High Performance and Novel Target Recognition Algorithm in the Sparse Measurement Space

ABSTRACT: Our main objective in this project is to propose a novel Structured Sparse Priors for Target Recognition (SSPTR) system and to demonstrate that discriminative applications such as data clustering or target detection, tracking, and classification can be solved effectively and directly on the compressed measurement domain without the need to recover the original data. Our proposed sparse-representation discriminative algorithms have, at worst, the same level of complexity as popular sparse recovery algorithms in CS signal reconstruction while yielding comparable clustering/detection/classification accuracy as state-of-the-art discriminative strategies applying on original data. Here the crucial observation is that a test sample can be reasonably approximated as a linear combination of training samples belonging to the same class, with no contributions from training samples of other classes. Therefore, the sparse code which is often recovered via either basis pursuit or matching pursuit naturally encodes discriminative information that is crucial to classification tasks. In other words, the semantic (label) information of the signal of interest is directly captured in and instantaneously available from the sparse representation. Moreover, we also propose to further improve our baseline sparse-representation-based classification approach by the development of a novel unifying robust discriminative framework based on sparse representations directly on the collected measurements via context-aware and observable data-adaptive dictionaries and available domain-knowledge priors.; BENEFIT: Target/pattern detection, classification, and recognition applications will benefit more by incorporating such class-specific discriminative information than merely by conventional sparse signal recovery followed by a conventional classification strategy. Hence, we focus on maximizing the discriminability within the sparse recovery process by enforcing meaningful adaptive class-specific priors/constraints directly in the data measurement domain along with adaptive sparse representations in the measurement space explicitly for the purpose of image understanding and classification. ???Our proposed system can be used for missile seekers and other military surveillance and reconnaissance applications. We expect our software will have a unit price of $300 per device. With an estimated sales of over 20,000 units in the next decade, the military market potential results in more than 6 million dollars in the next decade. ????Besides military applications, ARLLCs technology will have many users in the commercial world. For example, border patrol, security monitoring in buildings and parking lots, coastal patrol, urban development monitoring, vegetation monitoring, hurricane damage assessment, and many others can benefit from our technology. We expect the commercial market size will be at least 20 million dollars over the next decade. ?