The ATR(automatic target recognition) system has been developed during last several decades for various military purposes. Because of life-threatening situations in which the ATR system is used, the system is compelled to be equipped with real-time operating recognition module. And also the kind of target objects is so various, it makes the size of database very large. Furthermore, to handle the size variation of input object, some strategy is needed to construct the database in which features of various objects can be easily retrieved. In the time point of view, size of feature vector plays an important role in the total recognition time. So it is needed to reduce the size of feature vector. For this, transform coefficients can be used to pack up signal energy of pixels in spatial domain. And to adapt various sizes of input objects, feature vector spaces should be composed in a hierarchical manner.
The ATR system in this thesis uses an IR(infra-red) sensored image as an input. Because of the poor edginess of IR sensored imagery, the spatial distribution of the intensity is used as a feature for recognition, and all pixels in the object region is considered.
In the thesis, we propose a system using wavelet transform to satisfy both issues on reducing size of feature vector and constructing hierarchical feature spaces. Wavelet transform shows relatively good energy compaction ratio, and transformed coefficients are structured suitable for multi-resolutional feature retrieving. At the matching phase of the recognition process, we propose the PNZL(paired nonzero list) matching scheme. PNZL matching scheme eliminates zero coefficients of transform domain, without affecting hierarchical structure of feature space. So the scheme can reduce the computational cost and the size of database effectively.
Simulation results shows that the proposed system can handle various sizes of input object and also achieve much shorter matching time than conventional template matching system while preserving the same recognition rate.