile are introduced. The characteristics of range profile that vary with target aspect are analyzed deeply. Then we discussed the pattern invariable transform and feature extraction way, named K-L transform.In chapter three, we mainly analyze some statistic properties of pattern samples and discuss how to choose training samples and testing samples. They are usually ignored in many references. We firstly analyze the statistic properties of pattern samples. Next we consider how to choose the testing samples and training samples to ensure they have the same statistic properties. It's important to ensure the reliability of the recognition result in the following chapter.In chapter four, we consider radar range profile classification based on fuzzy min-max neural networks . We use the merging of the same kind of target's fuzzy set hyperboxes to form the radar target's range profile feature's orbit We use three kinds of planes' range profile to do classifying experiment. The results indicate that using fuzzy min-max neuron network classifier to classify radar target range profile has high recognition rate.In chapter five, we consider radar target range profile recognition based on open sample sets. This is another kind of pattern recognition, i.e. pattern recognition based on open sample sets, which is more difficult and more useful than pattern recognition based on close set.In this chapter we not only consider the rejection of confusers, but also consider how to utilize the rejected samples, since the collection of targets' samples is not easy, especially the range profiles of enemy's targets.In this chapter we consider two kinds of recognition based on open sample sets. One is that we know prior that the n testing samples come from one pattern class which may have been trained by the classifier or not, the other is that we do not know prior that the n testing samples come from one pattern class which may have been trained by the classifier or not. In each case we first consider how to reject the confusers. We also consider how to utilize the rejected samples. We use some methods to judge whether there exists a new kind of pattern in those refused patterns. When there is a new kind of pattern, we can use the refused patterns to retrain the network with other pattern classes' patterns we have got previously. When retraining has completed, the new target just now will becomes a known pattern class. We can use the network to classify patterns and find other new kinds ofpatterns.In chapter six, we use a support vector machine to classify radar targets. We can modify the kernel of the support vector machine based on the pattern samples, and point out the difficult when we use support vector machine to classify radar targets.There are usually three kinds of kernels used in using support vector machine to classify patterns. In this chapter we compare the difference when use these three kinds of kernels to classify radar targets. Usually we do not consider how to use the samples' information when use the three kernels to classify. In this chapter we modify the kernel of the support vector machine based on the pattern samples. When we use the modified kernel to classify radar targets, we find it can shorten the training time, and reduce the number of support vectors, and also get a good recognition result.In chapter seven, we discussed two problems about combination of multiple classifiers, one is that the dependence of classifiers used to combine; another is how about the relationship between the recognition result of combination of multiple classifiers and that of each classifier used to combine. We got the result that choosing a few of classifiers from many classifiers to combine can lower the dependence of the chosen classifiers. We proved that sometime the recognition result of combination of multiple classifiers may be better than that of each classifier used to combine, while sometime it will be worse than the recognition result of each classifier used to combine.