An Automatic Target Recognition (ATR) system is a sensor which is usuallyable to recognize targets or objects based on gathered data. The applicationof automatic target recognition technology is a critical element of robotic warfare.ATR systems are used in unmanned aerial vehicles and cruise missiles.There are many systems which are able to collect data (e.g. radar sensor,electro-optic sensor, infra-red devices) which are commonly used to collectinformation and detect, recognise and classify potential targets. Despite significanteffort during the last decades, some problems in ATR systems havenot been solved yet.This Ph.D. tried to understand the variation of the information content intoan ATR system and how to measure as well as how to preserve informationwhen it passes through the processing chain because they have not beeninvestigated properly yet. Moreover the investigation focused also on thedefinition of class-separability in ATR system and on the definition of thedegree of separability. As a consequence, experiments have been performedfor understanding how to assess the degree of class-separability and how thechoice of the parameters of an ATR system can affect the final classifier performance(i.e. selecting the most reliable as well as the most informationiiiiipreserving ones).As results of the investigations of this thesis, some important results havebeen obtained: Definition of the class-separability and of the degree of classseparability(i.e. the requirements that a metric for class-separability hasto satisfy); definition of a new metric for assessing the degree of classseparability;definition of the most important parameters which affect theclassifier performance or reduce/increase the degree of class-separability (i.e.Signal to Clutter Ratio, Clutter models, effects of despeckling processing).Particularly the definition of metrics for assessing the presence of artefactsintroduced by denoising algorithms, the ability of denoising algorithms inpreserving geometrical features of potential targets, the suitability of currentmathematical models at each stage of processing chain (especially for cluttermodels in radar systems) and the measurement of variation of informationcontent through the processing chain are some of them most important issueswhich have been investigated.
展开▼
