This paper addresses how the image processing steps involved in computational imaging can be adapted to specific image-based recognition tasks, and how significant reductions in computational complexity can be achieved by leveraging the recognition algorithm's robustness to defocus, poor exposure, and the like. Unlike aesthetic applications of computational imaging, recognition systems need not produce the best possible image quality, but instead need only satisfy certain quality thresholds that allow for reliable recognition. The paper specifically addresses light field processing for barcode scanning, and presents three optimizations which bring light field processing within the complexity limits of low-powered embedded processors.
Scott McCloskey, Scott McCloskey,
"Application-driven computational imaging", Proc. SPIE 9836, Micro- and Nanotechnology Sensors, Systems, and Applications VIII, 983604 (17 May 2016); doi: 10.1117/12.2225755; https://doi.org/10.1117/12.2225755