Feasibility Studies of Nearest Neighbor Residual Vector Quantizer Classifiers fora Collection of Signal and Sensor Waveforms: Automatic Target Recognition in SAR Images

摘要

This executive summary contains a concise overview of the grant purpose, problemstatement and proposed solution, the research objective, and the technical approach used to achieve this objective. Experimental setups, performance results, and conclusions are also summarized. The purpose of this ONE grant is to support the evaluation of the performance of a particular joint compression/classification algorithm called nearest neighbor residual vector quantizer (NN-RVQ) classification on data obtained from a variety of sensor types and for a variety of applications. NN-RVQ is based on a recent mathematical development called direct sum successive approximations (DSSA). DSSA can be used as a technical foundation for data compression or pattern recognition algorithms, or for a single algorithm that does both. DSSA uses an unconventional mathematical data analysis/synthesis process to construct structured pattern dictionaries that can be efficiently searched (in terms of computation and memory). These patterns can be used as codevectors in vector quantizers (VQs) used for data compression, and as templates in nearest neighbor classifiers used for data classification. The purpose of this grant is to assess the performance of NN-RVQs when they are used for classification, compression, or joint classification and compression of various types of sensor data.