Many optical systems are used for specific tasks such as classification. Of these systems, th e majority are designedto maximize image quality for human observers; however, machine learning classification algorithmsdo not require the same data representation used by humans. In this work we investigate compressive opticalsystems optimized for a specific m a chine s e nsing t a s k. T wo c o mpressive o p t ical a r chitectures a r e e x a mined: anarray of prisms and neutral density filters w h e re e a ch p r i sm a n d n e u tral d e n sity fi lt er pa ir re al iz es on e datumfrom an optimized compressive sensing matrix, and another architecture using conventional optics to imagethe aperture onto the detector, a prism array to divide the aperture, and a pixelated attenuation mask in theintermediate image plane.We discuss the design, simulation, and tradeoffs of these compressive imaging systems built for compressedclassification of the MNIST data set. To evaluate the tradeoffs of the two architectures, we present radiometric andraytrace models for each system. Additionally, we investigate the impact of system aberrations on classi-ficationaccuracy of the system. We compare the performance of these systems over a range of compression. Classificationperformance, radiometric throughput, and optical design manufacturability are discussed.
展开▼
