TRACKING INFORMATION IN SAR IMAGE FORMATION AND CLASSIFICATION ALGORITHMS

• Main Authors: Rajagopal, Abhejit, Radzicki, Vincent, Chandrasekaran, Shivkumar, Lee, Hua
• Other Authors: UCSB, Dept Electrical & Comp. Eng.
• Language: en_US
• Published: International Foundation for Telemetering 2017
• Online Access: http://hdl.handle.net/10150/626979; http://arizona.openrepository.com/arizona/handle/10150/626979

Traditional target detection pipelines involve two sequential steps: the formation of a range-profile or likely-image, and the classification of likely targets within that image. Although it has been shown that target tracking in the RaDAR image-domain can be unnecessarily noisy, with more accurate and efficient implementations involving a direct analysis of the measured wavefield, image formation remains a desirable output in many applications due to its highly descriptive and interpretable nature. In this paper, we outline a mechanism for formalizing and accelerating this procedure in application-specific use cases. Enabled by recent advances in deep learning, we present a pipeline for automatically selecting an “optimal” filtered back-projection model, forming a likelyimage, and performing target recognition and classification. The architecture allows practitioners to track and optimize the flow of information throughout the pipeline, enabling applications that utilize only intermediate outputs of the algorithm.